Research On Representation Learning Methods For Course Recommendation In Heterogeneous Information Networks

In the face of massive course resources provided by online education platforms,it is a common concern for the academia and enterprises to recommend suitable courses for learners quickly and accurately.However,traditional course recommendation methods suffer from serious issues such as cold starts and data sparsity,which results in poor performance in course recommendation.Nowadays,with the growth of online education data and the in-depth study of heterogeneous information networks,the representation learning approaches in heterogeneous information networks are booming,which provides novel solutions for course recommendation.Therefore,the research on representation learning methods for course recommendation in heterogeneous information networks has become one of the research hotspots in recent years.The representation learning methods for course recommendation in heterogeneous information networks are faced with significant challenges,which is outlined as follows: Firstly,as implicit feedback data are failed to be used in algorithms,it leads to poor performance in course recommendation;secondly,since explicit feedback data are poorly utilized in recommender systems,unsatisfied effect is reached in recommendation results;Finally,the cold start problem and high variance problem hamper the recommendation methods to achieve sati factoring performance in practical application.Aiming at the problems above,this thesis endeavors to study the representation learning methods in heterogeneous information networks for course recommendation on the Massive Open Online Courses(MOOC)platform.Furthermore,a series of experiments are conducted to validate effectiveness of the proposed methods,which are learner-course link prediction and course recommendation.The main innovations of the thesis are outlined as follows:(1)A representation learning method in heterogeneous information network with the name of Multi-Granular Online Learning Behavior Data(MGOLBD)is proposed to solve the problem of poor performance in recommendation algorithms due to its failure in utilizing implicit feedback data fully.Firstly,a learner-learner similarity matrix and a course-course similarity matrix are generated based on multiple similarity calculations,an implicit feedback matrix of learning behaviors is generated based on learners’ implicit learning behavior feedback data,and an Online Learning Behavior Heterogeneous Information Network(OLBHIN)containing direct and indirect interaction information is constructed based on the multi-granularity learning behaviors obtained from these matrices;secondly,we use Graph Convolutional Network(GCN)to learn the representation of meta-paths and use the attention mechanism to learn the importance of meta-paths,and weighted fusion of multiple meta-path representations to generate high-quality learner representations and course representations to enhance the expressive power of OLBHIN further;finally,the learned representations are used for course recommendation to verify the effectiveness of the proposed method.(2)A representation learning method in heterogeneous information network with the name of Multi-Source Teaching Data(MSTD)is proposed to address its inadequate utilization of explicit feedback data in course recommendation.Firstly,a Learning Behavior Heterogeneous Information Network(LBHIN)is constructed by sequentially constructing an explicit learning behavior feedback matrix,a learner-learner similarity matrix,and a course-course similarity matrix based on learner online review data,and fusing the implicit feedback matrices between different nodes;secondly,the learned representations are weighted and summed twice to obtain high-quality representations of learners and courses using a mean encoder to encode meta-path instances;finally,the obtained learner representations and course representations are used for learner-course link prediction to validate the effectiveness of the proposed method.(3)A representation learning method of Multi-Source Heterogeneous Data Fusion(MSHDF)is proposed for online education in heterogeneous information networks,which solves the problems of cold start and high variance to some extent.Firstly,fusing the above LBHIN and OLBHIN to build an Online Education Heterogeneous Information Network(OEHIN),which aims to solve the problem of inefficient course recommendation due to the cold start problem;secondly,the representation of meta-paths is learned using GCN,the weights of meta-paths are obtained using the multi-headed attention mechanism,and the final representation of nodes is obtained by weighting and summing the representations of meta-paths to reduce the variance while preventing overfitting;finally,the learned learner representations and course representations are used for course recommendation to verify the effectiveness of the proposed method.In this thesis,we conduct extensive experiments on the Xuetang X dataset,the MOOCCube dataset,and the dataset of 706 course reviews on the Xuetang X platform,respectively.Moreover,a variety of evaluation metrics are introduced to assess the effectiveness of the methods.The experimental results show that the series of approaches proposed in this thesis have improved the performance of course recommendations and achieved the expected recommendation effect.