Research On Implicit Feedback Recommendation Algorithms Based On Heterogeneous Information Network

In social media,the main task of the recommendation system is to comprehensively analyze the expression of users' preferences based on users' historical records,and find information that users may be interested in,so as to improve users' experience in using internet services.Recommendation system appears frequently on the homepage of major media,which not only improves the service level of the whole media application,but also promotes the application research of recommendation system.This thesis builds a heterogeneous information network based on the historical data of users in social media,considers real-life behaviors and habits to design a network traversal strategy,and comprehensively analyzes the social features of users in social networks and the expression of interests of users in multiple fields for the recommendation,to alleviate the data sparsity problem in the recommendation system.The main contributions of this thesis are as follows:1.We propose a social recommendation model based on a heterogeneous information network.First,based on the user's historical behavior records,establish a heterogeneous information network model,and define the explicit social relationship and the implicit social relationship,and then analyze the impact of different social relationships on the user.For explicit social relations,we propose a social influence calculation method to converts direct social relations into a list of social items for each user.For implicit social,the social characteristics expressed by users are analyzed based on heterogeneous network embedding method.Combined with item characteristics,the social characteristics are expressed as user-item feature relations matrix to analyze the influence of implicit social relationships on the recommendation list.Finally,combine the two social relationships and establish a heterogeneous social recommendation model to alleviate the problem of data sparsity.According to the experimental results on the three data sets,the model proposed in this thesis has a better recommendation effect than the benchmark methods.2.We propose a cross-domain recommendation model based on a heterogeneous information network.This thesis considers that users will express their hobbies in different fields,as well as the commonness and differences between different domains.This thesis first establishes a multi-domain heterogeneous information network model based on the historical behavior records of two domains with common users.Then,the heterogeneous network embedding algorithm is used to establish a unique feature distribution for each user in different fields,and analyze the user's expression in different fields.This thesis proposes a feature filtering method based on the differences and commonalities between features,so that the source domain information can serve the recommendation task of the target domain,so as to alleviate the data sparsity of users in a single domain.We test on two data-sets with common users,compare with the classic methods,the model proposed in this thesis shows the best experimental effect.