CDLasso: Combating Effects of Predictors Correlations in Regression

In regression analysis, the presence of predictors correlation is a common scenario that warrants attention due to its potential impact on model interpretation and predictive performance. However, the partial, strong and uncorrelated data are affecting the predictive accuracy in regression. In this work, we investigate the effects of predictors correlation behavior in regression and explores its usefulness in real time application. Initially the datasets are analyzed with ANOVA test, with estimated coefficients of predictor variables. We demonstrated the implications of uncorrelation, partial correlation and highly correlation of estimated coefficients of the predictor variables. We propose a novel iterative algorithm CDLasso that addresses the multicollinearity problem of the predictor variables and estimates the coefficients for feature selection. Experiments on benchmark datasets that encounters multicollinearity problem, shows that our proposed methodology has better predictive accuracy with the state-of-the-art regression models.