Service and network providers actively evaluate and derive Quality of Experience (QoE) metrics within their systems, which necessitates suitable monitoring strategies. Objective QoE monitoring involves mapping Quality of Service (QoS) parameters into QoE scores, such as calculating Mean Opinion Scores (MOS) or Good-or-Better (GoB) ratios, by using appropriate mapping functions. Alternatively, individual QoE monitoring directly assesses user experience based on self-reported feedback. We discuss the strengths, weaknesses, opportunities, and threats of both approaches. Based on the collected data from individual or objective QoE monitoring, providers can calculate the QoE metrics across all users in the system, who are subjected to a range of varying QoS conditions. The aggregated QoE across all users in the system for a dedicated time frame is referred to as system QoE. Based on a comprehensive simulation study, the expected system QoE, the system GoB ratio, as well as QoE fairness across all users are computed. Our numerical results explore whether objective and individual QoE monitoring lead to similar conclusions. In our previous work [Hoss2024], we provided a theoretical framework and the mathematical derivation of the corresponding relationships between QoS and system QoE for both monitoring approaches. Here, the focus is on illustrating the key differences of individual and objective QoE monitoring and the consequences in practice.

System QoE: Assessment of QoE of Users in a System

The term “System QoE” refers to the assessment of user experience from a provider’s perspective, focusing on the perceived quality of the users of a particular service. Thereby, providers may be different stakeholders along the service delivery chain, for example, network service provider and, in particular, Internet service provider, or application service provider. QoE monitoring delivers the necessary information to evaluate the system QoE, which is the basis for appropriate actions to ensure high-quality services and high QoE, e.g., through resource and network management.

Typically, QoE monitoring and management involves evaluating how well the network and services perform by analyzing objective metrics like Quality of Service (QoS) parameters (e.g., latency, jitter, packet loss) and mapping them to QoE metrics, such as Mean Opinion Scores (MOS). However, QoE monitoring involves a series of steps that providers need to follow: 1) identify relevant QoE metrics of interest, like MOS or GoB ratio; 2) deploy a monitoring framework to collect and analyze data. We will discuss this in the following.

The scope of system QoE metrics is to quantify the QoE across all users consuming the service for a dedicated time frame, e.g., one day, one week, or one month. Thereby, the expected QoE of an arbitrary user in the system, the ratio of all users experiencing Good-or-Getter (GoB) quality or Poor-or-Worse (PoW) quality, as well as the QoE fairness across all users are of interest. The users in the system may achieve different QoS on network level, e.g., different latency, jitter, throughput, since resources are shared among the users. The same is also true on application level with varying application-specific QoS parameters, for instance, video resolution, buffering time, or startup delays for video streaming. The varying QoS conditions manifest then in the system QoE. Fundamental relationships between the system QoE and QoS metrics were derived in [Hoss2020].

Expected system QoE: The expected system QoE is the average QoE rating of an arbitrary user in the system. The fundamental relationship in [Hoss2020] shows that the expected system QoE may be derived by mapping the QoS as experienced by a user to the corresponding MOS value and computing the average MOS over the varying QoS conditions. Thus, a MOS mapping function is required to map the QoS parameters to MOS values.

System GoB and System PoW: The Mean Opinion Score provides an average score but fails to account for the variability in users and the user rating diversity. Thus, users obtaining the same QoS conditions, may rate this subjectively differently. Metrics like the percentage of users rating the experience as Good or Better or as Poor or Worse provide more granular insights. Such metrics help service providers understand not just the average quality, but how quality is distributed across the user base. The fundamental relationship in [Hoss2020] shows that the system GoB and PoW may be derived by mapping the QoS as experienced by a user to the corresponding GoB or PoW value and computing the average over the varying QoS conditions, respectively. Thus, a GoB or PoW mapping function is required.

QoE Fairness: Operators must not only ensure that users are sufficiently satisfied, but also that this is done in a fair manner. However, what is considered fair in the QoS domain may not necessarily translate to fairness in the QoE domain, making the need to apply a QoE fairness index. [Hoss2018] defines the QoE fairness index as a linear transformation of the standard deviation of MOS values to the range [0;1]. The observed standard deviation is normalized with the maximal standard deviation, being theoretically possible for MOS values in a finite range, typically between 1 (poor quality) and 5 (excellent quality). The difference between 1 (indicating perfect fairness) and the normalized standard deviation of MOS values (indicating the degree of unfairness) yields the fairness index.

The fundamental relationships allow different implementations of QoE monitoring in practice, which are visualized in Figure 1 and discussed in the following. We differentiate between individual QoE monitoring and objective QoE monitoring and provide a qualitative strengths-weaknesses-opportunities-threats (SWOT) analysis.

Individual QoE Monitoring

Individual QoE monitoring refers to the assessment of system QoE by collecting individual ratings, e.g., on a 5-point rating scale, from users through their personal feedback. This approach captures the unique and individual nature of user experiences, accounting for factors like personal preferences and context. It allows optimizing services in a personalized manner, which is regarded as a challenging future research objective, see [Schmitt2017, Zhu2018, Gao2020, Yamazaki2021, Skorin-Kapov2018].

The term “individual QoE” was nicely described by in [Zhu2018]: “QoE, by definition, is supposed to be subjective and individual. However, we use the term ‘individual QoE’, since the majority of the literature on QoE has not treated it as such. […] The challenge is that the set of individual factors upon which an individual’s QoE depends is not fixed; rather this (sub)set varies from one context to another, and it is this what justifies even more emphatically the individuality and uniqueness of a user’s experience – hence the term ‘individual QoE’.”

Strengths: Individual QoE monitoring provides valuable insights into how users personally experience a service, capturing the variability and uniqueness of individual perceptions that objective metrics often miss. A key strength is that it gathers direct feedback from a provider’s own users, ensuring a representative sample rather than relying on external or unrepresentative populations. Additionally, it does not require a predefined QoE model, allowing for flexibility in assessing user satisfaction. This approach enables service providers to directly derive various system QoE metrics.

Weaknesses: Individual QoE monitoring is mainly feasible for application service providers and requires additional monitoring efforts beyond the typical QoS tools already in place. Privacy concerns are significant, as collecting sensitive user data can raise issues with data protection and regulatory compliance, such as with GDPR. Additionally, users may use the system primarily as a complaint tool, focusing on reporting negative experiences, which could skew results. Feedback fatigue is another challenge, where users may become less willing to provide ongoing input over time, limiting the validity and reliability of the data collected.

Opportunities: Data from individual QoE monitoring can be utilized to enhance individual user QoE through better resource and service management. From a business perspective, offering a personalized QoE can set providers apart in competitive markets and the data collected has monetization potential, supporting personalized marketing. Data from individual QoE monitoring enables deriving objective metrics like MOS or GoB, to update existing QoE models or to develop new QoE models for novel services by correlating it with QoS parameters. Those insights can drive innovation, leading to new features or services that meet evolving customer needs.

Threats: Individual QoE monitoring accounts for factors outside the provider’s control, such as environmental context (e.g., noisy surroundings [Reichl2015, Jiménez2020]), which may affect user feedback but not reflect actual service performance. Additionally, as mentioned, it may be used as a complaint tool, with users disproportionately reporting negative experiences. There is also the risk of over-engineering solutions by focusing too much on minor individual issues, potentially diverting resources from addressing more significant, system-wide challenges that could have a broader impact on overall service quality

Objective QoE Monitoring

Objective QoE monitoring involves assessing user experience by translating measurable QoS parameters on network level, such as latency, jitter, and packet loss, and on application level, such as video resolution or stalling duration for video streaming, into QoE metrics using predefined models and mapping functions. Unlike individual QoE monitoring, it does not require direct user feedback and instead relies on technically measurable parameters to estimate user satisfaction and various QoE metrics [Hoss2016]. Thereby, the fundamental relationships between system QoE and QoS [Hoss2020] are utilized. For computing the expected system QoE, a MOS mapping function is required, which maps a dedicated QoS value to a MOS value. For computing the system GoB, a GoB mapping function between QoS and GoB is required. Note that the QoS may be a vector of various QoS parameters, which are the input values for the mapping function.

Recent works [Hoss2022] indicated that industrial user experience index values, as obtained by the Threshold-Based Quality (TBQ) model for QoE monitoring, may be accurate enough to derive system QoE metrics. The TBQ model is a framework that defines application-specific thresholds for QoS parameters to assess and classify the user experience, which may be derived with simple and interpretable machine learning models like decision trees.

Strengths: Objective QoE monitoring relies solely on QoS monitoring, making it applicable for network providers, even for encrypted data streams, as long as appropriate QoE models are available, see for example [Juluri2015, Orsolic2020, Casas2022]. It can be easily integrated into existing QoS monitoring tools already deployed, reducing the need for additional resources or infrastructure. Moreover, it offers an objective assessment of user experience, ensuring that the same QoS conditions for different users are consistently mapped to the same QoE scores, as required for QoE fairness.

Weaknesses: Objective QoE monitoring requires specific QoE models and mapping functions for each desired QoE metric, which can be complex and resource-intensive to develop. Additionally, it has limited visibility into the full user experience, as it primarily relies on network-level metrics like bandwidth, latency, and jitter, which may not capture all factors influencing user satisfaction. Its effectiveness is also dependent on the accuracy of the monitored QoS metrics; inaccurate or incomplete data, such as from encrypted packets, can lead to misguided decisions and misrepresentation of the actual user experience.

Opportunities: Objective QoE monitoring enables user-centric resource and network management for application and network service providers by tracking QoS metrics, allowing for dynamic adjustments to optimize resource utilization and improve service delivery. The integration of AI and automation with QoS monitoring can increase the efficiency and accuracy of network management from a user-centric perspective. The objective QoE monitoring data can also enhance Service Level Agreements (SLAs) towards Experience Level Agreements (ELAs) as discussed in [Varela2015].

Threats: One risk of Objective QoE monitoring is the potential for incorrect traffic flow characterization, where data flows may be misattributed to the wrong applications, leading to inaccurate QoE assessments. Additionally, rapid technological changes can quickly make existing QoS monitoring tools and QoE models outdated, necessitating constant upgrades and investment to keep pace with new technologies. These challenges can undermine the accuracy and effectiveness of objective QoE monitoring, potentially leading to misinformed decisions and increased operational costs.

