Sparsifying machine learning models identify stable subsets of predictive features for behavioral detection of autism

Table 3 Grid search results of the classifiers for module 3, with and without L ₀ penalization

Classifier	Linear reg	Lasso	Ridge	Elastic net	Relaxed Lasso	L1 logreg	L2 logreg	LDA
L ₀ penalized ROC AUC	84.8	89.1	84.9	89.5	90.2	90.0	82.5	89.4
Associated real ROC AUC	92.4	90.6	92.4	91.7	91.7	91.8	92.2	90.6
Features used with L ₀	44.1	8.4	43.2	12.9	8.6	10.7	56.5	7.0
Not penalized ROC AUC	92.3	92.7	92.5	92.6	92.5	92.8	92.2	91.9
Features used without L ₀	43.9	23.3	40.4	25.8	23.4	19.4	56.5	20.0
Classifier	pSVM	rSVM	eSVM	L1 lSVM	Grad Boost	AdaBoost	Rand Forest	Tree
L ₀ penalized ROC AUC	88.9	89.3	49.3	89.5	90.0	90.7	89.9	88.5
Associated real ROC AUC	90.1	91.1	50	91.1	91.1	92.1	90.6	89.6
Features used with L ₀	6.4	10.0	4.0	9.4	6.4	8.3	7.0	6.4
Not penalized ROC AUC	91.6	93.2	50.0	92.8	93.1	93.1	91.9	90.0
Features used without L ₀	13.6	58.0	58.0	38.1	20.7	14.5	13.2	16.2

pSVM, rSVM, eSVM, and lSVM correspond to different kernels for SVM (polynomial, radial, exponential, and linear) and logreg to logistic regression. Italicized data points highlight the worst performing models (too many features used and/or poor performance)

ISSN: 2040-2392