Multi-factor disentangled graph neural networks for session-based new item recommendation

Recent studies have shown that graph neural networks for session-based recommendation systems typically recommend old items, making it difficult to recommend new items to users, leading to the phenomenon of the "information cocoon". To address this issue, this paper introduces a Multi-Factor Disentangled Graph Neural Network for Session-Based New Item Recommendation (MFD-GNN), which considers both the embedding of new items and user intent from a multi-factor perspective. First, item embeddings from sessions are generated across multiple factors using a disentangled network. By leveraging item classification and attribute information, new item embeddings are inferred through zero-shot learning. Attention weights are assigned to each factor to capture user intent across different factors, enabling reasonable recommendations for new items. Experiments are conducted on two publicly available datasets, and the results are compared with those of leading recommendation models. The findings demonstrate that the proposed method surpasses current models in performance. These experimental outcomes confirm the approach's effectiveness and its advantages over existing methods.