DAGNN: Demand-aware Graph Neural Networks for Session-based Recommendation Liqi Yang 1 , Linhao Luo 1 , Fengxin Li 2 , Xiaofeng Zhang 1 , and Xinni Zhang 1 1 Harbin Institute of Technology, Shenzhen, China 2 PING AN INSURANCE (GROUP) COMPANY OF CHINA liqiyanglqy@sina.com luolinhao@stu.hit.edu.cn LIFENGXIN682@pingan.com.cn zhangxiaofeng@hit.edu.cn zhangxinni.hit@gmail.com Abstract. Session-based recommendations have been widely adopted for vari- ous online video and E-commerce Websites. Most existing approaches are intu- itively proposed to discover underlying interests or preferences out of the anony- mous session data. This apparently ignores the fact these sequential behaviors usually reflect session user’s potential demand, i.e., a semantic level factor, and therefore how to estimate underlying demands from a session is challenging. To address aforementioned issue, this paper proposes a demand-aware graph neural networks (DAGNN). Particularly, a demand modeling component is designed to first extract session demand and the underlying multiple demands of each ses- sion is estimated using the global demand matrix. Then, the demand-aware graph neural network is designed to extract session demand graph to learn the demand- aware item embedddings for the later recommendations. The mutual information loss is further designed to enhance the quality of the learnt embeddings. Exten- sive experiments are evaluated on several real-world datasets and the proposed model achieves the SOTA model performance. Keywords: Recommendation System · Session-based recommendation · Graph Nerual Networks. 1 Introduction Session-based recommendation has been widely applied in various online video sites and E-commerce applications, and thus attracts a vast amount of research efforts. The prominent characteristics of session recommendation issue is that the length of an anonymous session is quite short which poses a great challenge to the conventional recommendation techniques. Without loss of generality, most session-based recommendations are technically de- signed to first explore users’ preferences hidden in each session, and then best match the feature representations between target item and the extracted preferences[5,10,27]. In the literature, there exist a good number of rela ted works to extract session pref- erences, such as attention-based method [15]. STAMP[17] is proposed to investigate how the preference of the last item affects the recommendation performance. Alterna- tively, chained-based methods are proposed to explicitly capture the order dependen- cies between session items for recommendation, e.g. FPMC [21] and GRU4Rec [10]. arXiv:2105.14428v1 [cs.IR] 30 May 2021