Numerical Results: Visualizing the Differences

In this section, we explore and visualize the obtained system QoE metrics, which are based on collected data either through i) individual QoE monitoring or ii) objective QoE monitoring. The question arises if the two monitoring approaches lead to the same results and conclusions for the provider. The obvious approach for computing the system QoE metrics is to use i) the individual ratings collected directly from the users and ii) the MOS scores obtained through mapping the objectively collected QoS parameters. While the discrepancies are derived mathematically in [Hoss2024], this article presents a visual representation of the differences between individual and objective QoE monitoring through a comprehensive simulation study. This simulation approach allows us to quantify the expected system QoE, the system GoB ratio, and the QoE fairness for a multitude of potential system configurations, which we manipulate in the simulation with varying QoS distributions. Furthermore, we demonstrate methods for utilizing data obtained through either individual QoE monitoring or objective QoE monitoring to accurately calculate the system QoE metrics as intended for a provider.

For the numerical results, the web QoE use case in [Hoss2024] is employed. We conduct a comprehensive simulation study, in which the QoS settings are varied. To be more precise, the page load times (PLTs) are varied, such that the users in the system experience a range of different loading times. For each simulation run, the average PLT and the standard deviation of the PLT across all users in the system are fixed. Then each user gets a randomly assigned PLT according to a beta distribution in the range between 0s and 8s with the specified average and standard deviation. The PLTs per user are sampled from that parameterized beta distribution.

For a concrete PLT, the corresponding user rating distribution is available and follows in our case a shifted binomial distribution, where the mean of the binomial distribution reflects the MOS value for that condition. To mention this clearly, this binomial distribution is a conditional random variable with discrete values on a 5-point scale: the user ratings are conditioned on the actual QoS value. For the individual QoE monitoring, the user ratings are sampled from that conditional random variable, while the QoS values are sampled from the beta distribution. For objective QoE monitoring, only the QoS values are used, but in addition, the MOS mapping function provided in [Hoss2024] is used. Thus, each QoS value is mapped to a continuous MOS value within the range of 1 to 5.

Figure 2 shows the expected system QoE using individual QoE monitoring as well as objective QoE monitoring depending on the average QoS as well as the standard deviation of the QoS, which is indicated by the color. Each point in the figure represents a single simulation run with a fixed average QoS and fixed standard deviation. It can be seen that both QoE monitoring approaches lead to the same results, which was also formally proven in [Hoss2024]. Note that higher QoS variances also result in higher expected system since for the same average QoS, there may be some users with larger QoS values, but also some users with lower QoS values. Due to the non-linear mapping between QoS and QoE this results in higher QoE scores.

Figure 3 shows the system GoB ratio, which can be simply computed with individual QoE monitoring. However, in the case of objective QoE monitoring, we assume that only a MOS mapping function is available. It is tempting to derive the GoB ratio by deriving the ratio of MOS values which are good or better. However, this leads to wrong results, see [Hoss2020]. Nevertheless, the GoB mapping function can be approximated from an existing MOS mapping function, see [Hoss2022, Hoss2017, Perez2023]. Then, the same conclusions are then derived through objective QoE monitoring as for individual QoE monitoring.

Figure 4 considers now QoE fairness for both monitoring approaches. It is tempting to use the user rating values from individual QoE monitoring and apply the QoE fairness index. However, in that case, the fairness index considers the variances of the system QoS and additionally the variances due to user rating diversity, as shown in [Hoss2024]. However, this is not the intended application of the QoE fairness index, which aims to evaluate the fairness objectively from a user-centric perspective, such that resource management can be adjusted and to provide users with high and fairly distributed quality. Therefore, the QoE fairness index uses MOS values, such that users with the same QoS are assigned the same MOS value. In a system with deterministic QoS conditions, i.e., the standard deviation diminishes, the QoE fairness index is 100%, see the results for the objective QoE monitoring. Nevertheless, the individual QoE monitoring also allows computing the MOS values for similar QoS values and then to apply the QoE fairness index. Then, comparable results are obtained as for objective QoE monitoring.

Conclusions

Individual QoE monitoring and objective QoE monitoring are fundamentally distinct approaches for assessing system QoE from a provider’s perspective. Individual QoE monitoring relies on direct user feedback to capture personalized experiences, while objective QoE monitoring uses QoS metrics and QoE models to estimate QoE metrics. Both methods have strengths and weaknesses, offering opportunities for service optimization and innovation while facing challenges such as over-engineering and the risk of models becoming outdated due to technological advancements, as summarized in our SWOT analysis. However, as the numerical results have shown, both approaches can be used with appropriate modifications and adjustments to derive various system QoE metrics like expected system QoE, system GoB and PoW ratio, as well as QoE fairness. A promising direction for future research is the development of hybrid approaches that combine both methods, allowing providers to benefit from objective monitoring while integrating the personalization of individual feedback. This could also be interesting to integrate in existing approaches like the QoS/QoE Monitoring Engine proposal [Siokis2023] or for upcoming 6G networks, which may allow the radio access network (RAN) to autonomously adjust QoS metrics in collaboration with the application to enhance the overall QoE [Bertenyi2024].

References

[Bertenyi2024] Berteny, B., Kunzmann, G., Nielsen, S., and Pedersen, K. Andres, P. (2024). Transforming the 6G vision to action. Nokia Whitepaper, 28 June 2024. Url: https://www.bell-labs.com/institute/white-papers/transforming-the-6g-vision-to-action/.

[Casas2022] Casas, P., Seufert, M., Wassermann, S., Gardlo, B., Wehner, N., & Schatz, R. (2022). DeepCrypt: Deep learning for QoE monitoring and fingerprinting of user actions in adaptive video streaming. In 2022 IEEE 8th International Conference on Network Softwarization (NetSoft) (pp. TBD). IEEE.

[Gao2020] Gao, Y., Wei, X., & Zhou, L. (2020). Personalized QoE improvement for networking video service. IEEE Journal on Selected Areas in Communications, 38(10), 2311-2323.

[Hoss2016] Hoßfeld, T., Schatz, R., Egger, S., & Fiedler, M. (2016). QoE beyond the MOS: An in-depth look at QoE via better metrics and their relation to MOS. Quality and User Experience, 1, 1-23.

[Hoss2017] Hoßfeld, T., Fiedler, M., & Gustafsson, J. (2017, May). Betas: Deriving quantiles from MOS-QoS relations of IQX models for QoE management. In 2017 IFIP/IEEE Symposium on Integrated Network and Service Management (IM) (pp. 1011-1016). IEEE.

[Hoss2018] Hoßfeld, T., Skorin-Kapov, L., Heegaard, P. E., & Varela, M. (2018). A new QoE fairness index for QoE management. Quality and User Experience, 3, 1-23.

[Hoss2020] Hoßfeld, T., Heegaard, P. E., Skorin-Kapov, L., & Varela, M. (2020). Deriving QoE in systems: from fundamental relationships to a QoE-based Service-level Quality Index. Quality and User Experience, 5(1), 7.

[Hoss2022] Hoßfeld, T., Schatz, R., Egger, S., & Fiedler, M. (2022). Industrial user experience index vs. quality of experience models. IEEE Communications Magazine, 61(1), 98-104.

[Hoss2024] Hoßfeld, T., & Pérez, P. (2024). A theoretical framework for provider’s QoE assessment using individual and objective QoE monitoring. In 2024 16th International Conference on Quality of Multimedia Experience (QoMEX) (pp. TBD). IEEE.

[Jiménez2020] Jiménez, R. Z., Naderi, B., & Möller, S. (2020, May). Effect of environmental noise in speech quality assessment studies using crowdsourcing. In 2020 Twelfth International Conference on Quality of Multimedia Experience (QoMEX) (pp. 1-6). IEEE.

[Juluri2015] Juluri, P., Tamarapalli, V., & Medhi, D. (2015). Measurement of quality of experience of video-on-demand services: A survey. IEEE Communications Surveys & Tutorials, 18(1), 401-418.

[Orsolic2020] Orsolic, I., & Skorin-Kapov, L. (2020). A framework for in-network QoE monitoring of encrypted video streaming. IEEE Access, 8, 74691-74706.

[Perez2023] Pérez, P. (2023). The Transmission Rating Scale and its Relation to Subjective Scores. In 2023 15th International Conference on Quality of Multimedia Experience (QoMEX) (pp. 31-36). IEEE.

[Reichl2015] Reichl, P., et al. (2015, May). Towards a comprehensive framework for QoE and user behavior modelling. In 2015 Seventh International Workshop on Quality of Multimedia Experience (QoMEX) (pp. 1-6). IEEE.

[Schmitt2017] Schmitt, M., Redi, J., Bulterman, D., & César, P. (2017). Towards individual QoE for multiparty videoconferencing. IEEE Transactions on Multimedia, 20(7), 1781-1795.

[Siokis2023] Siokis, A., Ramantas, K., Margetis, G., Stamou, S., McCloskey, R., Tolan, M., & Verikoukis, C. V. (2023). 5GMediaHUB QoS/QoE monitoring engine. In 2023 IEEE 28th International Workshop on Computer Aided Modeling and Design of Communication Links and Networks (CAMAD) (pp. TBD). IEEE.

[Skorin-Kapov2018] Skorin-Kapov, L., Varela, M., Hoßfeld, T., & Chen, K. T. (2018). A survey of emerging concepts and challenges for QoE management of multimedia services. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 14(2s), 1-29.

[Varela2015] Varela, M., Zwickl, P., Reichl, P., Xie, M., & Schulzrinne, H. (2015, June). From service level agreements (SLA) to experience level agreements (ELA): The challenges of selling QoE to the user. In 2015 IEEE International Conference on Communication Workshop (ICCW) (pp. 1741-1746). IEEE.

[Yamazaki2021] Yamazaki, T. (2021). Quality of experience (QoE) studies: Present state and future prospect. IEICE Transactions on Communications, 104(7), 716-724.

[Zhu2018] Zhu, Y., Guntuku, S. C., Lin, W., Ghinea, G., & Redi, J. A. (2018). Measuring individual video QoE: A survey, and proposal for future directions using social media. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 14(2s), 1-24.

Data-driven Quality of Experience (QoE) modelling using Machine Learning (ML) arose as a promising alternative to the cumbersome and potentially biased manual QoE modelling. However, the reasoning of a majority of ML models is not explainable due to their black-box characteristics, which prevents us from gaining insights about how the model actually related QoE influence factors and QoE. These fundamental relationships are highly relevant for QoE researchers and service and network providers though.

