سيستم توصيه‌گر مشاركتي افزايشي مبتني بر جلسه ناشناس

Many e-commerce platforms use session-based recommender systems (SBRS) to increase their product sales an‎d guide customers to find their favorite items. In recent years, graph neural networks (GNNs) based SBRS have achieved good results due to their ability to capture complex relationships between items clicked in a session. A number of these models use side information (such as item category) to enhance their performance, but none of them have focused on item bran‎d information an‎d none have explored the potential for combined use of side information. In addition, some of these models exploit collaborative information available in similar/neighboring sessions. They calculate session similarity by counting the number of shared items between two sessions an‎d do not consider other features such as the order of shared items. On the other han‎d, most of these models are trained offline only once an‎d create a static representation of users’ long-term preferences. In this paper, a model named Streaming Collaborative Graph Neural Networks (SC-GNN) for SBRS is presented, which combines two types of item side information (category an‎d bran‎d) to enrich the session data an‎d by improving the selec‎tion mechanism of neighboring sessions, increases the accuracy of the recommender. Also, in the case of user anonymity, it adds the aspect of streaming of input sessions so that after a period of user interactions, the updat‎e process of the recommender system is performed based on the sampling approach in the repository. Specifically, user interactions are modeled at two levels: (1) a global level where neighboring sessions are selec‎ted using a time- an‎d sequence-a`ware approach an‎d their collaborative information is extracted to improve the model performance; (2) a local level where the sequence of interactions in a session is transformed into a heterogeneous graph with three types of nodes including item, bran‎d, an‎d category an‎d nine different relationships between these nodes. The information from these two levels is then extracted an‎d aggregated using a graph convolution method. Experimental results on the Cosmetics an‎d Tmall datasets demonstrate the superiority of SC-GNN over the baseline models.

6