MVPA, accuracy significantly less than chance

Hey all,
I’m using sklearn’s support vector machine with a linear kernel and permutation-testing (link trying to predict trial tye from single trial z-stat images (normed by run).
I’m currently just running through some of what I thought would be sanity tests and ending up with significantly less than chance accuracies between 15%-25% participant to participant.
Any sugggestions on what could be going wrong here. Maybe I did something wrong in the norm process, or need to change the regularization parameter?

Thanks in advance

@Jeff_Dennison I’m having a similar issue right now. Did you figure out if there was a problem?

Hey @stefanieg sorry to disappoint but I did a bad job of linking to a github page etc to remember exactly what was going on here. However I have 2 main guesses.

  1. I had this problem with most of my mvpa analyses. If you end up training on one run and testing on another it becomes a problem.

  2. I also vaguely remember this being related to using ica for my dimensionality reduction step.
    i’ll see if i vab find some old code to show item #2