Neighborhood structure enhancement and denoising method for multi-behavior recommendation

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Traditional recommender systems often oversimplify user interactions, assuming a single behavior type.
• Real-world user engagement involves complex, multi-faceted behaviors like browsing, clicking, adding to cart, and purchasing.
• Existing multi-behavioral recommendation methods face challenges with unbalanced data, sparse information, and noise from auxiliary behaviors.

Purpose of the Study:

• To address limitations in current multi-behavioral recommendation systems, specifically unbalanced data and noise.
• To propose a novel method, Neighborhood Structure Enhancement and Denoising (NSED), for more accurate user preference modeling.
• To enhance the representation of neighbor nodes and mitigate the long-tail problem in recommendation systems.

Main Methods:

• Implemented a neighborhood-enhanced Graph Convolutional Network (GCN) to strengthen node representations.
• Employed a structural enhancement module to address the long-tail problem and improve neighbor information.
• Utilized cross-behavioral modeling by cascading structures to uncover dependencies between different user behaviors.
• Integrated a denoising module with contrastive learning to mitigate negative migration and refine auxiliary behavior information.

Main Results:

• NSED significantly improves user preference modeling by strengthening neighborhood structures and denoising cross-behavioral data.
• The method effectively captures dependencies among diverse user behaviors through cascading structures.
• User preferences learned under the target behavioral graph demonstrate high accuracy, while auxiliary behavioral graphs are effectively denoised.
• NSED achieved an average performance improvement of 10.4% and 10.67% over SOTA baselines on three public datasets.

Conclusions:

• NSED offers a robust solution for multi-behavioral recommendation systems by enhancing graph structures and reducing noise.
• The proposed method effectively models complex user preferences, leading to substantial performance gains.
• NSED successfully addresses key challenges including data imbalance, sparse information, and negative migration phenomena.
• The approach demonstrates superior performance compared to existing state-of-the-art methods, validated on multiple benchmark datasets.