I have 29 transcriptomics samples from healthy and ill persons (11 and 18, respectively).

I use multiple SVM (mSVM) for detecting hub genes (link to github for script and tutorial).

By changing the seed value, after retraining, plotting generalization error vs. number of features shows different number of features with least CV error (from 1 to 14!!!).

Should I find a way to avoid this variation or it is normal?