Algorithm Design And Analysis Of Second-Order Latent Factor Analysis Model Via Parallel Particle Swarm Optimization

In recommendation systems,the relationship between users and items is usually represented by a high-dimensional and incomplete matrix.Due to the high-dimensional and incomplete nature of the matrix,it is difficult to extract the inherent relationship between users and items directly from the high-dimensional and incomplete matrix.Latent factor analysis methods can approximate the original matrix through low-dimensional approximation,effectively achieving low-rank storage of high-dimensional incomplete matrices and predicting massive missing values.In order to more accurately predict the missing data in high-dimensional and incomplete matrices in sparse data scenarios,secondorder latent factor analysis models based on Hessian-free optimization,among others,are gradually gaining attention.However,the low-rank prediction performance of second-order latent factor analysis models based on Hessian-free optimization is largely affected by the selection of model hyper-parameter and cannot adapt to various data scenarios.Particle swarm optimization is widely used for hyper-parameter optimization because it can solve the optimization problem without explicitly representing the objective function.However,due to the slow convergence speed and the tendency to prematurely converge into local optima,directly using particle swarm optimization algorithm to optimize the hyper-parameter of the second-order latent factor analysis model can lead to more time costs and loss of prediction accuracy.Therefore,this paper proposes a second-order latent factor analysis model based on distributed parallel multi-phase and multi-elitist learning strategy particle swarm optimization,which allows the model to better adapt to various data scenarios.The main research contents of this paper are as follows:(1)A second-order latent factor analysis model based on distributed parallel particle swarm optimization is proposed.The model optimizes the hyper-parameter of the secondorder latent factor analysis model with distributed parallel particle swarm optimization algorithm.The position of each particle in the particle swarm is an abstract description of a set of hyper-parameter in the second-order latent factor analysis model.By iteratively sharing the best experience among individuals and the particle swarm,the hyper-parameter of the second-order latent factor analysis model can adapt to various data environments.(2)A second-order latent factor analysis model based on distributed parallel multi-phase and multi-elitist learning strategy particle swarm optimization is proposed.The model represents the hyper-parameter and latent space matrix of the second-order latent factor analysis model with the positions of particles.Different learning strategies are adopted in different update stages to iteratively optimize the hyper-parameter of the second-order latent factor analysis model,in order to avoid the dilemma of premature convergence.The idea of ensemble learning is incorporated to integrate the latent feature matrix of particles at different stages,in order to obtain more accurate prediction ability of the model for missing data in high-dimensional incomplete matrices.(3)Through experiments on high-dimensional incomplete matrices generated in multiple real industrial scenarios,the experimental results show that the second-order latent factor analysis model based on distributed parallel particle swarm optimization with multi-phase and multi-elitist learning strategy proposed in this paper can achieve adaptive optimization of hyper-parameter and has good prediction performance and convergence speed for missing data in high-dimensional and incomplete matrices.