The Research Of Personalized Recommender Algorithm Based On Item Cloud

In recent years, network data has surged exponentially with the rapid development of the science technology, which leads to very serious information overload issue. Thus, for the users, it is an important task to acquire their necessary information accurately and quickly from the mass of information in a short time. Recommender system as an important tool for information filtering has been widely used in many E-commerce sites. However, with the rapid development of science technology, the number of online shopping users has increased sharply, and new products continue coming out, which lead to the problems of cold start,extreme sparsity, scalability and other issues become increasingly prominent. In order to overcome the shortcomings of traditional recommender algorithms, this paper uses qualitative analysis and fuzzy cluster to build a complete personalized recommender system based on the item cloud. This system combine the advantages of cloud model and ordered rank cluster skillfully, and then propose a new algorithm of ordered rank cluster collaborative filtering recommendation based on item cloud.The personalized recommender system designed in this paper can be divided into three modules: data ore-processing, ordered rank clustering, prediction and recommendation. In the data ore-processing module, this paper uses cloud model to fit the distribution of the different items and their statistical characteristics, and then generate some missing values to fill the original user-item rating matrix by cloud generator; Next, this algorithm combined ordered rank cluster with recommender system, then do the preliminary classification among the whole item cloud according to the new rank criteria to output the ordered cloud vector, which is the foundation of the subsequent similarity calculation; Additionally, this paper adopts the "clustering-recommendation" mode to analyze the relationship between item cloud in the cluster, and then acquire the precise recommendations for the target users and items.Compared with the traditional recommendation algorithms, the new method has mainly made the following improvements:First, in this paper, it gives a reasonable assumption of the data missing mechanism by analyzing the raw data matrix and data missing principle, then use the cloud model to fit thedata distribution.Second, in order to improve the extremely sparsity, it proposes two kinds of data filling algorithms: One is based on the item distribution; the other is weighted by user rating reliability. Comparing two different filling mechanisms to inspect the influence of the data sparsity and cold start problems to recommend results.Third, this algorithm use empirical distribution function to find the critical points to determine the threshold of the rank density function, then transform the continue random numbers into discrete values, which is more scientifically to reflect the characteristics of the item rating.Forth, this paper is the first to combine the ordered rank cluster with recommender system, according to the newly defined sort criteria to item cloud not only can improve the recommendation accuracy but also greatly reducing the computation time. In the traditional recommendation algorithms require nn?2)1( times to calculate the similarity, but new algorithm only need n?1 times.Finally, in order to verify the effectiveness of the ordered rank cluster collaborative filtering recommender algorithm based on item cloud, the R language software is used to do the experiments on the Movie Lens Data and Jester Joke Data in this paper. The empirical results show that: the proposed new mechanism for filling data can effectively improve the extreme sparsity of the recommend system; ordered rank cluster algorithm based on the item cloud contain the uncertainty of the cloud model, which not only can improve cold start problem of the recommender system, but also can locate the homogeneous items more accurate and increase the recommendation precision; In addition, ordered rank cluster can reduce the computational complexity, enhance the recommendation system scalability.Therefore, the proposed algorithm has great significance for the personalized recommender systems research and development at present.