Twitch is a community driven live streaming video service which uses recommendations trained on implicit feedback data. While this data is available in large quantities, it is subject to various biases and shortcomings. In particular, Twitch watch history data is heavily affected by the unpredictable and irregular behavior of users and streams going online and offline. Two resulting issues are: accounting for negative examples caused by users and channels not being available simultaneously, and understanding the relative scale of importance of positive watch time feedback. In this paper, we propose two methods of loss weighting: one method to address the availability of channels, and another method to adjust for the preferences implied by differing amounts of minutes watched. We discuss the methods, offline experimentation, and the results of
adjusting evaluation metrics to fit the new loss weighting methods. Finally, we demonstrate success in a sitewide A/B test that increased recommended minutes watched by 7.9%.
Weighing dynamic availability and consumption for Twitch recommendations
2022
Research areas