With the emerging field of eXplainable Artificial Intelligence (XAI) and its recent technological advances, these issues can now be resolved. As a consequence, XAI enables data-driven QoE modelling to obtain generalizable QoE models and provides us simultaneously with the model’s reasoning on which QoE factors are relevant and how they affect the QoE score. In this work, we showcase the feasibility of explainable data-driven QoE modelling for video streaming and web browsing, before we discuss the opportunities and challenges of deploying XAI for QoE modelling.

Introduction

In order to enhance services and networks and prevent users from switching to competitors, researchers and service providers need a deep understanding of the factors that influence the Quality of Experience (QoE) [1]. However, developing an effective QoE model is a complex and costly endeavour. Typically, it requires dedicated and extensive studies, which can only cover a limited portion of the parameter space and may be influenced by the study design. These studies often generate a relatively small sample of QoE ratings from a comparatively small population, making them vulnerable to poor performance when applied to unseen data. Moreover, the process of collecting and processing data for QoE modelling is not only arduous and time-consuming, but it can also introduce biases and self-fulfilling prophecies, such as perceiving an exponential relationship when one is expected.

To overcome these challenges, data-driven QoE modelling utilizing machine learning (ML) has emerged as a promising alternative, especially in scenarios where there is a wealth of data available or where data streams can be continuously obtained. A notable example is the ITU-T standard P.1203 [2], which estimates video streaming QoE by combining manual modelling – accounting for 75% of the Mean Opinion Score (MOS) estimation – and ML-based Random Forest modelling – accounting for the remaining 25%. The inclusion of the ML component in P.1203 indicates its ability to enhance performance. However, the inner workings of P.1203’s Random Forest model, specifically how it calculates the output score, are not obvious. Also, the survey in [3] shows that ML-based QoE modelling in multimedia systems is already widely used, including Virtual Reality, 360-degree video, and gaming. However, the QoE models are based on shallow learning methods, e.g., Support Vector Machines (SVM), or on deep learning methods, which lack explainability. Thus, it is difficult to understand what QoE factors are relevant and how they affect the QoE score [13], resulting in a lack of trust in data-driven QoE models and impeding their widespread adoption by researchers and providers [14].

Fortunately, recent advancements in the field of eXplainable Artificial Intelligence (XAI) [6] have paved the way for interpretable ML-based QoE models, thereby fostering trust between stakeholders and the QoE model. These advancements encompass a diverse range of XAI techniques that can be applied to existing black-box models, as well as novel and sophisticated ML models designed with interpretability in mind. Considering the use case of modelling video streaming QoE from real subjective ratings, the work in [4] evaluates the feasibility of explainable, data-driven QoE modelling and discusses the deployment of XAI for QoE research.

The utilization of XAI for QoE modelling brings several benefits. Not only does it speed up the modelling process, but it also enables the identification of the most influential QoE factors and their fundamental relationships with the Mean Opinion Score (MOS). Furthermore, it helps eliminate biases and preferences from different research teams and datasets that could inadvertently influence the model. All that is required is a selective dataset with descriptive features and corresponding QoE ratings (labels), which covers the most important QoE influence factors and, in particular, also rare events, e.g., many stalling events in a session. Generating such complete datasets, however, is an open research question, but calls for data-centric AI [15]. By merging datasets from various studies, more robust and generalizable QoE models can theoretically be created. These studies need to have a common ground though. Another benefit is the fact that the models can also be automatically refined over time as new QoE studies are conducted and additional data becomes available.

XAI: eXplainable Artificial Intelligence

For a comprehensive understanding of eXplainable Artificial Intelligence (XAI), a general overview can be found in [5], while a thorough survey on XAI methods and a taxonomy of XAI methods, in general, is available in [6].

XAI methods can be categorized into two main types: local and global explainability techniques. Local explainability aims to provide explanations for individual stimuli in terms of QoE factors and QoE ratings. On the other hand, global explainability focuses on offering general reasoning for how a model derives the QoE rating from the underlying QoE factors. Furthermore, XAI methods can be classified into post-hoc explainers and interpretable models.

Post-hoc explainers [6] are commonly used to explain various black-box models, such as neural networks or ensemble techniques after they have been trained. One widely utilized post-hoc explainer is SHAP values [7], which originates from game theory. SHAP values quantify the contribution of each feature to the model’s prediction by considering all possible feature subsets and learning a model for each subset. Other post-hoc explainers include LIME and Anchors, although they are limited to classification tasks.

Interpretable models, by design, provide explanations for how the model arrives at its output. Well-known interpretable models include linear models and decision trees. Additionally, generalized additive models (GAM) are gaining recognition as interpretable models.

A GAM is a generalized linear model in which the model output is computed by summing up each of the arbitrarily transformed input features along with a bias [8]. The form of a GAM enables a direct interpretation of the model by analyzing the learned functions and the transformed inputs, which allows to estimate the influence of a feature. Two state-of-the-art ML-based GAM models are Explainable Boosting Machine (EBM) [9] and Neural Additive Model (NAM) [8]. While EBM uses decision trees to learn the functions and gradient boosting to improve training, NAM utilizes arbitrary neural networks to learn the functions, resulting in a neural network architecture with one sub-network per feature. EBM extends GAM by also considering additional pairwise feature interaction terms while maintaining explainability.

Exemplary XAI-based QoE Modelling using GAMs

We demonstrate the learned predictor functions for both EBM (red) and NAM (blue) on a video QoE dataset in Figure 1. All technical details about the dataset and the methodology can be found in [4]. We observe that both models provide smooth shape functions, which are easy to interpret. EBM and NAM differ only marginally and mostly in areas where the data density is low. Here, EBM outperforms NAM by overfitting on single data points using the feature interaction terms. We can see this, for example, for a high total stalling duration and a high number of quality switches, where at some point EBM stops the negative trend and strongly contrasts its previous trend to improve predictions for extreme outliers.

Using the smooth predictor functions, it is easy to apply curve fitting. In the bottom right plot of Figure 1, we fit the average bitrate predictor function of NAM, which was shifted by the average MOS of the dataset to obtain the original MOS scale on the y-axis, on an inverted x-axis using exponential (IQX), logarithmic (WQL), and linear functions (LIN). Note that this constitutes a univariate mapping of average bitrate to MOS, neglecting the other influencing factors. We observe that our predictor function follows the WQL hypothesis [10] (red) with a high R²=0.967. This is in line with the mechanics of P.1203, where the authors of [11] showed the same logarithmic behavior for the bitrate in mode 0.

As the presented XAI methods are universally applicable to any QoE dataset, Figure 2 shows a similar GAM-based QoE modelling for a web QoE dataset obtained from [12]. We can see that the loading behavior in terms of ByteIndex-Page Load Time (BI-PLT) and time to last byte (TTLB) has the strongest impact on web QoE. Moreover, we see that different URLs/webpages have a different effect on the MOS, which shows that web QoE is content dependent. Summarizing, using GAMs, we obtain valuable easy to interpret functions, which explain fundamental relationships between QoE factors and MOS. Nevertheless, further XAI methods can be utilized, as detailed in [4,5,6].

Discussion

In addition to expediting the modelling process and mitigating modelling biases, data-driven QoE modelling offers significant advantages in terms of improved accuracy and generalizability compared to manual QoE models. ML-based models are not constrained to specific classes of continuous functions typically used in manual modelling, allowing them to capture more complex relationships present in the data. However, a challenge with ML-based models is the risk of overfitting, where the model becomes overly sensitive to noise and fails to capture the underlying relationships. Overfitting can be avoided through techniques like model regularization or by collecting sufficiently large or complete datasets.

Successful implementation of data-driven QoE modelling relies on purposeful data collection. It is crucial to ensure that all (or at least the most important) QoE factors are included in the dataset, covering their full parameter range with an adequate number of samples. Controlled lab or crowdsourcing studies can define feature values easily, but budget constraints (time and cost) often limit data collection to a small set of selected feature values. Conversely, field studies can encompass a broader range of feature values observed in real-world scenarios, but they may only gather limited data samples for rare events, such as video sessions with numerous stalling events. To prevent data bias, it is essential to balance feature values, which may require purposefully generating rare events in the field. Additionally, thorough data cleaning is necessary. While it is possible to impute missing features resulting from measurement errors, doing so increases the risk of introducing bias. Hence, it is preferable to filter out missing or unusual feature values.

Moreover, adding new data and retraining an ML model is a natural and straightforward process in data-driven modelling, offering long-term advantages. Eventually, data-driven QoE models would be capable of handling concept drift, which refers to changes in the importance of influencing factors over time, such as altered user expectations. However, QoE studies are rarely conducted as temporal and population-based snapshots, limiting frequent model updates. Ideally, a pipeline could be established to provide a continuous stream of features and QoE ratings, enabling online learning and ensuring the QoE models remain up to date. Although challenging for research endeavors, service providers could incorporate such QoE feedback streams into their applications

Comparing black-box and interpretable ML models, there is a slight trade-off between performance and explainability. However, as shown in [4], it should be negligible in the context of QoE modelling. Instead, XAI allows to fully understand the model decisions, identifying relevant QoE factors and their relationships to the QoE score. Nevertheless, it has to be considered that explaining models becomes inherently more difficult when the number of input features increases. Highly correlated features and interactions may further lead to misinterpretations when using XAI since the influence of a feature may also depend on other features. To obtain reliable and trustworthy explainable models, it is, therefore, crucial to exclude highly correlated features.

Finally, although we demonstrated XAI-based QoE modelling only for video streaming and web browsing, from a research perspective, it is important to understand that the whole process is easily applicable in other domains like speech or gaming. Apart from that, it can also be highly beneficial for providers of services and networks to use XAI when implementing a continuous QoE monitoring. They could integrate visualizations of trends like Figure 1 or Figure 2 into dashboards, thus, allowing to easily obtain a deeper understanding of the QoE in their system.

Conclusion

In conclusion, the progress in technology has made data-driven explainable QoE modeling suitable for implementation. As a result, it is crucial for researchers and service providers to consider adopting XAI-based QoE modeling to gain a comprehensive and broader understanding of the factors influencing QoE and their connection to users’ subjective experiences. By doing so, they can enhance services and networks in terms of QoE, effectively preventing user churn and minimizing revenue losses.

