Continual Multimodal Contrastive Learning in Movie Recommendation Services

Personalized recommendation systems must adapt to rapidly changing user preferences and leverage diverse content modalities such as text, images, and audio. In this work, we propose CMCL—a Continual Multimodal Contrastive Learning framework designed for dynamic movie recommendation tasks. CMCL integrates multimodal movie features, including posters, descriptions, reviews, and trailers, to learn robust representations using self-supervised contrastive objectives. The framework combines cross-modal alignment with user-movie interaction modeling. It employs continual learning strategies such as knowledge distillation, experience replay, and meta-learning to retain past knowledge and adapt to new data. Evaluated on MovieLens-25M, IMDb, and Netflix datasets, CMCL outperforms traditional collaborative filtering, deep multimodal models, and continual learning baselines like EWC and replay-based methods. Besides, forgetting analysis shows improved robustness against performance degradation over time. These results highlight CMCL’s effectiveness in bridging multimodal alignment and lifelong learning for personalized, evolving recommendation systems.