The Research Of Recommendation Algorithm Based On Relation And Content

With the increasing of information, we have been in era with mass data. Data, knowledge, goods, friend relationship, even micro-blog or intelligence mobile phone which connect Internet all have a lot of users’information, but for individual users, it is difficult to find the useful knowledge from the mass information.Fortunately, recommender systems emerges and tries to mine useful personalized information for individual by taking advantage of the knowledge about user selection or similarity between users or items.Its aim is to filter information.To date, a lot of recommendation systems have been designed and developed, such as the recommendation of products on ecommerce site, or the recommendation of friends in micro-blog and so on. These recommendation systems in engineering have compromise for time and the algorithmic complexity, which results in that the recommendation results sometimes do not meet the users’ requirements. Usually, the original information in recommender systems cover the following parts:the evaluation information on products from users, the product content information, user’s own label information and so on. Collaborative filtering is a typical recommender algorithm which only uses the first kind of information. Meanwhile, the collaborative filtering works based on a hypothesis that users with same taste will approach to selecting same products. This hypothesis seems to be reasonable, but it prefers to recommend hot goods, then the unpopular with high quality products will be rarely recommended. For content-based recommendation algorithms, they use the product content. The main idea is that if product A is selected by the user, the user will like other products which are similar to product A.Obviously, these kinds of information are all useful for recommendation systems. Thus, in this thesis, we combine the existing original information, and propose a bi-part graph to represent all information, and implement recommendation by applying the random walk with restart algorithm on this graph. Furthermore, we implement logistic regression algorithm to fast the recommendation process. A series of experiments on real world data set CiteULike have shown the performance of the proposed methods.