Study On Users' Interests Mining For Service Recommendation

With the development of economy, science and technology is developing rapidly too in our country. Now people have entered the era of big data. Many people do not know which one they want, when they are faced with large number of data. In order to solve the problem, recommender system appears.It aims to help people to discovery what they are interested in and find out what they are supposed to do. It recommends what users are interested in is based on user characteristics; user's browsing history and user's score and so on. Today the main use of recommendation system is in the aspect o electronic commerce. The person who provides the service becomes the initiative person by the recommendation, and gives them more profitable. Consumers are not so confused in the choice of time.Collaborative filtering is one of the most successful and widely used techniques among recommender systems. First calculating the similarity between users or services. According to the nearest neighbor analysis, the final results will be recommended to the user. However, this algorithm suffers from a range of problems, such as sparsely, multiple-content, stability and group recommendation.The main problems are improved in this paper. The main works of this thesis are as follows:Firstly, collaborative filtering based on mixed attributes is proposed. When people go shopping or other behavior, people may leave their personal information and browsing history and so on. In this paper, we will make full use of above information, including user's attributes, user's ratings, and user's browsing history. When using user attributes, we get the key attribute after many experiments; User's rating is the main use of the user's common score, and use the improved SVD(Singular Value Decomposition) method for data preprocessing; time stamp function is joined and used to achieve dynamic recommendation. The three aspects are calculated the similarity. Finally, the similarity is calculated and the final similarity is obtained, then using the KNN(k-Nearest Neighbor, KNN) carry on analysis. At the same time, the cold start and data sparse problem are improved.Secondly, a single user based random forest is proposed. Recommender systems use the modified random forest when recommending. In the process of recommendation, the first is to preprocess the data. If data is discrete data using SVD preprocessing, continuous data are discrimination, which uses the CADD(Class-Attribute Dependent Discretizer)algorithm for processing, then the training and testing, finally getting the recommendation list.Thirdly, the cross validation method is proposed, which is combination of mixed attributes based on collaborative filtering and the recommendation results based on random forest. The data is divided into N parts, N-1 is used training, the rest is used test, and finally use the improved collaborative filtering to achieve the final recommendation.In the experiment, the accuracy and recall rate as the evaluation criteria. Through a large number of experiments,result demonstrates the method is effective.