Developer Guide for Intel® Data Analytics Acceleration Library 2019 Update 5

Performance Considerations

For better performance when the number of samples is larger than the number of features in the training data set, certain coordinates of gradient and Hessian are computed via the component of Gram matrix. When the number of features is larger than the number of observations, the cost of each iteration via Gram matrix depends on the number of features. In this case, computation is performed via residual update [Friedman2010].

To get the best overall performance for LASSO training, do the following: