With the transformation of college English teaching to online and offline integration and intelligent support, the problems of individual differences, insufficient memory retention and disconnection of context application in vocabulary learning have become more prominent. Aiming at college English blended vocabulary teaching, this paper constructs an intelligent education platform integrating learning and application, and proposes multi-source behavior event coding, scene-enhanced feature mapping, vocabulary knowledge graph modeling, deep learning mastery state recognition and scene-aware recommendation strategies. The platform converts preview browsing, quiz answering, review interval, reading and writing call and oral task feedback into trainable event sequences, and uses the graph node relationship, contextual attention mechanism and multi-factor scoring model to generate personalized recommendation results. The experiment collects 16 weeks of learning data from 218 students, forms 895205 effective learning events, and completes model training, graph query, cache scheduling and service deployment based on PyTorch, Neo4j, Redis and FastAPI. The results show that the Accuracy, Precision, Recall and F1-score of the proposed model reach 93.8%, 93.2%, 92.7% and 92.9%, respectively. After 8-week intervention, the learning completion rate increased from 78.6% to 93.1%, and the memory retention rate increased from 70.4% to 88.5%. This study provides a verifiable technical path for the data-driven personalized practice of college English vocabulary teaching and the optimization of intelligent education platform.
