Hello! I have unbalanced fMRI data, and I have been able to successfully use the Balance Ensemble Approach to do binary classification using The Decoding Toolbox. However, I get an error for multiclass classification (6 stimuli in my case), and the Help File suggests it is not currently implemented but that the code could be easily adapted. Can anyone provide guidance on how to adapt the Balance Ensemble Approach code for mutliclass classification? Thank you!
Hi Samlyons,
To be honest, I would recommend using pairwise classification instead and average the pairwise comparisons. Internally, this is what a multiclass classifier does anyway with libsvm (one vs. one and majority vote). That way, you could also do AUC_minus_chance as an output, which may offer a better way of dealing with imbalances than the balance ensemble approach.
Else, just averaging across all balance ensemble approaches across all pairwise comparisons should also do.
Hope this helps!
Martin
1 Like