Exploiting Category Information in Sequential Recommendation

In recent years, sequential recommender systems have been widely applied for service recommendations. However, most existing solutions do not take full advantage of one key factor that usually influences user behaviors: the category of services. It is necessary yet challenging to capture users' category preferences. Firstly, the complex inherent relationships that exist among categories are vital but difficult to mine and encode. Secondly, since interest preferences and category preferences are closely related, their dynamic evolution has to be studied simultaneously. To tackle the above challenges, we propose a novel Reciprocal Dual-Channel Network (RDCN) to capture users' comprehensive dynamic characteristics toward more accurate recommendations. For the former challenge, we devise a novel strategy to obtain the co-occurrence information of services and categories and jointly pre-train their embeddings. For the latter challenge, we design a Co-Guided Attention module and a Co-Guided GRU module to extract interest preferences and category preferences, respectively. Experimental results on three public datasets have demonstrated the necessity of exploiting the category information and the effectiveness of the proposed RDCN model.