Research Of Recommendation System Based On Multiple Personalized Information

With the rapid development of the Internet, online shopping, listening music, watching video as well as dating appear in most people’s lives. People used to listen to music online, shopping and like to watch the video their friends recommend to them. They tend to give a rating to the item they bought and give tags for the movies they watched. A lot of personalized information of users has left by the rich and colorful lives of people on the Internet. Now, recommendation system is facing data sparsity and cold start problems, and the traditional recommendation methods are more and more difficult to meet users’increasing personalized demands. So, many researchers try to analyze different information to obtain users’interests and provide better recommendation for users. At present, the information more commonly used are users’social friends relations, the tags users give to items, the trust relationship between users and the geographical location information and so on. However, with the investigation, we find that most researches make use of only one kind of information to analyze users’interest. That is, there is social network based recommendations, recommendation based on tags or recommendation based on geographical locations. We have to admit that all these researches are great and provide valuable experiences for us. These researches tell us the different influence of each kind of information on the final recommendation. However, from another perspective, if one kind of information is helpful to the final recommendation, the combination of these kinds of information could be better for the last recommendation and may alleviate the data sparsity and cold start issues further. Based on the idea, this paper carries out two different experiments. We try to integrate two kinds of information simultaneously into two different recommendation methods separately and verify whether more information really be more helpful to improve the recommendations than only one kind of information.The paper first integrates the friend relationships and tags of users’ into user-based collaborative filtering method (cfuser) step by step and then generate three different methods, such as social network based CF (sn_rating), tag based CF (cf_tag) and social network and tag based CF (cf_sn_tag). Sn_rating is different from cf_user. In sn_rating algorithm, the direct friends of users are their nearest neighbors while in cf_user, the ratings are used to get the nearest neighbors of users. Cf_tag algorithm profile users and items based on tags by TFIDF. The preferences of users to items are computed by users tag vectors and item tag vectors. Cf_sn_ag method expands the cold start users’ profile with their friends’ popular tags and then calculates the preference of users to items by their tag vectors. Based on the first experiment, a further experiment is conduct based on the matrix factorization method and an improved matrix factorization algorithm based on the weight of tags and genres of items is proposed. The user regularization and item regularization are both considered in this algorithm. In the algorithm, users’ direct friends are not used. Tags and genres of items are used to find the similar users which are not limited to users’ direct friends. At the same time, the algorithm utilizes LSI method to recognize the issues of synonymy and polysemy of tags and gradient descent method is used to find the locally optimal solution.We verify our methods on multiple real datasets with different evaluation metrics. The datasets we use are Movielens, last.frn and Douban which was crawled from douban.com. The information these datasets include are ratings, social friends, tags, genres of items and so on. The evaluation metrics we adopt are mainly classification accuracy and prediction accuracy. For the first experiment is Top-N recommendation, we make use of classification accuracy to evaluate different methods, such as precision, recall, fl and DCG. The second experiment predicts the rating of users to items, so we explore MAE and RMSE in evaluation process. We compare our methods with others in our experiments, and the results show that multiple kinds of information indeed be more helpful to analyze the interests of users, and could provide better and more accurate recommendations to users than only one kind of information. In detail, tags are easy to get more accurate recommendations and social network information could also be helpful to uncover the real interests of users. The genres of items which users have liked before could be useful to define the area of interest of users.