Category-Intent Fusion Graph Neural Network for Session-based Recommendation

Abstract

A user's interaction with an item is determined by a combination of intentions, such as a desire to purchase an electronic device while following a trend. However, these intentions are often unobservable, making it difficult to model user intentions and improve session recommendations. To tackle this problem, we propose a novel approach called the Category-Intent Graph Neural Network (CIGNN), which leverages the relationship between item categories and user intentions to provide accurate recommendations. We translate the category information into a compact representation, which represents the user's intent, and construct a category-intent fusion graph with item, category, and intent nodes. This graph connects multiple potential intents for each item in a session to capture user intentions and increase the expressiveness of item representations. The CIGNN model transfers information between intent, item, and category nodes, updating their representations alternately. Our experimental results on three benchmark datasets demonstrate the superiority of the CIGNN model over state-of-the-art (SOTA) methods in session-based recommendation (SBR).