Research On Online Course Recommendation Incorporating User Review Topics

The rapid development of modern science and technology has pushed human society into the information age in an all-round.The boundary of education between time and spaces has been constantly broken,and online education is developing quickly with the popularization of Internet technology.However,the development of online education not only brings massive learning resources to learners,but also brings many troubles.Users will have to spend a lot of time and efforts to search for the online course that they want under the continuous increase of online course resources.Personalized Recommendation technology is to simulate the process of salesmen recommending what to buy for customers according to their own characters.Collaborative Filtering Recommendation algorithm is the most widely used and studied personalized recommendation technology.Its main idea is to calculate the similarities between users or items based on the current existent history ratings,and predict users' preference for items based on this,so as to realize the entire process of recommendation.Comments are the carriers for network users to express their feelings and opinions on the platform,including experience,function requests,error reports and other information.If the comments on courses published by users can be taken into consideration,more accurate similarities between users or courses will be obtained,so as to realize more specific course recommendation.However,most of the user comments on the Internet are disorganized or even without content.Some comments are just the expression of emotions from the users,or some "long articles" may contain the key information that the user wants to express,but it takes a lot of manpower to read and understand and refine them.Therefore,other more economical,more effective and more comprehensive methods are needed to mine the topics of user comments and apply them to the collaborative filtering recommendation algorithm,so as to achieve more accurate course recommendation.Based on all above,this paper focuses the research object of online course recommendation on user comments and considers it into the traditional collaborative filtering recommendation based on course rating similarities,so as to improve the accuracy of recommendation.First of all,this paper calculates the similarities based on existing historical rating information according to traditional Collaborative Filtering Recommendation algorithm.Secondly,the topics are extracted from comments through topic modeling,and they are utilized as potential attributes to generate the course features portraits,so that course similarities are gained according to the portraits.For the topics extraction from user comments,Word2 Vec word vector model is added into the Latent Dirichlet Allocation(LDA)model which only considers word occurrence probability,to transform "Topic-Word" matrix generated by traditional Latent Dirichlet Allocation(LDA)model into "Topic-Word Vectors" matrix.And K Prototypes cluster is utilized to cluster the "Topic-Word Vectors" matrix via combining word vector(numeric feature)and subject classification(textual feature),so that the topics with semantic information is obtained to calculate the similarities based on review semantic topics.Then,Multivariable Linear Regression model is used to fit the similarity of course ratings with the similarities based on course comments' topics or that based on semantic topics of course comments,and the summarized weighted courses similarity scores are acquired so that the user ratings of the course can be predicted based on that.Finally,the results are analyzed by comparing collaborative filtering based on course ratings,collaborative filtering based on course ratings and topics,and collaborative filtering based on course ratings and semantic topics.It has been proven that topics of course comments added into course similarity calculations improves the prediction of course ratings and the effectiveness of online course recommendation.Moreover,the accuracy of online course rating prediction can be further improved after the course topics with word semantic information are added to the traditional course similarity calculation,which are extracted via the improved topic model.From the perspective of practical application,different from repeated purchase in general recommendation,users will never be interested in the courses that they have learned before.It is not enough if we just recommend courses based on the predicted ratings via collaborative filtering recommendation,so we need to associate it with other recommendation algorithm,such as Information Retrieval technology.The innovation of this paper can be summarized as follows.Firstly,the topics of course comments are used as potential attributes for the courses and the feature portrait of courses is generated,so that the similarities between courses based on the topics are obtained according to the feature portrait of the course,which will improve the accuracy of predicted rating about courses.Secondly,it considers the semantic relationship between words.Word2 Vec word vector model is utilized to transform the "Topic-Word" matrix generated by the traditional Latent Dirichlet Allocation(LDA)model into the "Topic-Word Vector" matrix,which will further improve the accuracy of predicted rating about courses.Thirdly,K prototypes Cluster is utilized to cluster the "Topic-Word Vectors" matrix via combining word vector(numeric feature)and subject classification(textual feature),so that topics with semantic information is obtained to calculate the similarity score based on semantic topics,which will make the course similarities closer to the reality.Fourthly,Multivariable Linear Regression model is used to fit the similarity of course ratings with the similarity based on improved topics from user comments,and the summarized weighted courses similarities are acquired so that the users' ratings of the course can be predicted based on that.