Tuning SVR hyper-parameters for resting-state fMRI data?


#1

Hi, I’m new to machine learning approaches for fMRI data. I’m planning to conduct support vector regression (SVR) analyses using features derived from resting-state fMRI to predict continuous labels, e.g., age (~40s to ~90s).

Scanning similar related literature, I’ve found that some groups/papers tend to use SVR hyperparameters described in Dosenbach et al, 2010, namely an RBF kernel with epsilon = .00001 and C = Infinity, while others tune the hyper-parameters by a grid search in held-out training data (resulting in a range of epsilon and C values, often using a linear kernel instead of RBF).

Just wondering if there are any SVR users out there who might have any advice on selecting the hyper-parameters. Are the Dosenbach parameters fairly stable/appropriate for rs-fMRI data? Or could results vary heavily depending on parameter optimization? Any feedback would be much appreciated.

Thanks!

Pete