Research On Open Online Course Recommendation Algorithm In MOOC

In the background of the pandemic of the Covid-19 epidemic,since the massive open online course(MOOC)platform can not be affected by region and time,learning courses on MOOCs has become a relatively simple and convenient way of learning,and people’s demand for online courses is increasing.At the same time,more and more teachers,experts,and scholars are also publishing their own courses on the MOOC platform,providing learners with more course choices.But this brings the problem of information overload,and it becomes a challenge to choose a course that suits people from the numerous course lists.To cope with this challenge,MOOC course recommendation(MCR)has become a research hotspot for experts and scholars.Course learning on the MOOC platform has the following characteristics:(1)There are various types and a large number of courses on the MOOC platform,but the quality of these courses is different.(2)The course learning in MOOCs is different from the traditional learning method of sitting in the classroom.Most of the learners use the fragmented time to study.(3)Most of the people who study on the MOOC platform do not know each other.Their history courses,educational backgrounds,and learning purposes are different.Therefore,the course recommendation task in MOOCs has the following challenges:(1)The course selection data of learners is very sparse.It is difficult to obtain a learner’s preference for a course due to the limited records of historical course selections.(2)Links between learners are difficult to express directly.Due to the flexibility of time and space for learners to study courses,and the diversity of course choices,the relationship between learners is different from that of classmates in previous offline classes.In order to solve the above-mentioned problems regarding the MOOC course recommendation task,this thesis makes adaptive improvements for the existing graph neural network and hypergraph neural network models by considering the course selection behavior sequence of different learners.The details are as follows:(1)This thesis proposes a course recommendation algorithm(CSGMC)based on the graph attention network.Different from the general course recommendation algorithm,the CSGMC model expands the courses’ relationships to the global perspective,takes into account the crosssequence course information,and constructs a course sequential graph(CSG)according to the chronological relationship.Then,two weighted directed graph attention layers are used to model the courses in the course sequential graph,and different weights are given through the attention mechanism to obtain the course embedding.Afterward,the short-term sequential relationships between courses are learned by a sequence encoder(SE).Finally,the learner’s preference for the course is obtained by combining the course embedding learned from the global perspective with the short-term sequence relationship.(2)On the basis of the above algorithm,a hyperedge-based graph neural network course recommendation algorithm(HGNN)is further proposed,which can take the relationship among learners into consideration.First of all,this thesis represents the learner by its historical course set and further represents the learner by the hyperedge in the hypergraph,and the overlapping relationship between two hyperedges is used to represent the learner’s relationship.Afterward,a hyperedge-based graph neural network is applied to infer the learner’s representation,and the attention mechanism is used to assign different weights to adjacent learners.The obtained learner representation and the sequence-level learner representation learned through the course sequential graph are subjected to an inner pooling operation to obtain the final learner representation,thereby inferring the learner preference information.(3)To evaluate the effectiveness of the two models presented in this thesis,experiments are carried out on the Xuetang X,which is a MOOC dataset.Meanwhile,we conduct experiments on the prevailing dataset Movie Lens to confirm the scalability of our solution.In the experimental results,it can be seen that the two models proposed in this thesis can outperform other existing related methods on both datasets,demonstrating the necessity of comprehensively considering the global sequence information and short-term sequence patterns of the courses,and the effectiveness of using the hyperedges in hypergraphs to model learners’ relations.