Research And Application On Anti-attack Collaborative Filtering Algorithm Based On User Credibility

Collaborative filtering technology which has been most widely used is one ofthe most mature personalized recommendation technology. It tries to calculate therelationship between users or items by analyzing users’ historical records, in orderto find the nearest neighbors for each user or project. Then, the nearest neighborsare used to recommend items for the target user, in order to achieve the purpose ofmining the user’s potential interest. But it has many problems such as datasparseness, cold start, scalability, easy to be attacked and so on. In this thesis, thesparseness and the easy to be attacked problem have been studied in-depth toimprove the traditional collaborative filtering algorithms, in order to have a higheraccuracy recommendation in the sparse data case, and to have better resistancewhen face a variety of common attacks.First of all, we pointed out that those algorithms, which use the missing datato calculate the similarity, will face the sparseness problem directly. While thepearson correlation similarity does not use the missing data, it does not face thesparseness problem directly. But, for the number of the common rated projects isnot the same, it has to face the problem called ‘volatility factor’.The distributionof the similarity calculation has been analyzed in different volatility factorsituations. Then we proposed a simple way to eliminate volatility-factor’sinfluence on similarity calculation according to the distribution. Next, thecollaborative filtering algorithms(including user-based and item-basedcollaborative filtering algorithms) which had been eliminated volatility-factor’sinfluence on similarity calculation were used to validate on Movielens data sets.From the experiment, we found that the new algorithms had a greatly improve onthe recommendation accuracy.Next, we proposed a simple method based on statistical methods to calculatethe users’ credibility, to strengthen the anti-attack capability of collaborativefiltering algorithm.Then, the users’ trustworthiness and traditional collaborative filteringalgorithms such as user-based and item-based collaborative filtering algorithmwere combined, and the volatility-factor’s influence was eliminated.We conductedan experiment to validate the algorithm’s recommended capacity and anti-attack capability. The result of the experiment showed that the improved algorithm had abetter performance than the original algorithm both in the recommendationaccuracy and the anti-attack capability.At last, the research in this thesis was applied to the project named "Thefourth side employment information platform".