Multi-behavioral recommendation model based on dual neural networks and contrast learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recommender Systems

Background:

• Recommender systems face challenges in capturing complex user-item dependencies and mitigating the smoothing problem in multi-layer neighborhood information.
• Understanding feature and temporal correlations between user behaviors is crucial for accurate recommendations.

Purpose of the Study:

• To propose a novel multi-behavior recommendation model, DNCLR, that leverages dual neural networks and contrast learning.
• To effectively capture both feature and temporal correlations within user behaviors.
• To alleviate the smoothing problem in recommender systems.

Main Methods:

• Utilized graph convolution networks (GCN) to learn user and item features across different behaviors.
• Employed a self-attention mechanism to learn feature correlations between behaviors.
• Incorporated a recurrent neural network (RNN) with attention to capture temporal correlations in user interaction sequences.
• Integrated contrast learning within the dual neural network architecture to refine user and item representations.

Main Results:

• The DNCLR model demonstrated significant improvements in recommendation accuracy compared to existing methods.
• Achieved HR@10 improvements of 2.5% on Yelp, 0.3% on ML20M, and 4% on Tmall datasets.
• Effectively captured complex dependencies between users, behaviors, and items.

Conclusions:

• The proposed DNCLR model effectively enhances recommendation accuracy by addressing complex behavioral dependencies.
• The integration of dual neural networks and contrast learning offers a robust approach for multi-behavior recommendation.
• DNCLR provides a promising solution for improving the performance of modern recommender systems.