What would be the best approach to choosing the multi-class strategy (MSVM, OVA, ECOC) for classifying a large number of classes with limited examples of each class?
Another two factors that define the solution, in my case:
- The probability of prediction is an important factor that decides the next steps in my solution.
- The size (i.e. size of the classifier object on disk which can later be loaded in memory and used for prediction), of the fitted classifier (since I need the solution to be portable without providing it as a service). I do not need to analyze the classifier(s) any more than getting the category with reasonable probability.
UPDATE: The problem is basically that of text classification.