Invariant Representation Learning in Multimedia Recommendation with Modality Alignment and Model Fusion.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recommender Systems

Background:

• Multimedia recommendation systems predict user preferences using multimodal data.
• Existing methods suffer from poor generalization due to learning spurious features.
• Prior invariant learning approaches lack proper data alignment, causing information loss.

Purpose of the Study:

• To propose a novel framework, M³-InvRL, to enhance recommendation system performance.
• To address generalization issues in multimodal recommendation systems.
• To improve the accuracy and robustness of user preference prediction.

Main Methods:

• Learning common and modality-specific representations using a novel contrastive loss.
• Extracting modality-specific features via mutual information constraints to prevent generalization issues.
• Generating invariant masks from heterogeneous environments for invariant representation learning.
• Integrating invariant-specific and shared invariant representations for model training and fusion.

Main Results:

• The proposed M³-InvRL framework effectively learns aligned and modality-specific representations.
• Invariant learning successfully identifies and utilizes robust features across different environments.
• Model merging reduces uncertainty and enhances generalization performance.
• Experiments on real-world datasets validate the proposed approach's effectiveness.

Conclusions:

• M³-InvRL significantly improves generalization ability in multimedia recommendation systems.
• The framework's approach to representation learning and invariant learning is effective.
• This work offers a promising direction for developing more robust and accurate recommender systems.