To improve the user engagement, especially under moderate to high traffic demand, it is important to understand the impact of the network and application QoS on user experience. This article comparatively evaluates the impact of impairments, with emphasis on rebufferings, startup delay, and bitrate changes, and their intensity and temporal dynamics, on user engagement in the context of video streaming. The analysis employed two large YouTube datasets. To characterize the user engagement and the impact of impairments, several new metrics were defined. We assessed whether or not there is a statistically significant relationship between different types of impairments and user engagement metrics, taking into account not only the characteristics of the impairments but also the covariates of the session (e.g., video duration, mean data rate). After observing the relationships across the entire dataset, we tested whether these relationships also persist under specific conditions with respect to the covariates. The introduction of several new metrics and of various covariates in the analysis are two innovative aspects of this work. We found that the presence of negative bitrate changes (BR-) is a stronger predictor of abandonment than rebufferrings (RB). Positive bitrate changes (BR+) in low resolution sessions are not well received. High rebufferring ratio has a prominent impact on the video watching percentage. These results can be used to guide the video streaming adaptation as well as suggest which parameters should be varied in controlled field studies.

To improve the quality of experience (QoE), especially under moderate to high traffic demand, it is important to understand the impact of the network and application QoS on user experience. This paper comparatively evaluates the impact of impairments, their intensity and temporal dynamics, on user-engagement in the context of video streaming. The analysis employed two large YouTube datasets. To characterize the user engagement and the impact of impairments, several new metrics were defined. We assessed whether or not there is a statistically significant relationship between different types of impairments and QoE and user engagement metrics, taking into account not only the characteristics of the impairments but also the covariates of the session (e.g., video duration, mean datarate). After observing the relationships across the entire dataset, we tested whether these relationships also persist under specific conditions with respect to the covariates. The introduction of several new metrics and of various covariates in the analysis are two innovative aspects of this work. We found that the number of negative bitrate changes (BR-) is a stronger predictor of abandonment than rebufferrings (RB). Even positive bitrate changes (BR+) are associated with increases in abandonment. Specifically, BR+ in low resolution sessions is not well-received. Temporal dynamics of the impairments have also an impact: a BR-that follows much later a RB appears to be perceived as a worse impairment than a BR-that occurs immediately after a RB. These results can be used to guide the design of the video streaming adaptation as well as suggest which parameters should be varied in controlled field studies.