References

[1] K. Brunnström, S. A. Beker, K. De Moor, A. Dooms, S. Egger, M.-N. Garcia, T. Hossfeld, S. Jumisko-Pyykkö, C. Keimel, M.-C. Larabi et al., “Qualinet White Paper on Definitions of Quality of Experience,” 2013.

[2] W. Robitza, S. Göring, A. Raake, D. Lindegren, G. Heikkilä, J. Gustafsson, P. List, B. Feiten, U. Wüstenhagen, M.-N. Garcia et al., “HTTP Adaptive Streaming QoE Estimation with ITU-T Rec. P. 1203: Open Databases and Software,” in ACM MMSys, 2018

[3] G. Kougioumtzidis, V. Poulkov, Z. D. Zaharis, and P. I. Lazaridis, “A Survey on Multimedia Services QoE Assessment and Machine Learning-Based Prediction,” IEEE Access, 2022.

[4] N. Wehner, A. Seufert, T. Hoßfeld, M. and Seufert, “Explainable Data-Driven QoE Modelling with XAI,” QoMEX, 2023.

[5] C. Molnar, Interpretable Machine Learning, 2nd ed., 2022. Available: https://christophm.github.io/interpretable-ml-book

[6] A. B. Arrieta, N. Diıaz-Rodriguez et al., “Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges Toward Responsible AI,” Information fusion, 2020.

[7] S. M. Lundberg and S.-I. Lee, “A Unified Approach to Interpreting Model Predictions,” NIPS, 2017.

[8] R. Agarwal, L. Melnick, N. Frosst, X. Zhang, B. Lengerich, R. Caruana, and G. E. Hinton, “Neural Additive Models: Interpretable MachineLearning with Neural Nets,” NIPS, 2021.

[9] H. Nori, S. Jenkins, P. Koch, and R. Caruana, “InterpretML: A Unified Framework for Machine Learning Interpretability,” arXiv preprint arXiv:1909.09223, 2019.

[10] T. Hoßfeld, R. Schatz, E. Biersack, and L. Plissonneau, “Internet Video Delivery in YouTube: From Traffic Measurements to Quality of Experience,” in Data Traffic Monitoring and Analysis, 2013.

[11] M. Seufert, N. Wehner, and P. Casas, “Studying the Impact of HAS QoE Factors on the Standardized Qoe Model P. 1203,” in ICDCS, 2018

[12] D. N. da Hora, A. S. Asrese, V. Christophides, R. Teixeira, D. Rossi, “Narrowing the gap between QoS metrics and Web QoE using Above-the-fold metrics,” PAM, 2018

[13] A. Seufert, F. Wamser, D. Yarish, H. Macdonald, and T. Hoßfeld, “QoE Models in the Wild: Comparing Video QoE Models Using a Crowdsourced Data Set”, in QoMEX, 2021

[14] D. Shin, “The effects of explainability and causability on perception, trust, and acceptance: Implications for explainable AI”, in International Journal of Human-Computer Studies, 2021.

[15] D. Zha, Z. P. Bhat, K. H. Lai, F. Yang, & X. Hu, “Data-centric ai: Perspectives and challenges”, in SIAM International Conference on Data Mining, 2023

The exponential growth in internet data traffic, driven by the widespread use of video streaming applications, has resulted in increased energy consumption and carbon emissions. This outcome is primarily due to higher resolution or higher framerates content and the ability to watch videos on various end-devices. However, efforts to reduce energy consumption in video streaming services may have unintended consequences on users’ Quality of Experience (QoE). This column delves into the intricate relationship between QoE and energy consumption, considering the impact of different bit rates on end-devices. We also consider other factors to provide a more comprehensive understanding of whether these end-devices have a significant environmental impact. It is essential to carefully weigh the trade-offs between QoE and energy consumption to make informed decisions and develop sustainable practices in video streaming services.

Energy Consumption for Video Streaming

In the past few years, we have seen a remarkable expansion in how online content is delivered. According to Sandvine’s 2023 Global Internet Phenomena Report [1], video usage on the Internet has increased by 24% in 2022 and now accounts for 65% of all Internet traffic. This surge in video usage is mainly due to the growing popularity of streaming video services. Videos have become an increasingly popular form of online content, capturing a significant portion of internet users’ attention and shaping how we consume information and entertainment online. Therefore, the rising quality expectations of end-users have necessitated research and implementation of video streaming management approaches that consider the Quality of Experience (QoE) [2]. The idea is to develop applications that can work within the energy and resource limits of end-devices, while still delivering the Quality of Service (QoS) needed for smooth video viewing and an improved user experience (QoE). Even though video streaming services are advancing so quickly, energy consumption is still a significant issue causing many concerns about its impact and the urgent need to boost energy efficiency [14].

The literature provides four main elements: the data centres, the data transmission networks, the end-devices and the consumer behaviour analysing of the energy consumption of video streaming [3]. In this regard, in [4], the authors present a comprehensive review of existing literature on the energy consumption of online video streaming services. Then, they outline the potential actions that can be taken by both service providers and consumers to promote sustainable video streaming, drawing from the literature studies discussed. Their summary of the current possible actions for sustainable video streaming, from both the provider’s and consumer’s perspective, is expressed in the following segments with some of the possible solutions:

• Data center: CDN (Content Delivery Network) can be utilized to offload contents/applications to the edge from the provider’s side and choose providers that prioritize sustainability from the consumer’s side.
• Data transmission network: Data compression/encoding algorithms from the provider’s side and no autoplay from the consumer’s side.
• End-Device: Produce energy-efficient devices from the provider’s size and prefer small-size (mobile) devices from the consumer’s side.
• Consumer behaviour: Reduce the number of subscribers from the provider’s size and prefer watching videos with other people than alone from the consumer’s side.

Finally, they noted that the end device and consumer behaviour are the primary contributors to energy costs in the video streaming process. This result includes actions such as reducing video resolution and using smaller devices. However, taking such actions may have a potential downside as they can negatively impact the QoE due to their effect on video quality. Therefore, in [5], they found that by sacrificing the maximum QoE and aiming for good quality instead (e.g., MOS score of 4=Good instead of MOS score 5=Excellent), significant energy savings can be achieved in video-conferencing services. This is possible by using lower video bitrates compared to higher bitrates which result in higher energy consumption, as per their logarithmic QoE model. Regarding this research, in [4], the authors propose identifying an acceptable level of QoE, rather than striving for maximum QoE, as a potential solution to reduce energy consumption while still meeting consumer satisfaction. They conducted a crowdsourcing survey to gather real consumer opinions on their willingness to save energy consumption while streaming online videos. Then, they analysed the survey results to understand how willing people are to lower video streaming quality in order to achieve energy savings.

Green Video Streaming: The Trade-Off Between QoE and Energy Consumption

To provide a trade-off between QoE and Energy Consumption, we looked at the connection between video bitrate of standard resolution, electricity usage, and perceived QoE for a video streaming service on four different devices (smartphone, tablet, laptop/PC, and smart TV) as taken from [4].

They calculated the energy consumption of streaming on devices which is provided in [6]: Q_i = t_i.(P_i+R_i.ƿ), in the given equation, Q_i represents the electricity consumption (in kWh) of the i-th device, t_i denotes the streaming duration (in hours per week) for the i-th device, P_i represents the power load (in kW) of the i-th device, R_i signifies the data traffic (in GB/h) for a specific bitrate, and ρ = 0.1 kWh/GB represents the electricity intensity of data traffic.

Then,  to estimate the perceived QoE based on the video bitrate, the authors employed a QoE model from [7], as noted in their analysis which is: QoE = a.br^b + c, where “br” represents the bitrate, and “a”, “b”, and “c” are the regression coefficients calculated for specific resolutions.

After taking this into account, we can establish a link between the QoE model, energy consumption, and the perceived QoE associated with video bitrate across various end-devices. Therefore, we implemented the green QoE model in [8] to provide a trade-off between the perceived QoE and the calculated energy consumption from above in the following way: f_γ(x)= 4/(log(x’_5)-log(x_1))*log(x)+ (log(x’_5)-5*log(x_1))/(log(x’_5)-log(x_1)). The given equation represents the mapping function between video bitrate and Mean Opinion Scores (MOS), considering both the minimum bitrate x_1 corresponding to MOS 1 and the maximum bitrate x_5 corresponding to MOS 5. Moreover, the factor γ, representing the greenness of a user, is considered in the context of maximum bitrate x’_5 = x_5/γ, which results in a MOS score of 5.

The model focuses on the concept of a “green user,” who considers the energy consumption aspect in their overall QoE evaluations. Thus, a green user might rate their QoE slightly lower in order to reduce their carbon footprint compared to a high-quality (HQ) user (or “non-green” user) who prioritizes QoE without considering energy consumption.

The numerical results for the energy consumption (in kWh) and the MOS scores depending on the video bitrate can be simplified with linear and logarithmic regressions, respectively. In Figure 1, the graph depicts a linear regression analysis conducted to examine the relationship between energy consumption (kWh) and bitrate (kbps). The y-axis represents energy consumption while the x-axis represents bitrate (kbps). The graph displays a straight-line trend that starts at 1.6 kWh and extends up to 3.5 kWh as the bitrate increases. The linear fitting function used for the analysis is formulated as: kWh = f(bitrate) = a * bitrate + c, where ‘a’ represents the slope and ‘c’ represents the y-intercept of the line.

Figure 1 visually illustrates how energy consumption tends to increase with higher bitrates, as indicated by the positive slope of the linear regression line in Figure 1. One notable observation is that as video bitrates increase, the electricity consumption of end-devices also tends to increase. This can be attributed to the larger amount of data traffic generated by higher-resolution video content, which requires higher bitrates for transmission. Consequently, smart TVs are likely to consume more energy compared to other devices. This finding is consistent with the results obtained from the linear regression model, as described in [4], further validating the relationship between bitrate and energy consumption.

As illustrated in Figure 2, the relationship between MOS and video bitrate (kbps) follows a logarithmic pattern. Therefore, we can use a straightforward QoE model to estimate the MOS if there is information about the video bitrate. This can be achieved by utilizing a logistic regression model MOS(x), where MOS = f(x) = a * log(x) + c, with x representing the video bitrate in Mbps, and a and c being coefficients, as provided in [9]. After, MOS and video bitrate (kbps) values in [4] are applied to the above-mentioned QoE green model equation regarding the logistic regression model, which is an extension of the logarithmic regression model [8]. This relationship allows to determine the green user QoE model and we exemplary show the green user QoE model for smart TV (using γ=2 in f_γ(x)).

