The Research And Application Of Collaborative Filtering Algorithm

With the further development of science and technology, information changes rapidly with time, and a lot of valuable data was soon hidden in the large data-tank easily. Therefore, how to let the Internet users quickly find the information which they are interested in becoming a research hotspot. Collaborative filtering recommendation system provides an effective solution to this problem. Nowadays, the recommendation system has got a very wide range of applications in Internet, retail, financial services and so on.Recommendation system is a kind of software which puts forward some recommendations to individual system user based on analyses of many former Internet users’scoring for each item, thus the impartiality and enthusiasm of network users’rating will directly influence the effect of the recommendation system. However, the most active network users could not score on all items, which leads to existing sparsity in the rating data, on the other hand, driven by interest, the phenomenon that a few Internet users with business background maliciously score on some projects is also inevitable, which will unavoidably result in the reduction of the credibility of the recommendation system. On the basis of the analyses of domestic and foreign research status, we have conducted some studies on the safety and cold start of the system subsequently.In general, malicious users have the characteristics of time concentration, high correlation, big difference in score distribution, etc. According to the characteristics of high correlation, we can use PCA-VarSelect algorithm to derive the suspect list of malicious attacking, as to the characteristics of big difference in score distribution and centralized time, we can first segment the time series in database and determine the items group as well as the time range being attacked respectively based on the score deviation so as to reduce the searching range, then analyze the scores in the attacking period, find the time anomalies, determine the types of attack, and finally, detect the outline of attacking by the moving windows. Experiments show that the algorithm has high detection accuracy for both single-target attacking and multiple-targets attacking simultaneously.There exists deviation in calculating similarity when the score data are sparse because the traditional similarity algorithm has neglected the characteristics of the common scoring quantity of users. The algorithm we adopt takes the common scoring quantity of users and arctan function into account to adjust the similarity degree dynamically which results in high exactitude in the nearest-neighbor selection.For new users, or the users who can’t find their similar neighbor, first we classify them based on their information characteristics; then according to the FTL model (follow the leader) thoughts used in social networks analysis, we employ the prediction algorithm based on expert credibility to gain the ranking of the experts specialized in the classification of the target users; at last we predict the score of the users by both the expert credibility and their scoring. Experiments show that the algorithm can not only improve the accuracy of the prediction score and the quality of recommendation, but also alleviate the cold start problem existing in the traditional algorithm. Finally, the paper designs a movie recommendation system by integrating the two algorithms above-mentioned together which consists of the attack-detection module, recommendation module, scoring and message module, etc. This system can recommend films for users in which they may be interested and at the same time isolate attacks from suspect users effectively, and as a result lead to high system fairness and optimize the liveliness and experience of the users to a certain extent.