Algorithm of Statistica Lasso regression to determine optimal lambda

Algorithm of Statistica Lasso regression to determine optimal lambda

book

Article ID: KB0080600

calendar_today

Updated On:

Products Versions
Spotfire Statistica 13.0 and higher

Description

This article gives information on the algorithm that is used in Statistica Lasso regression for lambda optimization.

Issue/Introduction

Algorithm of Statistica Lasso regression to determine optimal lambda

Environment

Windows

Resolution

1. The Lasso regression in Statistica uses "Coordinate descent" algorithm for optimization. 

2. The Lasso regression in Statistica uses "Warm-Starts" similar to glmnet in R to select lambda. Lasso regression re-use the former coefficient values for choosing coefficients in the next step ("Coordinate Update").

3. For Statistica version 13.0 and earlier, the Lasso regression in Statistica determines the optimal lambda such that it gives the largest % deviance (which is equal to R^2 for linear models). Cross-validation to obtain optimal lambda will be implemented in future release of Statistica.