YouTube Viewing Experience Modeling

• Activity Number: 28
• Type: Topic Contributed
• Date/Time: Sunday, August 4, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Jin Cao*+
• Companies: Bell Labs
• Keywords: Latent variables

Abstract

Using Youtube traffic data collected from a major US wireless provider, we are developing a model to characterize a user's YouTube viewing experience using predictors of network condition and user/video specific metrics. We use the ``download completion'' as a substitute metric for user experience. The network condition collected using WNG framework includes throughput, loss and delay. Using one week of YouTube data, we modeled ``download completion'' as a combination of two latent variables: ``satisfactory network condition'' and ``enjoyable video''. We fitted the above two components with the selected variables both semi-parametrically and parametrically. Using this model, we are able to determine the impact of network condition alone on viewing experience, while taking into account of user's interest and video quality. It also helps us to understand aspects of network condition that are vital for the user experience.

This is joint work with Tian Bu, Sining Chen, and Sudarshan Vasudevan.