According to Figure 2, it is categorized users into two groups: those who prioritize high-quality (HQ) video regardless of energy consumption, and green users who prioritize energy efficiency while still being satisfied with slightly lower video quality. It can be observed that the MOS value changes in video quality on their smart TVs faster compared to other end-devices.  This is evident from the steeper curve in the smart TV section. On the other hand, when looking at the curve for tablets, it shows that changes in bitrate have a milder impact on MOS values. The outcome suggests that video streaming on smaller screens, such as tablets or laptops, may contribute less to the perception of quality changes. Considering that those small-screen devices consume less energy than larger screen devices, it may be preferable to use lower resolution videos instead of high-resolution ones. Analysing the relationship between laptops and tablets, it can be seen that low-resolution video streaming on laptops resulted in lower MOS scores compared to the tablet. From this result, it can be inferred that the choice of end-device and user behaviour plays a significant role in energy savings. Figure 2 indicates that the MOS values for the green user of a smart TV is comparable to the MOS values of an HQ user using a laptop.

Concerning this outcome, in [10], the authors presented the results of a subjective assessment aimed at investigating how different factors, such as video resolution, luminance, and end devices (TV, Laptop, and Smartphone), impact the QoE and energy consumption of video streaming services. The study found that, in certain conditions such as dark or bright environments, low device backlight luminance, or small-screen devices), users may need to strike a balance between acceptable QoE and sustainable (green) choices, as consuming more energy (e.g., by streaming higher-quality videos) may not significantly enhance the QoE.

Therefore, Figure 3 plots the trade-off relationship between energy consumption (kWh) and MOS for the end devices (such as smart TV, laptop and tablet). Thereby, we differentiate the HQ user and the green user, which presents some interesting results. First, a MOS score of 4 leads to comparable energy consumption results for green and HQ users. The relative differences are rather small. However, aiming for best quality (MOS 5) leads to significant differences. Furthermore, it is seen that the device type has a significant impact on energy consumption. Even for green users, which rate lower bitrates with higher MOS scores than HQ users, the energy consumption of the smart TV is much higher than for any quality (i.e. bitrate) for laptop and tablet users. Thus, device type and user behaviour are essential to strike the right balance between QoE and energy consumption.

Discussions and Future Research

Meeting the QoE expectations of end-users is essential to fulfilling the requirements of video streaming services. As users are the primary viewers of streaming videos in most real-world scenarios, subjective QoE assessment [11] provides a direct and dependable means to evaluate the perceptual quality of video streaming. Furthermore, there is a growing need to create objective QoE assessment models provided in [12]– [13]. However, many studies have focused on investigating the QoE obtained through subjective and objective models and have overlooked the consideration of energy consumption in video streaming.

Therefore, in the previous section, we have discussed how the different elements within the video streaming ecosystem play a role in consuming energy and emitting CO2.  The findings pave the way for an objective response to determining an appropriate optimal video bitrate for viewing, considering both QoE and sustainability considerations, which can be further explored in future research.

It is evident that addressing energy consumption and emissions is crucial for the future of video streaming systems, while ensuring that end-users’ QoE is not compromised poses a significant and ongoing challenge. Thus, potential solutions to prevent energy consumption increase in QoE while still satisfying the user include streaming videos on smaller screen devices and watching lower resolution videos that offer sufficient quality instead of the highest resolution ones. Here, it can be highlighted the importance of user behavior to prevent energy consumption. Additionally, trade-off models can be developed using the green QoE model (especially for smarTV) by identifying ideal bitrate values for energy savings and user satisfaction in the QoE.

Delving deeper into the dynamics of the video streaming ecosystem, it becomes increasingly clear that energy consumption and emissions are critical concerns that must be addressed for the sustainable future of video streaming systems. The environmental impact of video streaming, particularly in terms of carbon emissions, cannot be understated. With the growing awareness of the urgent need to combat climate change, mitigating the environmental footprint of video streaming has become a pressing priority.

As video streaming technologies evolve, optimizing energy-efficient approaches without compromising users’ QoE is a complex task. End-users, who expect seamless and high-quality video streaming experiences, should not be deprived of their QoE while addressing the energy and emissions concerns. The outcome opens a novel door for an objective answer to the question of what constitutes an appropriate optimal video bitrate for viewing that takes into account both QoE and sustainability concerns.

Future research in this area is crucial to explore innovative techniques and strategies that can effectively reduce the energy consumption and carbon emissions of video streaming systems without sacrificing the QoE. Additionally, collaborative efforts among stakeholders, including researchers, industry practitioners, policymakers, and end-users, are essential in devising sustainable video streaming solutions that consider both environmental and user experience factors [14].

In conclusion, the discussions on the relationship between energy consumption, emissions, and QoE in video streaming systems emphasize the need for continued research and innovation to achieve a sustainable balance between environmental sustainability and user satisfaction.

References

• [1] Sandvine. The Global Internet Phenomena Report. January 2023. Retrieved April 24, 2023
• [2] M. Seufert, S. Egger, M. Slanina, T. Zinner, T. Hoßfeld and P. Tran-Gia, “A Survey on Quality of Experience of HTTP Adaptive Streaming,” in IEEE Communications Surveys & Tutorials, vol. 17, no. 1, pp. 469-492, Firstquarter 2015, doi: 10.1109/COMST.2014.2360940., 2015.
• [3] Reinhard Madlener, Siamak Sheykhha, Wolfgang Briglauer,”The electricity- and CO2-saving potentials offered by regulation of European video-streaming services,” Energy Policy,vol. 161, p. 112716, 2022.
• [4] G. Bingöl, S. Porcu, A. Floris and L. Atzori, “An Analysis of the Trade-off between Sustainability,” in IEEE ICC Workshop-GreenNet, Rome, 2023.
• [5] T. Hoßfeld, M. Varela, L. Skorin-Kapov, P. E. Heegaard, “What is the trade-off between CO2 emission and video-conferencing QoE?,” ACM SIGMM Records, 2022.
• [6] P. Suski, J. Pohl, and V. Frick, “All you can stream: Investigating the role of user behavior for greenhouse gas intensity of video streaming,” in Proc. of the 7th Int. Conf. on ICT for Sustainability, 2020, pp. 128–138.
• [7] M. Mu, M. Broadbent, A. Farshad, N. Hart, D. Hutchison, Q. Ni, and N. Race, “A Scalable User Fairness Model for Adaptive Video Streaming Over SDN-Assisted Future Networks,” IEEE Journal on Selected Areas in Communications, vol. 34, no. 8, p. 2168–2184, 2016.
• [8] T. Hossfeld, M. Varela, L. Skorin-Kapov and P. E. Heegaard, “A Greener Experience: Trade-offs between QoE and CO2 Emissions in Today’s and 6G Networks,” IEEE Communications Magazine, pp. 1-7, 2023.
• [9] J. P. López, D. Martín, D. Jiménez and J. M. Menéndez, “Prediction and Modeling for No-Reference Video Quality Assessment Based on Machine Learning,” in 14th International Conference on Signal-Image Technology & Internet-Based Systems (SITIS), IEEE, pp. 56-63, Las Palmas de Gran Canaria, Spain, 2018.
• [10] G. Bingöl, A. Floris, S. Porcu, C. Timmerer and L. Atzori, “Are Quality and Sustainability Reconcilable? A Subjective Study on Video QoE, Luminance and Resolution,” in 15th International Conference on Quality of Multimedia Experience (QoMEX), Gent, Belgium, 2023.
• [11] G. Bingol, L. Serreli, S. Porcu, A. Floris, L. Atzori, “The Impact of Network Impairments on the QoE of WebRTC applications: A Subjective study,” in 14th International Conference on Quality of Multimedia Experience (QoMEX), Lippstadt, Germany, 2022.
• [12] D. Z. Rodríguez, R. L. Rosa, E. C. Alfaia, J. I. Abrahão and G. Bressan, “Video quality metric for streaming service using DASH standard,” IEEE Trans. Broadcasting, vol. vol. 62, no. 3, pp. 628-639, Sep. 2016.
• [13] T. Hoßfeld, M. Seufert, C. Sieber and T. Zinner, “Assessing effect sizes of influence factors towards a QoE model for HTTP adaptive streaming,” in 6th Int. Workshop Qual. Multimedia Exper. (QoMEX), Sep. 2014.
• [14] S. Afzal, R. Prodan, C. Timmerer, “Green Video Streaming: Challenges and Opportunities.” ACM SIGMultimedia Records, Jan. 2023.

In this column, we reflect on the environmental impact and broader sustainability implications of resource-demanding digital applications and services such as video streaming, VR/AR/XR and videoconferencing. We put emphasis not only on the experiences and use cases they enable but also on the “cost” of always striving for high Quality of Experience (QoE) and better user experiences. Starting by sketching the broader context, our aim is to raise awareness about the role that QoE research can play in the context of various of the United Nations’ Sustainable Development Goals (SDGs), either directly (e.g., SDG 13 “climate action”) or more indirectly (e.g., SDG 3 “good health and well-being” and SDG 12 “responsible consumption and production”).

The ambivalent role of digital technology

One of the latest reports from the Intergovernmental Panel on Climate Change (IPCC) confirmed the urgency of drastically reducing emissions of carbon dioxide and other human-induced greenhouse gas (GHG) emissions in the years to come (IPCC, 2021). This report, directly relevant in the context of SDG 13 “climate action”, confirmed the undeniable and negative human influence on global warming and the need for collective action. While the potential of digital technology (and ICT more broadly) for sustainable development has been on the agenda for some time, the context of the COVID-19 pandemic has made it possible to better understand a set of related opportunities and challenges.

