Based On Incremental Update Of The Adaptive Collaborative Filtering Algorithm

Useful information has been filtered out by recommendation system for people based on users'preference on the items and using knowledge discovery techniques recommends the contents for the users who may be most interested in these contents. As one of common knowledge discovery techniques, collaborative filtering is a personalized recommendation system's main tool. With the more and more complex site structure of recommended system, the increasing number of items and users, collaborative filtering algorithm exposed many problems in the implementation process. This thesis has researched three problems including sparse data processing, forecasting accuracy and scalability.Incremental updating adaptive collaborative filtering algorithm has been proposed, combing user-based and item-based collaborative filtering algorithm, while adding incremental updating mechanism. This algorithm has composed of two process models, adaptive collaborative filtering model and incremental updating model. In adaptive collaborative filtering, firstly, it has combined the item entity similarity and item attribute data to extend the user common data maximum. Secondly, the similarity adjustment factor and the threshold can increase the similarity between the objects which has been chosen as nearest neighbors and target object. Lastly, the algorithm has set a weighting factor to decide how to balance the relationship of the predicted scores between user-based method and item-based method. Incremental updating mechanism can adapt to the changing needs of the user interest, so that the mechanism can update the similarity between new rated item and other items in real time through a smaller system calculation after the user has submitted the new rates. Therefore, the mechanism which has eliminated the cost of scanning all items space at every time of the traditional method updating the recommended data can effectively improve scalability.The proposed algorithm has been verified on the open data sets MovieLens. The experimental results have shown that the IUACF algorithm is superior to the traditional user-based collaborative filtering algorithm and item-based collaborative filtering algorithms on the prediction accuracy and scalability. The experimental results have also tested out the optimal settings of the similarity adjustment factor and the threshold.