Research On Cold Start Problem In Personalized Recommender Systems

In recent years,with the development of new technologies such as big data,the internet of things and artificial intelligence,the amount of global data has grown dramatically.The abundant value and great potential contained in big data provide convenience for human beings,but also bring information overload problem.In order to help users quickly obtain valuable information from massive information,personalized recommender systems were born.With the rapid growth of user scale and item size in a recommender system,the user-item rating matrix becomes sparse,resulting in a decrease in the recommendation accuracy of the traditional collaborative filtering recommendation algorithms.When a new user or a new item is added to the system,due to the lack of its historical rating information,it can't be recommended,which causes cold start problem in the recommender system and affects user experience.In this paper,by analyzing the causes of cold start problems,knowledge graph is introduced into recommender system,then vector space model and random walk strategy are used to calculate the similarity between items,finally the accuracy of the recommendation results is improved.The main work of this paper includes:(1)By analyzing domain-specific knowledge graph construction technology,this paper designs and implements a method of constructing movie domain knowledge graph based on Linked Open Data(LOD).We analyze and study the related knowledge in movie field.Then we extract the knowledge ontology classes and relationships between entities in movie field,and accordingly generate the corresponding triples based on the Linked Open Data.Finally,we complete the construction of movie domain knowledge graph,and use a graph database to visualize it.(2)A similarity calculation method based on domain knowledge graph combining vector space model and random walk strategy(VSM-RW)is proposed to solve the cold start problem.In the knowledge graph,vector space model is used to calculate the similarity between items.These values are used as the weight of the edge to construct the transition probability matrix.Then with the help of random walk,the similar relationship between the items is further transmitted to obtain more accurate similarity.Thereby finding more accurate similar items for the target item.This method does not require historical rating information and adds rich semantic information to the recommender system through the knowledge graph,which makes each entity can calculate the similarity with other entities.It alleviates the impact of cold start problem on recommendation system.The experimental results show that VSM-RW algorithm has certain performance improvement compared with ItemCF,BPMF,and SVDpp algorithms.(3)Combining the method of using user-item rating matrix to calculate similarity in the traditional collaborative filtering recommendation algorithms with the VSM-RW algorithm,we describe the similarity between items from multiple angles,which help us further improve the recommendation accuracy.