Collaborative Filtering Research Based On User Characteristic Model And Interest Measure

With the development of the Internet, e-commerce technology is maturing and is widely applied to public life. People increasingly prefer to find the information they need on the network, so each site begins to shift their research focus in that according to the network user information which the site itself has to recommend the information they are interested. Collaborative filtering (Collaborative Filtering, CF) is the most successful personalized recommendation and it uses a known user evaluation to achieve the recommendation to the target user. Typical collaborative filtering algorithm is user-based collaborative filtering recommendation algorithm, its basic principle is to use the historical score data form user neighbors, according to the nearest neighbors’score data to recommend to the target user. With the increment of the number of users and projects, how to improve the algorithm scalability and recommendation quality is the main problem facing by the collaborative filtering technology.The paper proposed two optimized algorithms. The first one is a collaborative filtering algorithm based on user characteristic model. The second one is an optimized collaborative filtering recommendation based on Bayesian algorithm. The former algorithm optimized the traditional user-based collaborative filtering recommendation algorithm and started from the sparsity problem of user common scoring projects. At first, use an established formula to calculate the characteristic similarity of the user u and v according to user characteristics under the offline environment, to form a user characteristic similarity matrix. In order to improve the search speed, it established a feature neighbors sort linked list depending on the feature similarity measure. After finished creating the user characteristic model, statistics all projects of the two users evaluated, these projects could only be evaluated by one user or two users. For the project that was only evaluated by one user, you could predict the score value according to scores of the characteristic neighbor of the user who didn’t evaluate it. After this treatment, the project set evaluated by two users would increase significantly, and solved the problem that project set of common evaluation was small in the correlation similarity measure and the modified cosine similarity measure, and also solved the problem that all unrated projects were set zeros in the cosine similarity measure. In this way, it alleviated the sparsity problem to some extent in collaborative filtering recommendation, and improved the recommendation quality. The latter algorithm first introduced a threshold9to adjust the calculation of user similarity, and then grouped users combined with the user interest degree in the project, using the Bayesian algorithm to analyze user characteristics in order to estimate the probability of users liking unrated projects with different characteristic values. At last combined with the interest measure to determine the adjustment factor8and further optimized the user similarity formula, so that the calculation of users’nearest neighbor was more accurate. Finally, according to score information of the nearest neighbor to predict the score of the target users to the unrated project, and removed the highest forecast score in the first n terms to recommend to the user, and completed the recommendation.At last, the two optimization algorithms were combined. An optimized collaborative filtering algorithm was proposed based on user characteristic model and interest. Firstly, use the first method to fill the scoring matrix in order to alleviate the sparsity of the ratings data. Pre-processing improved the accuracy of the initial similarity calculation between users, and then combined with the degree of user interest in the project to group the users. Use the Bayesian algorithm to analyze user characteristics in order to estimate the probability of users with different characteristic values liking unrated projects. At last combined with the interest measure to determine the adjustment factor δ and further optimized the user similarity formula, so that the calculation of users’nearest neighbor was more accurate. At last, removed the highest forecast score in the first n terms to recommend to the user, and completed the recommendation.The experimental results show that compared with traditional user-based collaborative filtering recommendation algorithms, the proposed algorithm has higher validity and accuracy and improves the recommended quality of the system.