Deep Learning-Based Dual-Tower Contrastive Recommendation System

This study proposes a Deep Collaborative Dual-Tower Model (DCDT) to address the issue of insufficient feature interaction in multi-modal recommendation systems for educational scenarios. Through three key innovations, DCDT achieves precise recommendation of teaching resources: First, we construct a dual-path heterogeneous encoder that uses orthogonal constraints to decouple the user query and video content feature processing, and apply cross-modal attention mechanisms for semantic alignment. Second, a dynamic hard negative sampling strategy is designed, building high-quality training pairs based on the top-30% similarity threshold. Third, a hybrid-interaction layer is developed to enhance multi-modal correlations through feature concatenation and dot product operations. On our self-built linear algebra teaching video dataset, DCDT achieves 82.1% accuracy and 0.641 MRR, improving the F1 score by 19.8% compared to traditional collaborative filtering methods, with a response time of under 10ms. Ablation experiments show that the dual-tower architecture, with a feature pre-computation mechanism, improves recall efficiency by 8.6%. This study is the first to apply a contrastive learning framework to educational multi-modal recommendation systems, solving the conflict between heterogeneous feature fusion and real-time responses, and offering a new paradigm for the development of intelligent education systems. Future work will integrate knowledge graphs to enhance content understanding and explore federated learning frameworks to ensure data privacy.