Research On Hybrid Recommender Systems Based On The Dynamically Integrated Methodologies

With the vigorous development of Internet and pupularity of social networks, in-formation has been an explosive growth. Huge amounts of information are produced in the networks to provide users with many choices, which make them at loose ends. Although these information has the huge function and commercial value, it has be-come a hot area of research in the current that how to provide users with better service and generate greater value. With the advance of the Internet and the efforts of re-searchers, the recommender system technology has become a very effective way to reduce the information overload problem. It can dig out the useful information from a huge amount of internet information, to provide better service for users, such as ratings prediction, the list of recommended items and so on.In real world recommender systems, collaborative filtering (CF) techniques are the mostly applied methods. There are two typical types of CF, which are memory-based and model-based CF algorithms. In the real network, like film ratings, they can be predicted by the above two types of method. But for ratings data, the number and rat-ed preferences that users own are not same, which lead the difference of prediction accuracy for every method in the collaborative filtering. These two CF methods in fact pay attention to different parts of ratings data. Memory-based CF methods are adept at finding local similar users and perform worse when users have little ratings, while model-based CF algorithms emphasize achieving global optimization. But with the development of social network, trust relationship plays a more and more important rolein the service and can not be ignored. So the recommendation methods based on trust is proposed and it prove that they can also provide and make more accurate recommendations through the trust relationship in social networks. While trust-aware recommendation methods can make accurate recommendations using the trust rela-tionships even a user has few ratings, but they perform worse if users have few friends. Most social recommendation models combine these two methods by a static parame-ter to improve recommendation accuracy. However, due to the uneven distributions of the ratings and social relationships for each user, two types of above recommendation methods should have varying weights when make recommendations.According to the above problems, we do the following two aspects of the re-search:1) Due to the problems of data distribution imbalance, we firstly integrate a neigh-borhood approach and Probabilistic Matrix Factorization (PMF) into a hybrid CF model, DPMFNeg, which combines the advantages of memory-based methods and model-based CF algorithms. We explore the performance of our method on two test datasets--MoiveLens-100K and MoiveLens-1M. The results show that DPMFNeg performs better than other methods on those datasets in terms of Mean Absolute Error (MAE) and Root Mean Square Error (RMSE).2) Although traditional collaborative filtering algorithms are most effective recom-mendation methods when users have expressed many ratings and the above me-tioned hybrid recommendation model that we proposed can deal with the un-balance distribution of ratings data. Due to that it ignores the trust relationships of social network, we then propose a user adaptive hybrid recommendation model, which dynamically combines a trust-aware based method and Probabilistic Matrix Factorization with adaptive tradeoff parameters, named as DTMF. It can utilize the advantages of these two methods and learn combinative parameters automati-cally. We investigate our model’s performance on two social datasets-Epinions and Flixster. Experimental results show that DTMF performs better than other state-of-the-art methods on both datasets.Through our research, our proposed hybrid recommender models can alleviate the problems of uneven data distribution and imporve he accuracy of recommendation.