Collaborative filtering based on GNN with attribute fusion and broad attention.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Collaborative filtering (CF) is crucial for recommender systems.
• Traditional CF methods struggle with nonlinearity and higher-order feature interactions.
• Graph neural networks (GNNs) show promise but have limitations in attribute aggregation and exploiting higher-order information.

Purpose of the Study:

• To propose a novel GNN-based CF model, GNN-A², to address limitations in existing GNN CF methods.
• To improve recommendation performance by effectively distinguishing and aggregating attribute interactions.
• To leverage higher-order interaction information for more accurate predictions.

Main Methods:

• Developed GNN-A², a GNN-based CF method incorporating attribute fusion and broad attention.
• Implemented an inner interaction module with self-attention.
• Designed a cross interaction module with attribute fusion and a broad attentive cross module.

Main Results:

• GNN-A² demonstrated comparable performance in Area Under the Curve (AUC).
• Achieved optimal Normalized Discounted Cumulative Gain at rank 10 (NDCG@10) across three benchmark datasets (MovieLens 1M, Book-crossing, Taobao).
• Outperformed state-of-the-art (SOTA) models by up to 2.14% in NDCG@10.

Conclusions:

• GNN-A² effectively handles inner and cross interactions differently, extracting higher-order information for improved prediction.
• The proposed model offers a significant advancement in GNN-based collaborative filtering.
• Experimental results validate the superiority of GNN-A² on benchmark datasets.