Research On The Improvements Of Collaborative Filtering Personalized Recommendation Algorithm

Personalized recommendation technology is an effective tool to solve the problem ofinformation overload in the information age, it can be used as a stand-alone tool to helppeople solve problems, and can also be combined with some of the existing technologies,such as search engines, classified directory, so that the services they provide like the icing onthe cake. Common personalized recommendation technologies are the following: thecontent-based recommendation, collaborative filtering algorithm and the hybridrecommendation techniques and so on, the collaborative filtering technology is one of themost successful one. It is mainly based on the characteristics of the social attributes of theusers or objects to generate personalized recommendations by analyzing user’s historicalbehavior. Because the recommended procedure requires only the historical behaviorinformation of the user, and has nothing to do with the content attributes of the object, so it isvery simple to implement, and it has been widely used in e-commerce, movies, music services,individualized reading, search field and so on. Therefore, the study of the collaborativefiltering algorithm has great commercial value. But collaborative filtering algorithm in theapplication process is facing many problems, such as the scalability of the algorithm, coldstart problems, recommended precision, it is also worthy of our in-depth study.In order to improve the recommended precision of collaborative filtering algorithm,scholars through improving user similarity calculation method to effectively measure interestin the degree of consistency between users or objects, selecting the appropriate penalty factorto improve the accuracy of the similarity between users or objects in accordance with thecharacteristics of the data set, and combining with other recommended technologies toimprove the accuracy of recommendation. However, these methods have ignored how toselect the neighbors, as the final recommendations are generated based on the nearestneighbors, thus their quality is also directly determine the accuracy of the recommendationresult, and this is verified by experiment in the paper. This paper carried out a detailed studyto select the nearest neighbor:First, this paper proposes two close neighbors assessment indicators: the close neighborsuser/project group similarity and reference nearest neighbor ratio to measure the quality ofthe nearest neighbor. And we find the drawbacks of the calculating process by traditionalcollaborative filtering through experiments.Second, the traditional collaborative filtering algorithm to select a close neighbor group of either strong correlation, but it does not participate in the prediction score calculation process,or is able to participate in the calculation to predict ratings, but is not the real close neighborgroup to the target user/project with low correlation, led to some close neighbors in theprediction score calculation process are not to play a positive. For the lack of originalneighbor selection method, the paper proposes a dual threshold nearest neighbor lookupmethod. It not only considers how many close neighbors involved in the calculation, but alsoconsiders the nearest neighbor group correlation. And we apply the new way of finding closeneighbor group in the User-based and Item-based collaborative filtering, formed theDT-UBCF and the DT-IBCF algorithm.Then, based on the idea of dual threshold nearest neighbor group search method, the paperproposed searching nearest neighbor group by sampling strategy, it improves therecommendation accuracy of user-based collaborative filtering algorithm when application itin online, but the time overhead has increased.Finally, we verify the improved collaborative filtering algorithms in the dataset: Movielens,and the experimental results show that the recommendation accuracy of the DT-UBCF andDT-IBCF is better than traditional UBCF and of IBCF algorithms.