The Research Of Personalized Recommender Algorithms Based On Correlation And Associated Properties Preference

The development of the Internet has generated a lot of information available and therefore many choices for consumers. Personalized recommendation system help users find the information they are interested in from a big amount of information. We use the evaluation of users on resources to calculate the similarities between users, and then use evaluation of similar users on the resource to predict current user preference. In daily life, people often choose goods or movies based on the recommendation of friends around them. Based on this idea, collaborative filtering is used in network information service, using the evaluation of neighbor with similar preference to make recommendation for the target user. The collaborative filtering algorithms have been widely used in books, movies, music and other online recommendation systems. Group lens, Ringo, Amazon are using such recommendation methods.This article first used collaborative-based approach to make improvement in three aspects. Most of all, as the number of users increases, the complexity of the traditional recommendation algorithm becomes obvious, real-time recommendations are seriously affected. So firstly we use the idea of cloud model, identifying users with similar cloud characteristics to the target user, which had reduced the scope of the search for neighbors and improved the efficiency of the recommendation. Secondly, the score data sparse problem in the traditional algorithm and the problem of rare common rated items makes similarity between users the lack of credibility. To solve this problem, the concept of correlation is proposed. We reduce the scope of the search for neighbors again by finding the users which have a certain number of common rated items with the target user, which can thus ensure a certain correlation between users, so that the obtained similarity between users has greater credibility. Experiment show that the method not only improves the efficiency of the recommendation, but also has a higher recommendation accuracy.Moreover, based on content-based methods, we made two improvements over the traditional method from the perspective of the user preference. Most of all, the traditional method in the aspect of user preference only qualitatively analyze if the user likes an item, then according to preferences they give some weight to the rating. In this regard, we have made improvements by making more specific classification on the films the user has rated, which is regarded as the user's preference model. Then, using the Item Properties based on Bayesian probability theory, we calculate the probability of the item belonging to each classification. Then we combine the corresponding score in every classes and quantitatively analyze the user preference on the item. Secondly, the traditional method did an independent study on item properties and analyze whether the user likes the property. The reality there is a certain relationship of human's preference on the item's properties.For example the user likes movies with properties contain both science fiction and love. Based on this idea, we studied user preferences two item properties and predicted preference using association property preference (AP) theory.Considering the multiple factors that affect user ratings on items, to rely solely on the preference is more difficult to accurately predict the true score. We considered the impact of public ratings on the recommendation, then AP score and the general score in a weighted way, and the optimal weights obtained experimentally to predict the final score.