Research On Hybrid Recommendation Algorithms Based On Slope One Scheme

The innovation of information technology brings us into the information society,but the problem of "information overload" limits the better use of information.After the advent of search engines,personal recommendation system effectively alleviated the "information overload" problem.Collaborative filtering algorithm is currently the most widely used recommendation algorithm,but single collaborative algorithm has the problems of sparse data,scalability,cold start,explainability.These issues restrict the accuracy of recommendations system.In order to optimize the recommendation algorithm,more and more auxiliary information and hybrid models are introduced into the recommendation system.Content information helps to relate user to recommended item which can meet the personalized needs of users,and a hybrid model helps improve the performance of the recommendation system.The weighted Slope One algorithm belongs to memory-based collaborative filtering methods.It uses linear regression models to predict ratings.The greatest advantage of it is easy to implement,efficiency and high accuracy.However,Slope One algorithm fails to consider the internal connections between users and items,thereby affecting the quality of recommendations.For the above problems,based on the Slope One algorithm,in order to improve the performance of the recommendation system.This paper proposes corresponding improvement algorithms from the perspective of matrix factorization and content-based recommendation.This paper mainly introduces three methods: Slope One algorithm based on non-negative matrix factorization,Slope One algorithm based on item similarity,and Slope One algorithm based on tag genome.The main research contents are as follows:(1)In non-negative matrix factorization,NMF based on manifold learning has a good effect on improving clustering performance,and projection non-negative matrix factorization can ensure sparse expression.This paper proposes a Slope One method based on neighborhood preserving projective non-negative matrix factorization.We impose a nearest neighbor graph as regularization term on the projective nonnegative matrix factorization to preserve the neighbor relationship between data points in the original high-dimensional ambient space while improving the sparsity of the rating matrix,so that after dimensionality reduction,we can select more accurate nearest neighbors.Experiments with MovieLens dataset show it can improve accuracy and efficiency.(2)Collaborative filtering algorithm may introduce dissimilar users when selecting user neighbors,in order to alleviate this problem,we research on an improved nearest neighbor selection method,which exploit item similarity as weights when calculating user similarity.In the view of the problem of insufficient feature information of items,we ameliorate the calculation method of item similarity,exploiting more item attribute features as auxiliary information to calculate item attribute similarity as well as avoiding increasing computational complexity.Experiments with MovieLens dataset show it can alleviate cold start problems and improve explainability.(3)Tags are important intermediary entities relate user to recommended item.Different from traditional tagging systems,this paper introduces tag genome:a data structure that extends the traditional tagging model to provide enhanced forms of user interaction.It reflects the content similarity between items.For the purpose of ameliorating the accuracy of the recommendation system,we propose a hybrid algorithm combining tag information and item attributes.The algorithm defines the relevance of the tags and items as the tag similarity,and then integrates the predicted ratings by tags into the finally predicted rating.Experiments with different types of the MovieLens dataset show it can alleviate cold start problems and improve accuracy better.