A cross-domain hierarchical recurrent model for personalized session-based recommendations

Abstract Recently, much attention has been paid to personalized session-based recommendations, where detailed user information is available due to users’ automatic or active login. Nevertheless, most methods focus on a single-domain scenario, assuming that users are only active in a single domain. Consequently, they always suffer from lack of data due to ignoring the fact that users’ behaviors are scattered across domains. Therefore, we propose a novel model, called Cross-Domain Hierarchical Recurrent Model (CDHRM), to incorporate cross-domain sequential information by exploring correlations among users’ cross-domain behaviors. Specifically, we devise a cross-domain user-level recurrent neural network (RNN) to systematically depict users’ global interests by capturing the cross-domain inter-session dynamics. To separately capture intra-session dynamics of different domains, two domain-specific session-level RNNs, which can preserve behavioral differences, are constructed. Meanwhile, for achieving synchrony of interactions among domains, the user-level RNN exchanges information with different session-level RNNs in a chronological order. Moreover, fusion layers with different integration strategies are introduced to further capture behavioral differences. Finally, cross-domain user-level and session-level information are jointly exploited to predict users’ future behaviors. Empirical results show CDHRM outperforms the state-of-the-art methods on three cross-domain datasets and can work well even with non-overlapping and sparse item information across domains.