First of all, it has been observed that long-lasting lockdowns and restrictions due to the COVID-19 pandemic and its aftermath have triggered a drastic increase in internet traffic (see e.g., Feldmann, 2020). This holds particularly for the use of videoconferencing and video streaming services for various purposes (e.g., work meetings, conferences, remote education, and social gatherings, just to name a few). At the same time, the associated drastic reduction of global air traffic and other types of traffic (e.g., road traffic) with their known environmental footprint, has had undeniable positive effects on the environment (e.g., reduced air pollution, better water quality see e.g., Khan et al., 2020). Despite this potential, the environmental gains enabled by digital technology and recent advances in energy efficiency are threatened by digital rebound effects due to increased energy consumption and energy demands related to ICT (Coroamua et al., 2019; Lange et al., 2020). In the context of ever-increasing consumption, there has for instance been a growing focus in the literature on the negative environmental impact of unsustainable use and viewing practices such as binge-watching, multi-watching and media-multitasking, which have become more common over the last years (see e.g., Widdicks, 2019). While it is important to recognize that the overall emission factor will vary depending on the mix of energy generation technologies used and region in the world (Preist et al., 2014), the above observation also fits with other recent reports and articles, which expect the energy demands linked to digital infrastructure, digital services and their use to further expand and which expect the greenhouse gas emissions of ICT relative to the overall worldwide footprint to significantly increase (see e.g., Belkhir et al., 2018, Morley et al., 2018, Obringer et al., 2021). Hence, these and other recent forecasts show a growing and even unsustainable high carbon footprint of ICT in the middle-term future, due to, among others, the increasing energy demand of data centres (including e.g., also the energy needed for cooling) and the associated traffic (Preist et al., 2016).

Another set of challenges that became more apparent can be linked to the human mental resources and health involved as well as environmental effects. Here, there is a link to the abovementioned Sustainable development goals 3 (good health and well-being) and 12 (sustainable consumption and production). For instance, the transition to “more sustainable” digital meetings, online conferences, and online education has also pointed to a range of challenges from a user point of view.  “Zoom fatigue” being a prominent example illustrates the need to strike the right balance between the more sustainable character of experiences provided by and enabled through technology and how these are actually experienced and perceived from a user point of view (Döring et al., 2022; Raake et al., 2022). Another example is binge-watching behavior, which can in certain cases have a positive effect on an individual’s well-being, but has also been shown to have a negative effect through e.g., feelings of guilt and goal conflicts  (Granow et al., 2018) or through problematic involvement resulting in e.g., chronic sleep issues  (Flayelle, 2020).

From the “production” perspective, recent work has looked at the growing environmental impact of commonly used cloud-based services such as video streaming (see e.g., Chen et al., 2020, Suski et al., 2020, The Shift Project, 2021) and the underlying infrastructure consisting of data centers, transport network and end devices (Preist et al., 2016, Suski, 2020, Preist et al., 2014). As a result, the combination of technological advancements and user-centered approaches aiming to always improve the experience may have undesired environmental consequences. This includes stimulating increased user expectations (e.g., higher video quality, increased connectivity and availability, almost zero-latency, …) and by triggering increased use, and unsustainable use practices, resulting in potential rebound effects due to increased data traffic and electricity demand. 

These observations have started to culminate into a plea for a shift towards a more sustainable and humanity-centered paradigm, which considers to a much larger extent how digital consumption and increased data demand impact individuals, society and our planet (Widdicks et al., 2019, Priest et al., 2016, Hazas & Nathan, 2018). Here, it is obvious that experience, consumption behavior and energy consumption are tightly intertwined.

How does QoE research fit into this picture?

This leads to the question of where research on Quality of Experience and its underlying goals fit into this broader picture, to which extent related topics have gained attention so far and how future research can potentially have an even larger impact.

As the COVID-19 related examples above already indicated, QoE research, through its focus on improving the experience for users in e.g., various videoconferencing-based scenarios or immersive technology-related use cases, already plays and will continue to play a key role in enabling more sustainable practices in various domains (e.g., remote education, online conferences, digital meetings, and thus reducing unnecessary travels, …) and linking up to various SDGs. A key challenge here is to enable experiences that become so natural and attractive that they may even become preferred in the future. While this is a huge and important topic, we refrain from discussing it further in this contribution, as it already is a key focus within the QoE community. Instead, in the following, we, first of all, reflect on the extent to which environmental implications of multimedia services have explicitly been on the agenda of the QoE community in the past, what the focus is in more recent work, and what is currently not yet sufficiently addressed. Secondly, we consider a broader set of areas and concrete topics in which QoE research can be related to environmental and broader sustainability-related concerns.

Traditionally, QoE research has predominantly focused on gathering insights that can guide the optimization of technical parameters and allocation of resources at different layers, while still ensuring a high QoE from a user point of view. A main underlying driver in this respect has traditionally been the related business perspective: optimizing QoE as a way to increase profitability and users/customers’ willingness to pay for better quality  (Wechsung, 2014). While better video compression techniques or adaptive video streaming may allow the saving of resources, which overall may lead to environmental gains, the latter has traditionally not been a main or explicit motivation.

There are however some exceptions in earlier work, where the focus was more explicitly on the link between energy consumption-related aspects, energy efficiency and QoE. The study of Ickin, 2012 for instance, aimed to investigate QoE influence factors of mobile applications and revealed the key role of the battery in successful QoE provisioning. In this work, it was observed that energy modelling and saving efforts are typically geared towards the immediate benefits of end users, while less attention was paid to the digital infrastructure (Popescu, 2018). Efforts were further also made in the past to describe, analyze and model the trade-off between QoE and energy consumption (QoE perceived per user per Joule, QoEJ) (Popescu, 2018) or power consumption (QoE perceived per user per Watt, QoEW) (Zhang et al., 2013), as well as to optimize resource consumption so as to avoid sources of annoyance (see e.g., (Fiedler et al., 2016). While these early efforts did not yet result in a generic end-to-end QoE-energy-model that can be used as a basis for optimizations, they provide a useful basis to build upon.

A more recent example (Hossfeld et al., 2022) in the context of video streaming services looked into possible trade-offs between varying levels of QoE and the resulting energy consumption, which is further mapped to CO₂ emissions (taking the EU emission parameter as input, as this – as mentioned – depends on the overall energy mix of green and non-renewable energy sources). Their visualization model further considers parameters such as the type of device and type of network and while it is a simplification of the multitude of possible scenarios and factors, it illustrates that it is possible to identify areas where energy consumption can be reduced while ensuring an acceptable QoE.

Another recent work (Herglotz et al., 2022) jointly analyzed end-user power efficiency and QoE related to video streaming, based on actual real-world data (i.e., YouTube streaming events). More specifically, power consumption was modelled and user-perceived QoE was estimated in order to model where optimization is possible. They found that optimization is possible and pointed to the importance of the choice of video codec, video resolution, frame rate and bitrate in this respect.

These examples point to the potential to optimize at the “production” side, however, the focus has more recently also been extended to the actual use, user expectations and “consumption” side (Jiang et al., 2021, Lange et al., 2020, Suski et al., 2020, Elgaaied-Gambier et al., 2020) Various topics are explored in this respect, e.g., digital carbon footprint calculation at the individual level (Schien et al., 2013, Preist et al., 2014), consumer awareness and pro-environmental digital habits (Elgaaied-Gambier et al., 2020; Gnanasekaran et al., 2021), or impact of user behavior (Suski et al., 2020). While we cannot discuss all of these in detail here, they all are based on the observation that there is a growing need to involve consumers and users in the collective challenge of reducing the impact of digital applications and services on the environment (Elgaaied-Gambier et al., 2020; Priest et al., 2016).

QoE research can play an important role here, extending the understanding of carbon footprint vs. QoE trade-offs to making users more aware of the actual “cost” of high QoE. A recent interview study with digital natives conducted by some of the co-authors of this column  (Gnanasekaran et al., 2021) illustrated that many users are not aware of the environmental impact of their user behavior and expectations and that even with such insights, substantial drastic changes in behavior cannot be expected. The lack of technological understanding, public information and social awareness about the topic were identified as important factors. It is therefore of utmost importance to trigger more awareness and help users see and understand their carbon footprint related to e.g., the use of video streaming services (Gnanasekaran et al., 2021). This perspective is currently missing in the field of QoE and we argue here that QoE research could – in collaboration with other disciplines and by integrating insights from other fields – play an important role here.

In terms of the motivation for adopting pro-environmental digital habits, Gnanasekaran et al., (2021) found that several factors indirectly contribute to this goal, including the striving for personal well-being. Finally, the results indicate some willingness to change and make compromises (e.g., accepting a lower video quality), albeit not an unconditional one: the alignment with other goals (e.g., personal well-being) and the nature of the perceived sacrifice and its impact play a key role. A key challenge for future work is therefore to identify and understand concrete mechanisms that could trigger more awareness amongst users about the environmental and well-being impact of their use of digital applications and services, and those that can further motivate positive behavioral change (e.g., opting for use practices that limit one’s digital carbon footprint, mindful digital consumption). By investigating the impact of various more environmentally-friendly viewing practices on QoE (e.g., actively promoting standard definition video quality instead of HD, nudging users to switch to audio-only when a service like YouTube is used as background noise or stimulating users to switch to the least data demanding viewing configuration depending on the context and purpose), QoE research could help to bridge the gap towards actual behavioral change.

Final reflections and challenges for future research

We have argued that research on users’ Quality of Experience and overall User Experience can be highly relevant to gain insights that may further drive the adoption of new, more sustainable usage patterns and that can trigger more awareness of implications of user expectations, preferences and actual use of digital services. However, the focus on continuously improving users’ Quality Experience may also trigger unwanted rebound effects, leading to an overall higher environmental footprint due to the increased use of digital applications and services. Further, it may have a negative impact on users’ long-term well-being as well.

We, therefore, need to join efforts with other communities to challenge the current design paradigm from a more critical stance, partly as “it’s difficult to see the ecological impact of IT when its benefits are so blindingly bright” (Borning et al., 2020). Richer and better experiences may lead to increased, unnecessary or even excessive consumption, further increasing individuals’ environmental impact and potentially impeding long-term well-being. Open questions are, therefore: Which fields and disciplines should join forces to mitigate the above risks? And how can QoE research — directly or indirectly — contribute to the triggering of sustainable consumption patterns and the fostering of well-being?

Further, a key question is how energy efficiency can be improved for digital services such as video streaming, videoconferencing, online gaming, etc., while still ensuring an acceptable QoE. This also points to the question of which compromises can be made in trading QoE against its environmental impact (from “willingness to pay” to “willingness to sacrifice”), under which circumstances and how these compromises can be meaningfully and realistically assessed. In this respect, future work should extend the current modelling efforts to link QoE and carbon footprint, go beyond exploring what users are willing to (more passively) endure, and also investigate how users can be more actively motivated to adjust and lower their expectations and even change their behavior.

These and related topics will be on the agenda of the Dagstuhl seminar  23042 “Quality of Sustainable Experience” and the conference QoMEX 2023 “Towards sustainable and inclusive multimedia experiences”.

