Disentangled graph-text collaborative filtering with prototypical cold-start for scientific article recommendation

Area of Science:

• Information Science
• Computer Science
• Artificial Intelligence

Background:

• Scientific article recommendation systems often struggle to balance collaborative filtering's reliance on user interaction data with content-aware methods' use of textual information.
• Existing graph-based models excel at warm-start scenarios but fail to represent new articles effectively, while content-based methods may neglect valuable graph structures.

Purpose of the Study:

• To propose a novel framework, Disentangled Graph-Text Collaborative Filtering (DGT-CF), that integrates collaborative filtering and content-based approaches for enhanced scientific article recommendation.
• To address the limitations of existing methods in handling both user interaction patterns and semantic article content, particularly for unseen articles.

Main Methods:

• DGT-CF employs a disentangled approach, separating collaborative and textual representations learned through distinct branches.
• The collaborative branch utilizes LightGCN, while the content branch employs a dual-stream encoder for article text.
• Key integration interfaces include score-level fusion, a content bridge, orthogonality regularization, and a prototype-based cold-start module for zero-shot ranking of new articles.

Main Results:

• DGT-CF demonstrated consistently strong performance in warm-start recommendation scenarios across CiteULike-A and CiteULike-T datasets.
• The framework significantly improved cold-start ranking compared to established collaborative filtering, content-aware, graph contrastive, and language model baselines.
• Performance was further validated on the MIND-small dataset, confirming its robustness even with out-of-domain data.

Conclusions:

• The proposed DGT-CF framework effectively combines collaborative and content information for scientific article recommendation.
• DGT-CF offers a robust solution for both warm-start and cold-start recommendation challenges, outperforming existing state-of-the-art methods.
• The disentangled architecture and prototype-based cold-start module are key innovations enabling superior performance, especially for recommending new or unseen articles.