Design of Personalized Intelligent Teaching Resource Recommendation System Integrating Multimodal Contrastive Learning

Traditional teaching resource recommendation systems lack multimodal data fusion, resulting in poor recommendation results. This study proposes a personalized intelligent teaching resource recommendation system, which aligns the spatiotemporal features of video, text, and knowledge graphs through a cross-modal attention mechanism, and designs multi-level contrastive learning: the MoCo framework is used locally to enhance visual-text consistency; hypergraph contrastive learning is used globally to mine high-order associations, and adversarial training is combined to improve robustness. Capsule networks are used to separate user multimodal interests; a dual-tower structure is used to fuse collaborative filtering and semantic matching; and multi-task optimization is used to predict click-through rate and knowledge mastery. Online $A / B$ tests show that the average pass rate of the experimental group increased by 11.1%; the learning time was shortened; and the Gini coefficient was reduced (0.28 vs. 0.52), which verifies the effectiveness of the system in accurate recommendation, efficiency improvement, and resource diversity optimization.