ID-Guided Multimodal experts with contrastive diffusion for sequential recommendation.

Area of Science:

• Artificial Intelligence
• Computer Science

Background:

• Multimodal sequential recommendation leverages text and image data for enhanced user-item interaction modeling.
• Existing methods struggle with cross-modal inconsistencies and noise within features, limiting recommendation quality.

Purpose of the Study:

• To propose a novel framework, ID-Guided Multimodal Experts with Contrastive Diffusion (IMECD), to address limitations in current multimodal recommendation systems.
• To improve the quality of multimodal representations by resolving semantic inconsistencies and suppressing irrelevant signals.

Main Methods:

• Implemented an ID-guided multimodal mixture of experts module to dynamically guide feature extraction using long-term user preferences.
• Introduced a modality-specific vector quantization module for denoising sequential features.
• Developed a contrastive diffusion generation module to mitigate generation bias using sequence representations and contrastive loss.

Main Results:

• The proposed IMECD framework demonstrated consistent outperformance against state-of-the-art baselines across four benchmark datasets.
• The ID-guided module effectively resolved cross-modal semantic inconsistency and suppressed irrelevant signals.
• Modality-specific vector quantization successfully denoised sequential features, enhancing representation quality.

Conclusions:

• IMECD offers a significant advancement in multimodal sequential recommendation by effectively handling cross-modal inconsistencies and data noise.
• The framework's ability to leverage ID embeddings and advanced techniques like contrastive diffusion contributes to superior recommendation performance.