关键词: 多属性信息, 属性关联, 会话推荐, 图神经网络, 注意力机制

Abstract: In the era of the Internet information explosion, the efficiency of information acquisition has dropped sharply, which leads to an issue of information overload. In this context, how to support users to obtain valuable information from massive data has become a hot social concern. Recommender systems, as one type of decision support systems, effectively alleviate the information overload problem and have been widely used in online service platforms, such as social medias, e-commerce websites, and so on. However, conventional recommendation methods rely on users’ long-term historical behaviors, which leads to the fact that the recommendation performance suffers when users’ identity information and historical behaviors are unavailable. To overcome this limitation, session-based recommendation models users’ short-term interests in real-time and analyzes the current session sequences based on incomplete user information to provide dynamic recommendation. Hence, session-based recommendation has come to be popular now.
Since they have the advantage of modeling complex transitions among items, graph neural networks have been a hot technology in the session-based recommendation. However, existing studies have largely ignored the attribute information of items when learning item transitions, which results in inadequate session interest representation. Considering that category information is informative in understanding users’ session interests, some studies have combined category information to enrich item transitions and showed effective session-based recommendation performance. However, category information is only one type of item attributes;multi-attributes(e.g., brand, price) of items are not explored effectively in learning item transitions, resulting in inadequate representation of the session interest.
To this end, we propose a novel graph neural network named multi-attribute-aware graph neural network, short for MASR, for session-based recommendation. First, MASR models the attribute association with the multi-head self-attention mechanism to optimize the representations of all attributes. And then, an attribute-aware graph neural network is designed to learn item transitions in the session, which effectively improves item representations by synthesizing the multi-attribute information. Finally, the soft attention mechanism is used to integrate item representations in the session to fully obtain the multi-attribute-aware session representation to provide recommendation.
To validate the effectiveness and rationality of the proposed method, we perform a series of experiments on two publicly available benchmark datasets obtained from the Cosmetics site. The results of two ablation experiments confirm the effectiveness of considering both attribute association and multi-attribute information in item transitions. Based on our parametric experiments, we have determined that the optimal number of layers for the attribute-aware graph neural network on both datasets is 2. In addition, we conduct comparative experiments between MASR and four popular session-based recommendation models. The results of the experiment confirm the proposed method outperforms other mainstream models in terms of Precision, Hit Rate (HR), and Mean Reciprocal Rank (MRR) on both datasets. Specifically, our method achieves about 20% and 2% improvement in terms of MRR for two datasets.
Despite demonstrating superior performance in the session-based recommendation, the proposed method has certain limitations. Specifically, our method only concentrates on item transitions within the current session and has not yet effectively addressed item transitions across sessions. Therefore, it may lead to limited performance of the session-based recommendation when the length of a session is limited. In the future research, we will combine multi-attribute information with item transitions across sessions to further improve the performance of session-based recommendation.

Key words: multi-attribute information, attribute association, session-based recommendation, graph neural network, attention mechanism