Research On A New Feedback-control Algorithm In Collaborative Filtering

With the development the SNS and the Internet community, the information overload problem becomes much more serious. Recommender system is a key way to solve the information overload, and is a mature, good effective technology. However, the situation that the number of users and items is bigger and bigger makes the recommender system calculate the similarity matrix all the time and the work will take a long time for every time, and this is the scalability problem in collaborative filtering.The methods to solve the scalability problem in collaborative filtering are all based on the fixed model. Calculating the similarity wastes a lot of time in the CF algorithm. How to decrease the time of calculating the similarity is a new orient in CF research. In this paper, we create a novel way to solve the scalability problem in the model with the variable parameter using prediction errors which can be calculated by comparing the system prediction and the users'feedback. Prediction errors can be used by the feedback-control algorithm to modify the collaborative filtering model in order to change the result of the recommendation without calculating the similarity again. We do some work as following for the feedback-control based collaborative filtering.1. We review the development of the recommender system and collaborative, conclude the theory of collaborative filtering and its existing problems in order to lay a foundation for the following work.2. We establish the feedback-control based collaborative filtering model. The prediction error is used to control the collaborative filtering by using the feedback-control theory, and this work achieves the goal of feedback controlling the collaborative filtering.3. Based on the feedback-control based collaborative filtering model, we use the feedback-control principle and machine learning method to design two different feedback-control algorithms which modify the parameter of collaborative filtering, convert the fixed model into variable model, improves the results of the recommendation for the future time, decrease the system responds time.4. We apply the feedback-control based collaborative filtering algorithm in the Digg dataset. The experiment results show that the collaborative filtering based on the feedback-control algorithm is better than the traditional algorithm no matter whether in the time complexity and the precise based on MAE with little time increase.