Research On Recommender System Based On Product Attributes

Recommender system is an efficacious tool that helps the user find interesting itemsand overcome information overload problem. Based on a survey of its up-to-datedevelopment from technical and managerial perspective, this thesis delves down into therecommender process using the knowledge from the users’ interest and the attributes of theitems. This research is implemented from the perspective of information granularity andinformation source.The algorithm and flow of the recommender system for frequent purchased productbased on the attributes of the product is discussed firstly. On the premise of not increasingthe workload of the user, the algorithm that converts the user-item matrix into user-item’sattributes is proposed referring to the TFIDF algorithm in information retrieval research.Using this matrix, three recommendation approaches which are separately summing thepartial utilities of attributes, artificial neural network and attribute-based collaborativefilter are proposed. Experiments are conducted on the public data sets from the Internet.The results show that the proposed algorithm and approaches can increase accuracy ofprediction and help reduce the data sparsity and cold-start problems of recommendersystem.Conversational recommender system for infrequent purchased product is discussedsecondly, which is based on the users’ goals and the items’ attributes, and in whichadditional domain expert knowledge is encoded. Based on the means-end model, thisthesis describes a conversational recommender process that supports for users in findingsatisfying shopping products, for which the system lacks of users’ history shopping dataand the users lack of product knowledge, according to the users’ qualitative goals and thedomain experts’ quantitative product knowledge. The domain expert knowledge includesspecific names for the purpose, the functions and the attributes of the products and theirquantitative relation. It is inputted by the domain experts with the help of analyticalhierarchy process. According to the research method in design science, a prototype systemis designed, which can obtain the domain knowledge of computer experts and recommendappropriate computers to the users based on their goals. Experimental results show that the conversational recommender process proposed in this thesis abates the users’ perceivedcomplexity and improves their decision efficiency when they choose and buy shoppingproducts.Finally the algorithm and process of mining user preference for the product attributesfrom the online product reviews are discussed. This algorithm and process make use oftechnologies of natural language processing and data mining and some artificiality. It canobtain three type of information including user preference for the product attributes,weight for the online product reviews and scores for the product attributes. Themechanism of the preference for recommender systems is discussed also. It may be usedto build the user preference model or as a source of product domain knowledge.Experimental results aimed at computers show that there are no significant differencesbetween the scores for products get by the method proposed in this thesis and given by theexperts. This demonstrates the feasibility and effectiveness of the proposed algorithm andprocess.