Study On Collaborative Filtering Based On Heterogeneous Information Network Representation

With the rapid development of Internet technology,data in the network is being generated at a rate that we cannot imagine.The phenomenon of information overload is becoming increasingly serious.In order to explore and utilize information more efficiently,personalized recommendation technology has been put forward and widely researched and developed.In recent years,the research on Heterogeneous Information Network(HIN)has also aroused the attention of the majority of scholars.Many data mining tasks are also gradually using HIN to explore.HIN provides a new solution for the recommendation system.In the recommendation system,various types of objects(such as users,movies,directors,actors in movie recommendations,etc.)and rich relationships between them are often included.The heterogeneous information network can express the information precisely.HIN can collect this comprehensive,rich and valuable semantic information,resulting in recommendations better.The recommendation technology based on heterogeneous information networks can solve the key problems such as data sparsity and scalability in traditional recommendation systems to some extent.Although the previous research on heterogeneous information network could consider different types of objects together,the weights between objects are often ignored.Such weights information could reflect the closeness of the relationship between objects(such as user rate a film as a number in 1 to 5.The higher the score,the more the user prefers and the degree of relationship should be closer),which makes the inter-object semantics more interpretable.Therefore,when constructing a heterogeneous information network,this paper will also display the weight information and propose a random walk of the weighted meta-path.Besides,in addition to the types of objects in the recommendation system,objects may also contain rich attribute information.These attribute information also play an important role in the recommendation process.If we can use these attribute information properly,the effect of recommendation will be improved.Thus,this paper proposes AttrHIN2 vec,a heterogeneous information network representation learning method with attributes.This method combines the topological structure of nodes in heterogeneous information network with attribute information of nodes,and obtains more abundant feature data through heterogeneous Skip-gram model.Finally,these feature data are trained using BP neural network to generate a collaborative filtering recommendation model.The experimental data set of this paper adopts MovieLens-1M.The experiment is divided into two steps.The first step is to construct a heterogeneous information network which has weight and attribute information for the objects in the recommendation system,and use the AttrHIN2 vec method proposed in this paper to extract the potential features of the object;In the second step,potential features are used in BP neural networks to construct a ratring prediction model.Through experimental comparison and analysis,it is verified that the proposed method can improve the recommendation results effectively.