Graph Neural Network-Guided Contrastive Learning for Sequential Recommendation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recommender Systems

Background:

• Sequential recommendation systems often employ contrastive learning to address data sparsity by augmenting user sequences.
• A key limitation is the potential loss of semantic similarity in augmented data views, hindering performance.
• Existing methods lack robust mechanisms to ensure semantic coherence during data augmentation for sequential recommendation.

Purpose of the Study:

• To propose a novel Graph Neural Network-guided Contrastive learning framework for Sequential Recommendation (GC4SRec).
• To enhance the semantic consistency of augmented data views in contrastive learning for sequential recommendation.
• To improve the overall performance and mitigate data sparsity in sequential recommendation tasks.

Main Methods:

• Utilized Graph Neural Networks (GNNs) to derive meaningful user embeddings.
• Incorporated an encoder to calculate item importance scores for guided data augmentation.
• Developed a contrastive view construction method leveraging item importance scores and diverse augmentation techniques.

Main Results:

• GC4SRec demonstrated significant improvements on three public datasets.
• The model achieved a 1.4% increase in hit rate and a 1.7% increase in normalized discounted cumulative gain (NDCG).
• Experimental results validate the effectiveness of the proposed guided contrastive learning approach.

Conclusions:

• Graph neural network-guided contrastive learning (GC4SRec) effectively addresses the semantic similarity challenge in sequential recommendation.
• The proposed method enhances recommendation performance by improving key metrics like hit rate and NDCG.
• GC4SRec offers a promising solution for mitigating data sparsity issues in sequential recommendation systems.