Research on the application of knowledge mapping and knowledge structure construction based on adaptive learning model

This project has developed a geometry learning software that integrates multiple computer technologies to address the challenges of deep analysis of knowledge points and establishing connections in learning software. The software combines Long Short -Term Memory (LSTM) and Residual Neural Network (ResNet101) to encode text and image features. A self -attention mechanism is used to fuse information from both modalities, enabling decoding of geometric models and classification of corresponding knowledge points.This project uses LSTM and ResNet101 models to extract text and visual features for problem -solving using the Multi Mode Thinking Chain (CoT) method. Classification labels are utilized to generate text responses for problem -solving ideas. Furthermore, a recommendation module is proposed, which combines knowledge tracking and neural collaborative filtering algorithms to capture student behavior and knowledge point vectors. Implicit factors representing students' mastery of different knowledge points are used as inputs in neural collaborative filtering for personalized recommendations. The results demonstrate improvements in accuracy using the ResNet + LSTM multimodal algorithm, achieving a 13 % increase compared to single -modal classification. The multimodal CoT approach also outperforms language models like GPT3.5 and VisualBert by 10 %. Additionally, the combined algorithm of knowledge tracking and neural collaborative filtering shows a 13.3 % higher F1 value compared to ordinary algorithms, confirming the superiority of the adopted method in this project.