Research On Collaborative Filtering Algorithm Based On Distance Metric Learning

With the rapid development of computer technology,human beings have generated a lot of data while gaining convenience,they are faced with the problem of"information overload",that is,it is more difficult to mine valuable information from massive data.Recommender system was born under this background,it can obtain valuable information from a large amount of data,provide users with product information that meets their needs,and provide personalized recommendation services to users.In recommender system,the collaborative filtering algorithm is a well-known and commonly used recommendation algorithm,it models the user's historical behavior to predict the user's preference for the item.Due to the advantages of personalized recommendation and high degree of automation based on collaborative filtering algorithm,it has been widely concerned by academia and industry.Among them,matrix factorization occupies a very important position in collaborative filtering because of its high accuracy and scalability,and has become one of the most popular personalized recommendation algorithms.Matrix factorization uses dot product to predict user preferences,but dot product is simply a linear product combination and does not satisfy the triangle inequality,which may have a negative impact on the recommendation results.In view of the above shortcomings,this paper attempts to use a distance metric that satisfies the triangle inequality instead of the dot product in matrix factorization,and uses a distance metric algorithm to predict the user's preference for the item.In Chapter 3,Collaborative Filtering Combining Metric Learning and Matrix Factorization?MLMF?is proposed,it uses the distance metric algorithm to learn the location features of users and items in the same low-dimensional Euclidean space,so that users are closer to the items of interest,farther from uninteresting items.The MLMF algorithm retains the advantages of matrix factorization scalability and interpretability to a certain extent,but it is more intuitive and easier to understand than matrix factorization,and it can also effectively learn the neighborhood characteristics of users and items,MLMF mainly includes the following key points:?1?Unlike the general model,MLMF first converts the user's score to the distance deviation,and learns the positive correlation between user preferences and distance.?2?In addition,the l₂ regularization strategy used by traditional matrix factorization is not applicable to the algorithm in this paper.In order to prevent the model from overfitting and improve the recommendation accuracy,the MLMF algorithm uses a variety of regularization strategies,uses Norm clipping and de-correlation regularization?DeCov?strategy regularizes and uses Dropout to prevent overfitting.?3?In order to verify the performance of the MLMF algorithm,we compares the experimental results from multiple angles.First,experiment on multiple public datasets,and compare the accuracy of the MLMF algorithm with a variety of classic baseline algorithms.The experimental results show that the MLMF algorithm improves the accuracy of recommendations;in addition,because the algorithm can learn some neighborhood relations,therefore,this paper builds the TopN recommendation list by calculating the distance between the location features of the user and the item,the results show that the algorithm still has a good performance in TopN recommendation.?4?The T-SNE is used to reduce the dimensionality of the position feature vectors of users and items,and then visualized into a two-dimensional plane to observe the distribution of user and item features.At the same time,the article also carried out changes in the distance between users and items,and the results show that the MLMF algorithm can effectively learn user preferences for items.On the other hand,recommender system has a cold start problem,and new users and inactive users with no historical score records cannot effectively recommend.In this regard,this article also introduces a social trust relationship to alleviate the cold start problem in the recommender system.However,social trust relationships are usually directed relationships and cannot be directly used in distance metric learning.Moreover,traditional trust relationships are not constructed based on interest.Therefore,social recommendation based on trust relationships may not necessarily improve the accuracy of recommendations.In view of the above problems,Social Recommendation Algorithm Based on Distance Metric Factorization?SDMF?is proposed,it mainly includes the following:?1?The SDMF algorithm reconstructs the user's trust relationship,turning the directed social trust relationship into an undirected user relationship,which is more in line with the nature of the SDMF algorithm.?2?SDMF calculates the user's familiarity and similarity separately,and designs a new association method that comprehensively considers the familiarity and similarity of interests between users,rebuilds the user relationship matrix,and the reconstructed user relationship fully takes into account the interests of users and friends similarity,which allows the algorithm to recommend users more accurately.?3?The SDMF algorithm integrates the reconstructed user relationship into the distance metric algorithm through Co-regularization,and restricts users-items and users and friends with similar interests and hobbies.?4?Experiments on real public datasets show that SDMF can effectively learn user preferences.Compared with some classic social recommendation algorithms,SDMF has higher recommendation accuracy;in addition,on some cold-start users,SDMF still has the best performance,which shows that SDMF can also improve the cold start problem.In short,this paper improves and experiments on Collaborative Filtering Algorithm Based on Distance Metric Learning.Compared with some currently widely used recommendation algorithms,the algorithm in this paper can also have higher recommendation accuracy.Moreover,the Collaborative Filtering Algorithm Based on Distance Metric Learning also has certain scalability and interpretability,and has certain research significance.