References

Belkhir, L., Elmeligi, A. (2018). “Assessing ICT global emissions footprint: Trends to 2040 & recommendations,” Journal of cleaner production, vol. 177, pp. 448–463.

Borning, A., Friedman, B., Logler, N. (2020). The ’invisible’ materiality of information technology. Communications of the ACM, 63(6), 57–64.

Chen, X., Tan, T., et al. (2020). Context-Aware and Energy-Aware Video Streaming on Smartphones. IEEE Transactions on Mobile Computing.

Coroama, V.C., Mattern, F. (2019). Digital rebound–why digitalization will not redeem us our environmental sins. In: Proceedings 6th international conference on ICT for sustainability. Lappeenranta. http://ceur-ws.org. vol. 238

Döring, N., De Moor, K., Fiedler, M., Schoenenberg, K., Raake, A. (2022). Videoconference Fatigue: A Conceptual Analysis. Int. J. Environ. Res. Public Health, 19(4), 2061 https://doi.org/10.3390/ijerph19042061

Elgaaied-Gambier, L., Bertrandias, L., Bernard, Y. (2020). Cutting the internet’s environmental footprint: An analysis of consumers’ self-attribution of responsibility. Journal of Interactive Marketing, 50, 120–135.

Feldmann, A., Gasser, O., Lichtblau, F., Pujol, E., Poese, I., Dietzel, C., … & Smaragdakis, G. (2020, October). The lockdown effect: Implications of the COVID-19 pandemic on internet traffic. In Proceedings of the ACM internet measurement conference (pp. 1-18).

Daniel Wagner, Matthias Wichtlhuber, Juan Tapiador, Narseo Vallina-Rodriguez, Oliver Hohlfeld, and Georgios Smaragdakis.

Feldmann, A., Gasser, O., Lichtblau, F., Pujol, E., Poese, I., Dietzel, C., Wagner, D., Wichtlhuber, M., Tapiador, J., Vallina-Rodriguez, N., Hohlfeld, O., Smaragdakis, G. (2020, October). The lockdown effect: Implications of the COVID-19 pandemic on internet traffic. In Proceedings of the ACM internet measurement conference (pp. 1-18).

Fiedler, M., Popescu, A., Yao, Y. (2016), “QoE-aware sustainable throughput for energy-efficient video streaming,” in 2016 IEEE International Conferences on Big Data and Cloud Computing (BDCloud), Social Computing and Networking (SocialCom), Sustainable Computing and Communications (SustainCom)(BDCloud-SocialCom-SustainCom). pp. 493–50

Flayelle, M., Maurage, P., Di Lorenzo, K.R., Vögele, C., Gainsbury, S.M., Billieux, J. (2020). Binge-Watching: What Do we Know So Far? A First Systematic Review of the Evidence. Curr Addict Rep 7, 44–60. https://doi.org/10.1007/s40429-020-00299-8

Gnanasekaran, V., Fridtun, H. T., Hatlen, H., Langøy, M. M., Syrstad, A., Subramanian, S., & De Moor, K. (2021). Digital carbon footprint awareness among digital natives: an exploratory study. In Norsk IKT-konferanse for forskning og utdanning (No. 1, pp. 99-112).

Granow, V.C., Reinecke, L., Ziegele, M. (2018): Binge-watching and psychological well-being: media use between lack of control and perceived autonomy. Communication Research Reports 35 (5), 392–401.

Hazas, M. and Nathan, L. (Eds.)(2018). Digital Technology and Sustainability. London: Routledge.

Herglotz, C., Springer, D., Reichenbach,  M., Stabernack B. and Kaup, A. (2018). “Modeling the Energy Consumption of the HEVC Decoding Process,” in IEEE Transactions on Circuits and Systems for Video Technology, vol. 28, no. 1, pp. 217-229, Jan. 2018, doi: 10.1109/TCSVT.2016.2598705.

Hossfeld, T., Varela, M., Skorin-Kapov, L. Heegaard, P.E. (2022). What is the trade-off between CO2 emission and videoconferencing QoE. ACM SIGMM records, https://records.sigmm.org/2022/03/31/what-is-the-trade-off-between-co2-emission-and-video-conferencing-qoe/

Ickin, S., Wac, K., Fiedler, M. and Janowski, L. (2012). “Factors influencing quality of experience of commonly used mobile applications,” IEEE Communications Magazine, vol. 50, no. 4, pp. 48–56.

IPCC, 2021: Climate Change 2021: The Physical Science Basis. Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change [Masson-Delmotte, V., P. Zhai, A. Pirani, S.L. Connors, C. Péan, S. Berger, N. Caud, Y. Chen, L. Goldfarb, M.I. Gomis, M. Huang, K. Leitzell, E. Lonnoy, J.B.R. Matthews, T.K. Maycock, T. Waterfield, O. Yelekçi, R. Yu, and B. Zhou (eds.)]. Cambridge University Press, Cambridge, United Kingdom and New York, NY, USA, In press, doi:10.1017/9781009157896.

Jiang, P., Van Fan, Y., Klemes, J.J. (2021). Impacts of covid-19 on energy demand and consumption: Challenges, lessons and emerging opportunities. Applied energy, 285, 116441.

Khan, D., Shah, D. and Shah, S.S. (2020). “COVID-19 pandemic and its positive impacts on environment: an updated review,” International Journal of Environmental Science and Technology, pp. 1–10, 2020.

Lange, S., Pohl, J., Santarius, T. (2020). Digitalization and energy consumption. Does ICT reduce energy demand? Ecological Economics, 176, 106760.

Morley, J., Widdicks, K., Hazas, M. (2018). Digitalisation, energy and data demand: The impact of Internet traffic on overall and peak electricity consumption. Energy Research & Social Science, 38, 128–137.

Obringer, R., Rachunok, B., Maia-Silva, D., Arbabzadeh, M., Roshanak, N., Madani, K. (2021). The overlooked environmental footprint of increasing internet use. Resources, Conservation and Recycling, 167, 105389.

Popescu, A. (Ed.)(2018). Greening Video Distribution Networks, Springer.

Preist, C., Schien, D., Blevis, E. (2016). “Understanding and mitigating the effects of device and cloud service design decisions on the environmental footprint of digital infrastructure,” in Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems, pp. 1324–1337.

Preist, C., Schien, D., Shabajee, P. , Wood, S. and Hodgson, C. (2014). “Analyzing End-to-End Energy Consumption for Digital Services,” Computer, vol. 47, no. 5, pp. 92–95.

Raake, A., Fiedler, M., Schoenenberg, K., De Moor, K., Döring, N. (2022). Technological Factors Influencing Videoconferencing and Zoom Fatigue. arXiv:2202.01740, https://doi.org/10.48550/arXiv.2202.01740

Schien, D., Shabajee, P., Yearworth, M. and Preist, C. (2013), Modeling and Assessing Variability in Energy Consumption During the Use Stage of Online Multimedia Services. Journal of Industrial Ecology, 17: 800-813. https://doi.org/10.1111/jiec.12065

Suski, P., Pohl, J., Frick, V. (2020). All you can stream: Investigating the role of user behavior for greenhouse gas intensity of video streaming. In: Proceedings of the 7th International Conference on ICT for Sustainability. p. 128–138. ICT4S2020, Association for Computing Machinery, New York, NY, USA.

The Shift Project, Climate crisis: the unsustainable use of online video: Our new report on the environmental impact of ICT. https://theshiftproject.org/en/article/unsustainable-use-online-video/

Wechsung, I., De Moor, K. (2014). Quality of Experience Versus User Experience. In: Möller, S., Raake, A. (eds) Quality of Experience. T-Labs Series in Telecommunication Services. Springer, Cham. https://doi.org/10.1007/978-3-319-02681-7_3

Widdicks, K., Hazas, M., Bates, O., Friday, A. (2019). “Streaming, Multi-Screens and YouTube: The New (Unsustainable) Ways of Watching in the Home,” in Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, ser. CHI ’19. New York, NY, USA: Association for Computing Machinery, p. 1–13.

Zhang, X., Zhang, J., Huang, Y., Wang, W. (2013). “On the study of fundamental trade-offs between QoE and energy efficiency in wireless networks,” Transactions on Emerging Telecommunications Technologies, vol. 24, no. 3, pp. 259–265.

It is a natural thing that users of multimedia services want to have the highest possible Quality of Experience (QoE), when using said services. This is especially so in contexts such as video-conferencing and video streaming services, which are nowadays a large part of many users’ daily life, be it work-related Zoom calls, or relaxing while watching Netflix. This has implications in terms of the energy consumed for the provision of those services (think of the cloud services involved, the networks, and the users’ own devices), and therefore it also has an impact on the resulting CO₂ emissions. In this column, we look at the potential trade-offs involved between varying levels of QoE (which for video services is strongly correlated with the bit rates used), and the resulting CO₂ emissions. We also look at other factors that should be taken into account when making decisions based on these calculations, in order to provide a more holistic view of the environmental impact of these types of services, and whether they do have a significant impact.

Energy Consumption and CO2 Emissions for Internet Service Delivery

Understanding the footprint of Internet service delivery is a challenging task. On one hand, the infrastructure and software components involved in the service delivery need to be known. For a very fine-grained model, this requires knowledge of all components along the entire service delivery chain: end-user devices, fixed or mobile access network, core network, data center and Internet service infrastructure. Furthermore, the footprint may need to consider the CO₂ emissions for producing and manufacturing the hardware components as well as the CO₂ emissions during runtime. Life cycle assessment is then necessary to obtain CO₂ emission per year for hardware production. However, one may argue that the infrastructure is already there and therefore the focus will be on the energy consumption and CO₂ emission during runtime and delivery of the services. This is also the approach we follow here to provide quantitative numbers of energy consumption and CO₂ emission for Internet-based video services. On the other hand, quantitative numbers are needed beyond the complexity of understanding and modelling the contributors to energy consumption and C02 emission.

To overcome this complexity, the literature typically considers key figures on the overall data traffic and service consumption times aggregated over users and services over a longer period of time, e.g., one year. In addition, the total energy consumption of mobile operators and data centres is considered. Together with the information on e.g., the number of base station sites, this gives some estimates, e.g., on the average power consumption per site or the average data traffic per base station site [Feh11]. As a result, we obtain measures such as energy per bit (Joule/bit) determining the energy efficiency of a network segment. In [Yan19], the annual energy consumption of Akamai is converted to power consumption and then divided by the maximum network traffic, which results again in the energy consumption per bit of Akamai’s data centers. Knowing the share of energy sources (nonrenewable energy, including coal, natural gas, oil, diesel, petroleum; renewable energy including solar, geothermal, wind energy, biomass, hydropower from flowing water), allows relating the energy consumption to the total CO₂ emissions. For example, the total contribution from renewables exceeded 40% in 2021 in Germany and Finland, Norway has about 60%, Croatia about 36% (statistics from 2020).

