The Research On Graph Based Recommendation Model Importing User's Attention

With the fast development of Internet information,traditional retrieving techniques can no longer satisfy customers' needs of quickly and precisely acquiring desirable information.For this reason,personalized recommendation systems have attracted the interests of big internet companies,which have already put them into real application.As far as we know,almost all the recommendation systems depend on user's click log or visiting pages to guide recommendation.This kind of methods can only recommend the pages containing the similar contents to the information already visited by users or the pages containing some common information viewed by most of users.In fact,users are more inclined to obtain personalized recommendation results,which can satisfy their specific interests and preferences.The present personalized recommendation methods mostly rely on searching log,clicking log or the tags provided by users to identify user's characteristics(or namely,user model).Most of personalized recommendation methods treat users independently and can only take the features from one single user as the supports for recommendation,without combining the friendship among users that widely exist(namely,users' attention).Traditional personalized recommendation methods own the following limitations: 1)Hard to involve all kinds of interests and preferences from users: the various interests and preferences of users can hardly be fully express by the limited features.2)Hard to acquire the latent interests of users.3)Hard to integrate prior knowledge.4)Cannot change the recommendation results automatically along with the changes of user's interests.According to the above problems,this paper proposes a novel graph based recommendation model that introduces users' attention and validates its effectiveness in experiments.The proposed model mainly improves the problems existed in the traditional recommendation model from the following points: 1)Integrate the relation calculation between commodities and the similarity calculation between the reviews into an iterative process to obtain more reasonable relation calculating results between commodities.Experimental results show that after taking the relevance among products in to consideration,recommendation model can recommend relevant products to the users.2)Propose a graph based recommendation model,which just takes the mutual attention between users into consideration,and then regard users and commodities as nodes in the graph.Besides,it calculates the centrality of each node via the distribution of nodes.This centrality is treated as the probability of recommending the item in one node to the user in the same node.This model can fully exploit the user's interest by means of mutual attention between users.3)By setting the initial weight of the node to introduce prior knowledge into the recommendation model,and by adjusting the structure of the model according to prior knowledge to satisfy the user's preference,the recommendation results can be adjusted by the user's prior knowledge and satisfy the user's interest preferences.4)In order to improve the efficiency of our model,this paper proposes a local based calculation scheme,which can quickly change the recommendation results to improve the efficiency when the commodity dynamically update and the user's interests continue changing.5)The above method deals with the recommendation problem only via the structural network formed by user and commodity.To improve recommendation performance,this paper also imports textual features of users.From a multi-view perspective,it can achieve more accurate and reasonable recommending results.In summary,different from the recommendation system widely adopted by most of popular internet companies,the proposed graph based recommendation model presented by this paper first considers the interests and preferences exploited from the attentions among users.Besides,it can make use of the relevance between items for recommendation.Experimental results show that the proposed model is superior to the traditional recommendation models both in terms of accuracy and efficiency.At last,for verifying the performance of our proposed recommendation model,we propose a topic recommendation system.This system makes the use of the graph based model to predict the heat topic for users.System results demonstrate that,after using graph based model,it can detect the latent interests of users and recommend accurate topics to users.