A deep bi-directional prediction model for live streaming recommendation

Abstract

Live streaming becomes very popular in recent years. An accurate live streaming recommendation system is key to enhance user experience. As both viewer and anchor change their preferences dynamically, using existing recommendation approaches cannot fully capture their preferences. In this paper, we study how to model both viewer and anchor’s dynamic behaviors and predict their behaviors from two angles to improve recommendation accuracy. Existing works mainly focus on prediction from one angle, neglecting the prediction from another angle. How to simultaneously model the predictions from two angles to enhance their mutual learning and make recommendation based on them are not well studied in literature. To solve this problem, we propose a deep bi-directional prediction model to perform two prediction tasks: predicting viewer’s next favorite anchor and prediction anchor’s future audience simultaneously, based on which different recommendation methods are also developed. In the model, we develop multiple mechanisms such as shared embedding layer, cross attention and cross loss to help the two prediction tasks to learn from each other. Experiments conducted on real world datasets demonstrate that the proposed model performs better than state-of-the-art recommendation models and prediction from the two sides improves recommendation performance.