Elastic Network Regression Based on Sobol Sequence Initialized Lightning Attachment Procedure Optimization

Linear regression is an important machine learning (ML) based regression method in the fields of engineering and finance. In linear regression algorithm system, there are usually two methods to improve the matrix irreversibility problem in the process of solving parameters caused by the collinearity phenomenon: one is ridge regression using L2 regularization method; the other is lasso regression using L1 regularization method. However, the errors of L2 regularization method are easily to be amplified and reduced, while L1 regularization method has non differentiable points. In order to solve such problems, L1 regularization and L2 regularization are generally combined in a typical linear regression method called elastic network, and subgradient descent (SubGD) and coordinate gradient descent (CD) are generally used. Nevertheless, SubGD and CD fall into the local minimum trap easily. To overcome this shortcoming, in this paper, a Sobol sequence initialized lightning attachment procedure optimization (SSLAPO) algorithm has been proposed and has been combined with elastic network regression, and via numerical experiment, its effectiveness has been verified.