Cross-Domain Attentive Sequential Recommendations based on General and Current User Preferences (CD-ASR)

Abstract

Sequential Recommendations (SR) have become increasingly important because of their accuracy and consistency with real world scenarios, where a user interacts with a sequence of items over time. SR systems have the capability of modeling temporal information to extract user’s current preferences. However, data sparsity is a real challenge for SR models, causing them to sometimes function poorly and generate inaccurate recommendations. A practical solution to this problem is to transfer information from multiple source domains to tackle the sparsity in the target domain, an approach known as Cross-Domain Recommendations (CDR). Extracting users’ preferences from different domains in a sequential manner can help generate an effective CDR for sequential models. In this paper, we propose a Cross-Domain Attentive Sequential Recommendation model based on general and current user preferences (CD-ASR). We assume the user information from the source domains to be a user’s general information, and apply a general attention model to aggregate user source representational vectors. At the same time, we apply a self-attention sequential model to obtain user’s current preferences in the target domain. Implicitly, we utilize the aggregated user’s source vectors to transfer knowledge to produce more precise recommendation in the target domain. Our model shows superiority over other state-of-the-art SR models using three datasets extracted from Amazon dataset.