Reproduce “CaIRec: Calibrated Modality Imputation for Incomplete Multimodal Recommendation” on OpenResearch

About this paper

Ruiyu Liu, Xiaohao Liu, Miaomiao Cai, Yunshan Ma, See-Kiong Ng

CaIRec is a two-stage framework that addresses challenges in incomplete multimodal recommendation by first calibrating the structural coherence of imputed modalities and then adapting these completed representations and item relations for personalized ranking. The framework consistently outperformed baselines on Amazon datasets, achieving an average relative improvement of 3.77% in recommendation performance and demonstrating robustness to varying missing data rates.

Key insights

• Ensuring cross-modal structural coherence during modality imputation, beyond simple content recovery, is critical for generating effective multimodal item representations.
• Explicitly adapting imputed representations and reconstructing item-item relations for personalized ranking, through mechanisms like pseudo-missing alignment and completion-aware graphs, bridges the gap between imputation and recommendation.
• A two-stage decoupled optimization approach, separating structural imputation calibration from preference-oriented adaptation, allows for focused improvement of each specific challenge in incomplete multimodal recommendation.

Summary by alphaXiv.