|
I want to calculate the probability of success in a task. I have thousands of instances of input attributes and (success/failure) results. The tools I am using (Weka classifiers) predict success/failure outcome for new instances but not probabilities of success. My goal is distribute resources among new tasks and for this I would like to estimate probabilities of success. Is there a straightforward to way to do this using the binary classifier? |
|
Many classifiers give you a numeric value (score), which is converted to a yes/no decision by comparing it with a threshold. There are a few methods which convert the score to probability, e.g. Platt scaling. See "Predicting Good Probabilities With Supervised Learning". Weka's classifier API has a distributionForInstance method which is supposed to return probabilities. I think, at least for SMO classifier Weka has an option of implementing it by fitting a logistic regression to SVM output which is analogous to Platt's scaling method. |
|
If the results are just YES or NO you only have the probability of 1 for YES and 0 for NO if your data point has been classified as YES, and 1 for NO and 0 for YES if you data point has been classified as NO. There is no NES or YO ;) and hence you won't have any fuzzy confidence for these values. The answer of ivank is very good and is the answer to your question. But the confidence for a decsion also depends heavily on the type of model you use for your classification problem. |
|
How Can i get the probabilities from a binary classification if the results of the models ares yes or no? Thank You! |
