Social Attentive Network for Live Stream Recommendation

Live streaming platforms not only provide live videos but also allow social interactions between viewers via real-time chatting. However, none of existing research has studied the social impact for recommending live streams. In this work, we formulate a new personalized recommendation problem by factoring in both video and social contents (chats). Accordingly, we 1) design a new attention network ANSWER to identify viewers' attention on video and social contents, and 2) rank the channels based on the attentive features. We collect a real dataset from Twitch for evaluation. The experimental results manifest that ANSWER outperforms baselines by at least 26.6% in terms of NDCG@5.