Research On Credibility-based Collaborative Filtering Recommendation Algorithm

The spread of Internet technology makes e-commerce accepted by more and more people, with which people can purchase commodities without walking out the door. Recommendation system can simulate the behavior of shopkeeper to help customer finish the process of buying. What's more, it can recommend personalized goods for customer according he's interesting. Therefore, customer will be more satisfied with the buying process, and the sales of online shop will increase. Recently, recommendation system got great achievement both on theory research and application, and it is believed to get a better development in the future, getting more attention from scholars.The effect of recommend system depends on the technology it adopted including content-based, collaborative filtering, population statistic data based, utility-based, knowledge-based technology and hybrid method. Among these technologies, collaborative filtering algorithm is regarded as the best one for its feature that caring little about attributes of objects, neither strict demand of objects. So it is widely adopted on music recommendation, movie recommendation and online-learning resource recommendation. Collaborative filtering algorithm performs quite well when the density of data is high.This thesis made a lot research on User-based collaborative filtering algorithm and Item-based collaborative filtering algorithm and found that conventional collaborative filtering algorithm doesn't give effective answer to User-favor problem, User-interesting problem and User-credibility problem. So it couldn't perform exactly and efficiently when it meets great number of goods and low average credibility of customer. Current improved algorithms on collaborative filtering can solve some problems mentioned above to some extent, yet remaining some limitation at the same time. Based on the analysis above, this thesis proposes three conceptions: Class similarity, User interesting and User credibility, then combines three conceptions properly to improve the process of recommendation. This improvement can give a good answer to the problems. The proposed algorithm firstly determines some classes for customer according his interests and the class similarity, so the range of items to be predicted is limited. When it searches for neighbors for customer, it integrates similarity between target user and neighbor and the credibility of neighbor on current class together. Consequently, neighbors found in this way are similar and trustable to target user. Experiments in this thesis contain three parts. Beside the conventional MAE and MCT which is widely used to evaluate the accuracy and efficiency of recommendation algorithm, this thesis also proposes NOV to evaluate the novelty of the algorithm. A serial of experiments proves that the proposed algorithm is more exact than other algorithms. Besides, it is more efficient than others because it combines online calculation and offline calculation together, so that the range of items to be predicted is limited. Another advantage of the proposed algorithm is that the algorithm can still perform well with relatively low MAE when average user credibility is not high while the accuracy of other algorithms becomes bad in this condition.