With the explosive growth of information, recommender systems become a critical tool to alleviate the information overload problem in many online services such as e-commerce and media sharing websites. Conventional recommendation methods such as collaborative filtering rely on tracking user identities to model each individual user’s preferences, which may result in poor performance in scenarios where user identities cannot be tracked due to some reasons including anonymous users or privacy issues. Session-based recommendation (SBR) tackles this problem by assuming that users perform actions on a session basis, where a session is a sequence of actions in close temporal proximity. Under this assumption, users’ actions in the same session are highly correlated, and thus, the sequential and...[ Read more ]

With the explosive growth of information, recommender systems become a critical tool to alleviate the information overload problem in many online services such as e-commerce and media sharing websites. Conventional recommendation methods such as collaborative filtering rely on tracking user identities to model each individual user’s preferences, which may result in poor performance in scenarios where user identities cannot be tracked due to some reasons including anonymous users or privacy issues. Session-based recommendation (SBR) tackles this problem by assuming that users perform actions on a session basis, where a session is a sequence of actions in close temporal proximity. Under this assumption, users’ actions in the same session are highly correlated, and thus, the sequential and co-occurrence patterns in the active session can be utilized to more accurately model the current user’s preferences. Since SBR does not require user information and the “session-based” assumption is a common phenomenon, it is of great practical value and has received much attention in both academia and industry recently.

The key to building a successful session-based recommender system is to effectively utilize the properties of sessions by capturing both intra- and inter-session relationships. In this thesis, we introduce four studies for accurate session-based recommendation.

The first two studies focus on capturing intra-session relationships. In our first study, we consider the local invariance property in SBR, which states that the detailed order of user actions in local regions of sessions is not meaningful while the high-level order in the entire sessions reflect users’ intentions. We propose a model that can pay different attention to the ordering information in different levels of granularity by ignoring the insignificant detailed ordering information in some sub-sessions while keeping the high-level sequential information of the whole sessions. In our second study, we aim to improve the discrimination ability of graph neural network-based methods by addressing two information loss problems, namely the lossy session encoding and ineffective long-range dependency capturing problems. The first problem, lossy session encoding, says that different sessions are encoded to the same representation. The second problem, ineffective long-range dependency capturing, states that long-range dependencies among items cannot be explicitly captured due to the limited number of GNN layers. We propose a GNN model that does not have the two information loss problems by combining two novel GNN layers.

The other two studies focus on capturing inter-session relationships. In our third study, we consider the inter-session relationships in two levels, namely the item level and the session level. To capture the item-level inter-session relationships, we propose a GNN to automatically learn the importance of item co-occurrence patterns from a global graph that encodes the fine-grained information about item co-occurrences such as relative order and distance. To capture the session-level inter-session relationships, we propose four data augmentation techniques and adopt the constrastive learning framework to correctly cluster sessions with similar semantics. Lastly, we introduce our ongoing work that proposes a framework to help existing methods to more efficiently and effectively capture inter-session relationships when a social network among users is accessible. The proposed framework is able to integrate a variety of information such as user attributes and item categories when they are available. Existing methods can be plugged into the framework and achieve much better recommendation accuracy with the same inference efficiency.

We conduct extensive experiments on the commonly used public benchmark datasets and the results show that our methods are more effective than the state-of-the-art methods in terms of capturing intra- and inter-session relationships.