Research On Recommendation Algorithm Based On Spectral Clustering And Momentum Optimization

With the rapid development and wide popularization of the Internet,the scale of Internet users and social content has expanded explosively.It's gradually entered the era of information overload.The recommendation system reflects a strong advantage in information filtering and screening.It provides personalized services to users.This system can effectively improve the customer experience and improve the service quality of enterprises.So,it can bring huge economic benefits to enterprises and create value for users at the same time.Currently,recommendation systems can be divided into three categories: Content Based Recommender System(CB),Collaborative Filtering Recommender System(CF)and Hybrid Recommender Systems.CF has been widely applied,due to its simple form and good recommendation effect.This dissertation mainly analyzes and optimizes the Latent Factor Model(LFM)which is one of the CF.This dissertation makes an analysis of LFM.It continuously iterates to optimize and obtain the solution by the Stochastic Gradient Descent algorithm(SGD).Some scholars proposed a batch learning algorithm with parameters that added impulse and intermediate momentum in SGD to improve the learning rate and reduce the learning time.In this dissertation,we introduce Nesterov Accelerated Gradient(NAG),which is widely used in the field of deep learning,into LFM Algorithm,propose an optimizing LFM algorithm based on NAG(N-LFM).This algorithm designs a correction factor on the LFM.It can accelerate the iterative process of matrix reduction,achieve accelerated convergence,and improve the accuracy of the recommendation algorithm.In addition,considering the problem of recommendation efficiency,author adds the Spectral Clustering algorithm(SC)in the recommendation.Traditional collaborative filtering recommendation technology based on K-Means clustering algorithm to cluster users or projects.While this dissertation adopts spectral clustering algorithm to solve the problem of unstable effect.Through adding the SC,the space of searching the neighboring users is reduced.In this way,the computation is effectively reduced.And the sparsity,scalability and instantaneity of the recommendation system is improved.So,we propose the LFM Algorithm Optimization based on NAG(CF-NLFMSC)in this dissertation.The experiment conducts in the Movie Lens data set and proves that the improving algorithm CF-NLFMSC can achieve higher accuracy.It improves the accuracy and recall by 2.22% and 1.26% respectively,compared with the LFM in the experiment of Movie Lens100 K data set.The coverage rate increased by 4.02%.The improvement of coverage rate indicates that more products are recommended for users,which means that the user's satisfaction is more likely to be improved.The popularity is reduced by 1.09.It means that the algorithm can recommend more good items which is less popular and has a better performance in mining less popular items.In addition,the algorithm still performed well in the Movie Lens data sets of 1M,Film Trust data sets and Netflix data sets.The algorithm accuracy and recall rate in Movie Lens1 M data set achieved the improvement to 2.94% and 14.64% respectively.The accuracy and recall rate in Film Trust dataset's top-10 recommendations increased by 6.79% and 10.34%,and increased by 1.5% and 1.66% respectively in Netflix dataset.The experiments verify the algorithm is feasible and effective.So,it has the certain research value.