A detailed model of the total energy consumption of mobile network services and applications is provided in [Yan19]. Their model structure considers important factors from each network segment from cloud to core network, mobile network, and end-user devices. Furthermore, service-specific energy consumption are provided. They found that there are strong differences between the service type and the emerging data traffic pattern. However, key factors are the amount of data traffic and the duration of the services. They also consider different end-to-end network topologies (user-to-data center, user-to-user via data center, user-to-user and P2P communication). Their model of the total energy consumption is expressed as the sum of the energy consumption of the different segments:

• Smartphone: service-specific energy depends among others on the CPU usage and the network usage e.g. 4G over the duration of use,
• Base station and access network: data traffic and signalling traffic over the duration of use,
• Wireline core network: service specific energy consumption of a mobile service taking into account the data traffic volume and the energy per bit,
• Data center: energy per bit of the data center is multiplied by data traffic volume of the mobile service.

The Shift Project [TSP19] provides a similar model which is called the “1 Byte Model”. The computation of energy consumption is transparently provided in calculation sheets and discussed by the scientific community. As a result of the discussions [Kam20a,Kam20b], an updated model was released [TSP20] clarifying a simple bit/byte conversion issue. The suggested models in [TSP20, Kam20b] finally lead to comparable numbers in terms of energy consumption and CO₂ emission. As a side remark: Transparency and reproducibility are key for developing such complex models!

The basic idea of the 1 Byte Model for computing energy consumption is to take into account the time t of Internet service usage and the overall data volume v. The time of use is directly related to the energy consumption of the display of an end-user device, but also for allocating network resources. The data volume to transmit through the network, but also to generate or process data for cloud services, drives the energy consumption additionally. The model does not differentiate between Internet services, but they will result in different traffic volumes over the time of use. Then, for each segment i (device, network, cloud) a linear model E_i(t,v)=a_i * t + b_i * v + c_i is provided to quantify the energy consumption. To be more precise, the different coefficients are provided for each segment by [TSP20]. The overall energy consumption is then E_total = E_device + E_network + E_cloud.

CO₂ emission is then again a linear model of the total energy consumption (over the time of use of a service), which depends on the share of nonrenewable and renewable energies. Again, The Shift Project derives such coefficients for different countries and we finally obtain CO2 = k_country * E_total.

The Trade-off between QoE and CO2 Emissions

As a use case, we consider hosting a scientific conference online through video-conferencing services. Assume there are 200 conference participants attending the video-conferencing session. The conference lasts for one week, with 6 hours of online program per day.  The video conference software requires the following data rates for streaming the sessions (video including audio and screen sharing):

• high-quality video: 1.0 Mbps
• 720p HD video: 1.5 Mbps
• 1080p HD video: 3 Mbps

However, group video calls require even higher bandwidth consumption. To make such experiences more immersive, even higher bit rates may be necessary, for instance, if using VR systems for attendance.

A simple QoE model may map the video bit rate of the current video session to a mean opinion score (MOS). [Lop18] provides a logistic regression MOS(x) depending on the video bit rate x in Mbps: f(x) = m_1 log x + m_2

Then, we can connect the QoE model with the energy consumption and CO₂ emissions model from above in the following way. We assume a user attending the conference for time t. With a video bit rate x, the emerging data traffic is v = x*t. Those input parameters are now used in the 1 Byte Model for a particular device (laptop, smartphone), type of network (wired, wifi, mobile), and country (EU, US, China).

Figure 1 shows the trade-off between the MOS and energy consumption (left y-axis). The energy consumption is mapped to CO₂ emission by assuming the corresponding parameter for the EU, and that the conference participants are all connected with a laptop. It can be seen that there is a strong increase in energy consumption and CO₂ emission in order to reach the best possible QoE. The MOS score of 4.75 is reached if a video bit rate of roughly 11 Mbps is used. However, with 4.5 Mbps, a MOS score of 4 is already reached according to that logarithmic model. This logarithmic behaviour is a typical observation in QoE and is connected to the Weber-Fechner law, see [Rei10]. As a consequence, we may significantly save energy and CO₂ when not providing the maximum QoE, but “only” good quality (i.e., MOS score of 4). The meaning of the MOS ratings is 5=Excellent, 4=Good, 3=Fair, 2=Poor, 1=Bad quality.

Figure 2, therefore, visualized the gain when delivering the video in lower quality and lower video bit rates. In fact, the gain compared to the efforts for MOS 5 are visualized. To get a better understanding of the meaning of those CO₂ numbers, we express the CO₂ gain now in terms of thousands of kilometers driving by car. Since the CO₂ emission depends on the share of renewable energies, we may consider different countries and the parameters as provided in [TSP20]. We see that ensuring each conference participant a MOS score of 4 instead of MOS 5 results in savings corresponding to driving approximately 40000 kilometers by car assuming the renewable energy share in the EU – this is the distance around the Earth! Assuming the energy share in China, this would save more than 90000 kilometers. Of course, you could also save 90 000 kilometers by walking – which requires however about 2 years non-stop with a speed of 5 km/h. Note that this large amount of CO₂ emission is calculated assuming a data rate of 15 Mbps over 5 days (and 6 hours per day), resulting in about 40.5 TB of data that needs to be transferred to the 200 conference participants.

Discussions

Raising awareness of CO₂ emissions due to Internet service consumption is crucial. The abstract CO₂ emission numbers may be difficult to understand, but relating this to more common quantities helps to understand the impact individuals have. Of course, the provided numbers only give an impression, since the models are very simple and do not take into account various facets. However, the numbers nicely demonstrate the potential trade-off between QoE of end-users and sustainability in terms of energy consumption and CO₂ emission. In fact, [Gna21] conducted qualitative interviews and found that there is a lack of awareness of the environmental impact of digital applications and services, even for digital natives. In particular, an underlying issue is that there is a lack of understanding among end-users as to how Internet service delivery works, which infrastructure components play a role and are included along the end-to-end service delivery path, etc. Hence, the environmental impact is unclear for many users. Our aim is thus to contribute to overcoming this issue by raising awareness on this matter, starting with simplified models and visualizations.

[Gna21] also found that users indicate a certain willingness to make compromises between their digital habits and the environmental footprint. Given global climate changes and increased environmental awareness among the general population, such a trend in willingness to make compromises may be expected to further increase in the near future. Hence, it may be interesting for service providers to empower users to decide their environmental footprint at the cost of lower (yet still satisfactory) quality. This will also reduce the costs for operators and seems to be a win-win situation if properly implemented in Internet services and user interfaces.

Nevertheless, tremendous efforts are also currently being undertaken by Internet companies to become CO₂ neutral in the future. For example, Netflix claims in [Netflix2021] that they plan to achieve net-zero greenhouse gas emissions by the close of 2022. Similarly, also economic, societal, and environmental sustainability is seen as a key driver for 6G research and development [Mat21]. However, the time horizon is on a longer scope, e.g., a German provider claims they will reach climate neutrality for in-house emissions by 2025 at the latest and net-zero from production to the customer by 2040 at the latest [DT21]. Hence, given the urgency of the matter, end-users and all stakeholders along the service delivery chain can significantly contribute to speeding up the process of ultimately achieving net-zero greenhouse gas emissions.

References

• [TSP19] The Shift Project, “Lean ict: Towards digital sobriety,” directed by Hugues Ferreboeuf, Tech. Rep., 2019, last accessed: March 2022. Available online (last accessed: March 2022)
• [Yan19] M. Yan, C. A. Chan, A. F. Gygax, J. Yan, L. Campbell,A. Nirmalathas, and C. Leckie, “Modeling the total energy consumption of mobile network services and applications,” Energies, vol. 12, no. 1, p. 184, 2019.
• [TSP20] Maxime Efoui Hess and Jean-Noël Geist, “Did The Shift Project really overestimate the carbon footprint of online video? Our analysis of the IEA and Carbonbrief articles”, The Shift Project website, June 2020, available online (last accessed: March 2022) PDF
• [Kam20a] George Kamiya, “Factcheck: What is the carbon footprint of streaming video on Netflix?”, CarbonBrief website, February 2020. Available online (last accessed: March 2022)
• [Kam20b] George Kamiya, “The carbon footprint of streaming video: fact-checking the headlines”, IEA website, December 2020. Available online (last accessed: March 2022)
• [Feh11] Fehske, A., Fettweis, G., Malmodin, J., & Biczok, G. (2011). The global footprint of mobile communications: The ecological and economic perspective. IEEE communications magazine, 49(8), 55-62.
• [Lop18]  J. P. López, D. Martín, D. Jiménez, and J. M. Menéndez, “Prediction and modeling for no-reference video quality assessment based on machine learning,” in 2018 14th International Conference on Signal-Image Technology & Internet-Based Systems (SITIS), IEEE, 2018, pp. 56–63.
• [Gna21] Gnanasekaran, V., Fridtun, H. T., Hatlen, H., Langøy, M. M., Syrstad, A., Subramanian, S., & De Moor, K. (2021, November). Digital carbon footprint awareness among digital natives: an exploratory study. In Norsk IKT-konferanse for forskning og utdanning (No. 1, pp. 99-112).
• [Rei10] Reichl, P., Egger, S., Schatz, R., & D’Alconzo, A. (2010, May). The logarithmic nature of QoE and the role of the Weber-Fechner law in QoE assessment. In 2010 IEEE International Conference on Communications (pp. 1-5). IEEE.
• [Netflix21] Netflix: “Environmental Social Governance 2020”,  Sustainability Accounting Standards Board (SASB) Report, (2021, March). Available online (last accessed: March 2022)
• [Mat21] Matinmikko-Blue, M., Yrjölä, S., Ahokangas, P., Ojutkangas, K., & Rossi, E. (2021). 6G and the UN SDGs: Where is the Connection?. Wireless Personal Communications, 121(2), 1339-1360.
• [DT21] Hannah Schauff. Deutsche Telekom tightens its climate targets (2021, January). Available online (last accessed: March 2022)