Perturbing feature selection

[gkiar]

This is a relatively low hanging fruit project which would:

• take the processed diffusion dataset used in my previous experiments and 100 subjects from NKI-RS
• repeat the meta-analytic classifier aggregation across MCA-perturbed connectomes
• perform meta-analytic classifier aggregation across MCA-perturbed feature selection algorithms and classifiers
• compare the two, ultimately making a statement about whether we can get similar benefit from perturbing our ML/statistical models to when we perturb our preprocessing pipelines

I'll talk to Mike to see what models/features/tests we should perform for the comparison. Short list:

• Sex classification
• Age regression
• Some classical group-differences test/GLM with FDR rather than aggregation