Research On Cold Start Recommendation Algorithm Based On Attribute-Fused Matrix Factorization

The recommendation system is designed to help users pick out information that is of interest to users from vast amounts of information.It automatically mines users'interests and preferences through history,and finally provides personalized service for each user.The collaborative filtering recommendation algorithm plays an important role in promoting the service recommendation ecology.However,in practical applications,collaborative filtering algorithms also have problems of data sparsity,cold start,mobility,and whether user privacy is violated.The most serious and most problematic issue is the cold start problem.The cold boot problem is divided into three types,namely,the user cold start problem,the item cold start problem,and the system cold start problem.Utilizing item and user information is often an effective way to solve cold start problems.For item cold start,effective use of item attribute information is one of the effective ways to solve this problem.However,for the user's cold start problem,due to the increasingly serious privacy leakage,users are more and more concerned about personal privacy,so it is often impossible to use the user's personal information as a solution to the cold start problem.Collaborative filtering plays an important role in promoting the service recommendation ecosystem,and matrix factorization technology has proven to be one of the most effective recommendation methods.This paper proposes a solution to the problem of cold start of the item and the cold start of the user in the recommendation system.For the cold start problem of new items,this paper proposes a similarity calculation method for the fusion item attributes,which is used to predict the user's preference for new services,combine the predicted preferences with the matrix factorization model,and finally use the model to Predict user ratings for new users.For the cold start recommendation of new users,this paper starts with the location information,combined with Hofstede's cultural dimension theory to calculate the cultural distance of new users,through cultural distance and matrix.The decomposition model is dynamically combined to finally complete the recommendation for new users.The main contributions of this paper are as follows:(1)For the cold start problem of new items,this paper uses the attribute information of the item to replace the traditional scoring information for similarity calculation,and uses the current user's historical record information to explore the user's preferences,through the user's preference information and item attributes.Perform a preliminary match,then incorporate the matching results into the matrix factorization model,and use the trained model to complete the recommendation for the new item.In the process of calculating user preferences,considering that the user's preferences may change over time,this paper also introduces a time penalty factor to reduce the weight of the preferences calculated by the user's long history.(2)For the cold start problem of new users,this paper uses the location information of users to calculate the cultural distance between users and services using Hofstede's cultural dimension theory without infringing the user's privacy.Through the search for inspiring services,preliminary predictions for new users are completed.The preliminary prediction results are dynamically combined with the matrix factorization technique,and the model parameters are updated by fitting the training set.Use the final model trained to complete the prediction of new users.(3)Through a large number of experiments on the real data set,it can be proved that for the cold start problem of new items,our method is obviously superior to the comparison method in terms of accuracy,scalability and data sparseness;for new users Cold start problem,the method adopted in this paper can accurately complete the recommendation of cold start users without infringing user privacy,and can also obtain very good recommendation effect in data sparse environment.