A novel personalized recommendation algorithm by exploiting individual trust and item's similarities A novel personalized recommendation algorithm

Recently, incorporating user information, social network information, item information and user ratings to improve recommendation performance has attracted great attention. However, most of the existing recommendation methods suffer from the following problems: (1) They only use user information, the item's information, or user ratings to make recommendations, thus, they demonstrate low recommendation accuracy. (2) They employ decision trees to execute user information and item information, therefore, they ignore the correlation between attributes, resulting in a decrease in recommendation performance. (3) It is difficult to cope with the problem of data sparsity. To address these problems, we propose a novel personalized recommendation algorithm by exploiting the individual trust and item's similarities, named PRAITIS. In PRAITIS framework, the zero-mean spherical Gaussian prior is first applied to the item feature vector, user feature vector, and user-rating-data matrix to obtain their latent feature space and the user-rating-data space. Then, Bayesian inference is used to get the posterior probability of potential features. Finally, under the condition that the hyper-parameters are fixed, the framework of the algorithm is obtained by maximizing the log-posterior probability of three potential features. In order to verify the effectiveness of the proposed algorithm, a series of experiments done on two real-world datasets (e.g., Douban and Epinions) show that the proposed algorithm is superior to the state-of-the-art recommendation algorithms in terms of recommendation accuracy and quality.