Research On Hybrid Recommendation Algorithm Based On Item Description And User Behavior

Today,the rapid development of information technology provides users with a large amount of information.Users can find all kinds of information through the Internet without leaving their homes.This satisfies users' information needs.However,with the increasing supply of resources on the Internet,information has exploded and caused "information overload." This has created huge challenges for users to obtain information they are interested in.An important way to solve information overload is personalized recommendation technology,which can help users to obtain information that they are really interested in.This thesis studies the background and state-of-the-art of the recommendation system,and focuses on exploring the problems and challenges in the recommendation system.In order to solve the problem of cold start and data sparsity in collaborative filtering algorithms,this thesis proposes the integration of content information and user behavior which is a hybrid recommendation algorithm.The main work is as follows:Analyze Saveski et al.'s hybrid recommendation algorithm for item description and user behavior.This scheme is executed when the default item description information is rich enough,but there are many items describing the short and small situations in actual application scenarios.To address this problem,a kind of improved item description information is proposed,to construct user matrix and item description matrix respectively,and then to fuse the recommendation scheme of user behavior.Experiments are carried out on MovieLens dataset,and have verified the new scheme has positive effects for the cold start problem and the recommendation.Propose a hybrid recommendation scheme based on user behavior and item type description combined with topic vectors.For items with limited description,select the key features from the item's tag information as an extension of the item description information,and then use the theme vector model to calculate new one.The thesis describes the theme vector,and thus calculates the similarity of the item,and is integrated with the similarity calculated based on the user behavior,thereby providingthe user with a recommendation.The results of the experiment verify that when the description of the item is limited,the method of augmenting item description information and adopting the theme vector has an improvement over the baseline method in terms of accuracy and cold start of the item.