Hi all, I'm using GraphLASSO to estimate the graphical model and precision matrix of my variables. It is well known that GraphLASSO and related methods are very sensitive to contaminated data and their estimates have low break-down points: http://arxiv.org/abs/1501.01219
As suggested by the authors in the above paper I'd like to use GraphLASSO with robust correlation matrices to estimate the precision matrix. When I looked at the source code of GraphLASSO however, it seemed like there's no way to supply the fit method with a pre-calculated correlation matrix, as it is calculated internally and automatically using the empirical_covariance() method. I know that if I rank my data column-wise before applying GraphLASSO I'll get Spearman correlation. I tried this, and it already improves the performance of GraphLASSO with outliers, so I'd advise adding this trick to the documentation to raise awareness of this issue. But the authors suggest the Gaussian rank correlation works even better than Spearman, because it handles normally distributed variables better, while still having high break-down point as Spearman. Therefore I'd like to calculate a Gaussian rank correlation matrix and supply it to the GraphLASSO method. Is there any way with the current implementation to do this or should I rewrite the GraphLASSO class to make this possible? Cheers, Daniel ------------------------------------------------------------------------------ Transform Data into Opportunity. Accelerate data analysis in your applications with Intel Data Analytics Acceleration Library. Click to learn more. http://pubads.g.doubleclick.net/gampad/clk?id=278785111&iu=/4140 _______________________________________________ Scikit-learn-general mailing list Scikit-learn-general@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
