Multi-level sequence denoising with cross-signal contrastive learning for sequential recommendation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Sequential recommender systems (SRSs) suggest items based on user history.
• Existing methods struggle with noisy data, either over-weighting or discarding items.
• This limitation hinders accurate next-item prediction.

Purpose of the Study:

• Propose a novel model, Multi-level Sequence Denoising with Cross-signal Contrastive Learning (MSDCCL).
• Enhance sequential recommendation by effectively denoising user interaction sequences.
• Improve the accuracy and robustness of next-item suggestions.

Main Methods:

• Developed a target-aware user interest extractor for long and short-term interests.
• Introduced a multi-level sequence denoising module using soft and hard signal strategies.
• Extended curriculum learning by simulating human learning patterns.

Main Results:

• MSDCCL significantly outperforms state-of-the-art baselines on five public datasets.
• The model demonstrates superior effectiveness in handling noisy sequences.
• Experimental results validate the proposed denoising and interest extraction techniques.

Conclusions:

• MSDCCL offers a robust solution for noisy data in sequential recommendation.
• The model can be integrated with existing recommendation systems to boost performance.
• The approach advances the field of personalized item suggestion.