Research Of Personalized Learning Recommendation Technology

The teaching model of Massive Open Online Course(MOOC)provides people with convenient learning scenarios.The rapid growth of online course resources leads to the problem of information overload when learners face a large amount of course resource data.In order to help people quickly find suitable courses that they are really interested in,and at the same time make high-quality online courses known to more people,it is necessary to use the recommender system to analyze the learner's historical behavior data,mine their potential preference information,and realize the personalized recommendation for the learner.However,when there are few historical interaction records between users and items,traditional recommendation algorithms usually have problems such as cold start,which limits the performance of the algorithm.In view of the above problems,considering that there are a lot of semantic information in heterogeneous information networks,an attempt is made to combine the heterogeneous information network with the matrix factorization model,and the method of combining explicit and implicit feedback is used to improve the performance of the model.The research contents of this dissertation mainly include:(1)The data on the MOOC platform is modeled by heterogeneous networks,and the semantic association between users and courses is obtained through meta-paths,and then the content feature representations under different meta-paths are learned through graph convolutional neural networks.In order to make full use of the feature information of the neighbor nodes of each layer of the graph convolutional neural network,the original single serial graph convolution layer hierarchical propagation structure is changed to a crosslayer connection structure.By cascading each graph convolutional layer,neighbor feature representations of different orders are obtained.Finally,in order to distinguish the contribution of content feature representation under different meta paths,the global feature representation is obtained by fusing them through the attention mechanism,and then applied to the matrix decomposition model.(2)Research on recommendation algorithms under different user feedback mechanisms.Firstly,a weighted probability sampling algorithm based on the popularity and similarity of courses is designed to sample negative samples,and training by paired ranking learning method.Then,an implicit scoring mechanism is proposed to convert the user's historical log data into scoring data,and use the method of mean square error loss for training.Finally,combine the training results of the above two to get the final recommendation list.In order to verify the feasibility and effectiveness of the model,experiments are carried out on MOOCCube dataset and self-made IMOOC dataset.At the same time,compared with the classical recommendation algorithm,considering the influence of important parameters and components in the model on the performance of the model,different ablation experiments and comparison experiments are designed,and the experimental results are comprehensively analyzed and evaluated.