High-dimensional sparse matrices contain rich information such as user preferences and community tendencies.The Latent Factor analysis model has been proven to efficiently extract the useful information from high-dimensional sparse matrices.The stochastic gradient descent algorithm is an efficient algorithm for constructing the Latent Factor analysis model.Since the optimization algorithm of stochastic gradient descent algorithm can accelerate the convergence of the model and improve the prediction accuracy of the model,this paper proposes five kinds of Latent Factor analysis optimization models based on the standard stochastic gradient descent algorithm.In order to get better prediction accuracy,this paper will aggregate the all Latent factor optimization models to an ensemble.The main research contents of this paper are as follows:（1）Two methods based on FOBOS and SPGD are proposed to construct the elastic network latent factor analysis models.In the model construction,L_{1} and L_{2} are used to constrain the objective function.It can not only prevent the model from over-fitting and improve the sparsity of the model data but also increases the accuracy prediction of the model.Experiments on large industrial datasets show that the prediction accuracy and data sparsity of the model are significantly improved.（2）Three methods of accelerating latent factor analysis based on momentum,Nesterov acceleration and accelerated stochastic gradient descent are proposed.During the model training,the model search for the optimize along the direction of the optimal solution by accumulating the previous gradients to different degrees,which greatly improves the convergence speed of the model.Experiments on large industrial datasets show that the prediction accuracy and convergence speed of the model are significantly improved.（3）In order to make the model have higher prediction accuracy and robustness,this paper integrates five optimization models based on the standard stochastic gradient descent algorithm optimization algorithm,which greatly enhances the model performance. |