@techreport{oai:ipsj.ixsq.nii.ac.jp:00214565,
 author = {Anyu, Cai and Shin'ichi, Konomi and Anyu, Cai and Shin'ichi, Konomi},
 issue = {8},
 month = {Dec},
 note = {In recommender systems, capturing the rich sequential information in historical interaction sequences is essential for user representation learning. However, most of the existing methods only focused on modeling the user's historical information into one fixed-length vector, which brings two limitations. First, one fixed-length vector is insufficient to model a user's varying interests. This simplified user modeling leads to monotonous recommendations, making users bored and trapped in the information cocoon. Second, since the user's interests are naturally dynamic and evolving, merely considering past information can only capture the user's outdated interests and predict the interest changes based on a fixed period (i.e., they model the elapsed times since the historical interactions into a same interval). As a result, they cannot give different predictions when users access the system after different time intervals, resulting in suboptimal recommendation performance. In this paper, we seek to explicitly model the users' diverse interests and elapsed time within a sequential interests modeling framework to explore the causality between users' multi-interests and the influence of different elapsed times. We propose a Dynamic Evolutionary Multi-Interest network for sequential RECommendation (DemiRec), which models user's historical interactions into an interest sequence and predicts evolved interests based on sequential information and elapsed time. Extensive empirical studies on real-world datasets demonstrate that the proposed method significantly outperforms state-of-the-art multi-interest baselines., In recommender systems, capturing the rich sequential information in historical interaction sequences is essential for user representation learning. However, most of the existing methods only focused on modeling the user's historical information into one fixed-length vector, which brings two limitations. First, one fixed-length vector is insufficient to model a user's varying interests. This simplified user modeling leads to monotonous recommendations, making users bored and trapped in the information cocoon. Second, since the user's interests are naturally dynamic and evolving, merely considering past information can only capture the user's outdated interests and predict the interest changes based on a fixed period (i.e., they model the elapsed times since the historical interactions into a same interval). As a result, they cannot give different predictions when users access the system after different time intervals, resulting in suboptimal recommendation performance. In this paper, we seek to explicitly model the users' diverse interests and elapsed time within a sequential interests modeling framework to explore the causality between users' multi-interests and the influence of different elapsed times. We propose a Dynamic Evolutionary Multi-Interest network for sequential RECommendation (DemiRec), which models user's historical interactions into an interest sequence and predicts evolved interests based on sequential information and elapsed time. Extensive empirical studies on real-world datasets demonstrate that the proposed method significantly outperforms state-of-the-art multi-interest baselines.},
 title = {DemiRec: Dynamic Evolutionary Multi-Interest Network for Sequential Recommendation},
 year = {2021}
}