An adaptive program recommendation system for multi-user sharing environment

More and more accounts or devices are shared by multiple users in video applications, which makes it difficult to provide recommendation service. Existing recommendation schemes overlook multiuser sharing scenarios, and they cannot make effective use of the mixed information generated by multi-user when exploring users' potential interests. To solve these problems, this paper proposes an adaptive program recommendation system for multi-user sharing environment. Specifically, we first design an offline periodic identification module by building multi-user features and periodically predicting target user in future sessions, which can separate the profile of target user from mixed log records. Subsequently, an online recommendation module with adaptive time-varying exploration strategy is constructed by jointly using personal information and multi-user social information provided by identification module. On one hand, to learn the dynamic changes in user-interest, a time-varying linear upper confidence bound (LinUCB) based on personal information is designed. On the other hand, to reduce the risk of exploration, a time-invariant LinUCB based on separated multi-user social information from one account/device is proposed to compute the quality scores of programs for each user, which is integrated into the time-varying LinUCB by cross-weighting strategy. Finally, experimental results validate the efficiency of the proposed scheme.