Item correlation modeling in interaction sequence for graph convolutional session recommendation

Abstract

Session-based recommendation aims to predict the user's follow-up behavior based on the user's short-term behavior in the session and is mostly used in scenarios where the user visits the website anonymously. The key problem of session-based recommendation is to model collaborative relationships between items to improve the recommendation performance. Previous methods primarily focused on discovering the interaction sequence of items to learn correlations between items and only explored the information of adjacent interactive items without taking into account historical interactive information. Actually, if two items share the same historical interaction information, they can be considered to have strong relationships. For example, in a session where the interaction sequences of two items are $I_1$-$I_2$-$I_3$ and $I_1$-$I_2$-$I_4$, the historical interaction information of item $I_3$ and item $I_4$ are both $I_1$-$I_2$. Thus, item $I_3$ and item $I_4$ can be deemed as similar Toward this end, this paper proposes a new session-based recommendation model named IHGCN to jointly explore the adjacent interactive information and the historical interactive information by analyzing the interaction sequence of items in the session. To learn the item feature from the interactive information, the proposed method introduces the graph structure to model correlations between items and then uses graph convolutional networks to abstract the item features from the item correlation graph The proposed IHGCN method introduces a fine-grained attention mechanism in the feature dimension level to discover the global preference of users and then integrate the local preference of users. Based on the learned item features and user preference features, the final recommendation results are obtained. Experiments conducted on three public datasets show the superiority of the proposed method to the state-of-the-art methods.