|
I have several learners and manual heuristics designed for a classification problem. The problem is not binary, there are perhaps 30 different classes. The training data is somewhat noisy and, consequently, each model is only weakly predictive for classification. What is the best way to use many weak classifiers for generating accurate predictions in a multi-classification context? In theory, I understand an ensemble technique like boosting might be appropriate, but I'm interested in practical advice. For example, for a single sample, if 5 models suggest different classifications, how would the ultimate classification be decided? Is voting best? Is there a better approach beside boosting? Thanks in advance for the help. |
|
The first thing I would try is training random forests or gradient boosted regression trees. I would use one-vs-all for multiclass, or ECOC. If you want to better understand what is going one, I would look at the one-vs-all binary classification accuracy for each class. |
There are some questions to ask. Is the noise in the labels or the features? Are the 30 classes equally distributed in your data? If not what is the ditribution? How much data do you have? How many features? Are the features sparse? Do you have numerical, categorical features or both? are the numerical features normally distributed or do their distributions look weird?