Adaptive self-supervised learning for sequential recommendation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recommender Systems

Background:

• Sequential recommendation models often suffer from insufficient interaction data, leading to sparsity issues.
• Current self-supervised learning (SSL) methods in sequential recommendation are limited to single approaches and simple data augmentation.
• Existing models fail to fully leverage the combined power of diverse SSL techniques and graph augmentation strategies for improved item representation learning.

Purpose of the Study:

• To propose a novel multi-task sequential recommendation framework, Adaptive Self-supervised Learning for sequential Recommendation (ASLRec).
• To address limitations in current SSL-based sequential recommendation by adaptively combining contrastive and generative methods.
• To enhance item representation learning by exploring combined graph augmentation schemes and multiple loss functions.

Main Methods:

• ASLRec adaptively combines contrastive and generative self-supervised learning methods.
• The framework applies diverse perturbations at graph topology and node feature levels to create augmented graph views.
• Joint training utilizes multiple loss functions (contrastive, generative, mask, prediction) and adds uniform noise to mitigate popularity bias.

Main Results:

• ASLRec achieves state-of-the-art performance on three benchmark datasets compared to 14 competitive methods.
• The hit rate (HR) improved by over 14.39% and normalized discounted cumulative gain (NDCG) increased by over 18.67%.
• The model effectively mitigates interaction noise and data sparsity, learning more robust item representations.

Conclusions:

• The proposed ASLRec framework demonstrates superior performance in sequential recommendation tasks.
• Adaptive combination of multiple SSL methods and graph augmentation strategies is effective for learning better item representations.
• ASLRec offers a significant advancement in addressing data sparsity and noise in sequential recommendation systems.