In recommendation system,high-dimensional and sparse matrix is often used to describe the relationship between users and items.The latent factor analysis(LFA)model can estimate the missing values in a high-dimensional sparse matrix,and achieve a personalized recommendation effect for users through missing prediction.The stochastic gradient descent method is an effective strategy for solving the latent factor analysis model.However,the latent factor analysis model based on stochastic gradient descent algorithm largely depends on the selection of hyper-parameter learning rate.So,the particle swarm optimization(PSO)algorithm has been successfully applied to the adaptive tuning of hyper-parameters of the latent factor analysis model because of its fast convergence and easy implementation.However,due to the standard PSO algorithm is easy to fall into its local optimal solution,its premature convergence results in the loss of prediction accuracy of missing values.Therefore,this thesis proposes two adaptive latent factor analysis models of improved PSO algorithm to solve the above defects,and so as to solve the complex and cumbersome parameter learning rate tuning problem.The main contents of this thesis are as follows(1)An adaptive LFA model based on the combination of Adam and PSO is proposed innovatively.In the PSO algorithm,the “gradient” of each particle position is estimated.The “gradient” value of the particle position is estimated through the individual position transformation and the swarm position transformation of the particle,and then the particle position is updated by the Adam algorithm,and the the Adam-PSO optimization algorithm is constructed.The Adam-PSO-LFA model can effectively realize the adaptive tuning of the hyper-parameter learning rate of the LFA model.(2)An adaptive LFA model based on the combination of Generalized Momentum and PSO is proposed innovatively.In the evolution process of particles,the momentum and “gradient” of particle velocity are calculated,and the “gradient” of particle velocity is expressed in the form of update increment,which is compatible with the generalized momentum method,and the optimization algorithm of Generalized Momentum-PSO combination is constructed.The Generalized Momentum-PSO-LFA model can effectively realize the adaptive tuning of the hyper-parameter learning rate of the LFA model.(3)In this thesis,a large number of experiments on four high-dimensional and sparse matrices generated by industrial applications show that the two models proposed in this thesis achieve the adaptive tuning of the super parameter learning rate of the LFA model without increasing the computational burden,and show good performance in terms of prediction accuracy and convergence speed. |