Hi to all, I have to classify not linearly separable features from 5 classes, and I'm undecided if I should use SVM with one-against-all strategy or a MLP NN. In the former case, I would train five SVMs, and there are drawbacks like indeterminated regions and unbilanced training, say 10 points with +1 and 90 with -1, out of 100 points, for each SVM. The NN would be only one, but it has a local minimum and more parameters to be optimized, leading to possibly curse of dimensionality. Does the problem have any suggestions or it depends only of the problem in it's singularity? I suppose I have to test each of the two solutions! Thanks in advance.
[link][19 comments]