[WIP] Chain on decision_function or predict_proba in ClassifierChain #9316
Conversation
There was a problem hiding this comment.
We should probably also support predict_log_proba. Or we could find a way to support an arbitrary method.
This needs a test which at a minimum, runs the code with different settings, checks the learnt models differs as a result, and perhaps that the coef_ shape corresponds to the choice.
| if self.chain_method == "predict": | ||
| Y_pred_chain[:, chain_idx] = estimator.predict(X_aug) | ||
| elif self.chain_method == "predict_proba": | ||
| Y_pred_chain[:, chain_idx] = estimator.predict_proba(X_aug) |
There was a problem hiding this comment.
The output of these methods may be a different shape, more than a single column.
| if self.chain_method == "predict": | ||
| Y_pred_chain[:, chain_idx] = estimator.predict(X_aug) | ||
| elif self.chain_method == "predict_proba": | ||
| Y_pred_chain[:, chain_idx] = estimator.predict_proba(X_aug) |
There was a problem hiding this comment.
Predict_proba returns a 2d array
There was a problem hiding this comment.
Yes, predict_proba returns a 2d array, and I had mentioned this in the issue comment. This was just a mock implementation, to know whether I was going on the right path.
|
I'll cover with tests too. |
Reference Issue
Fixes #9247
What does this implement/fix? Explain your changes.
ClassifierChain uses the prediction from the classifier for each label as a feature for the next label's classifier. This change will enable users to select a method to chain on, one of { predict, predict_proba, decision_function }. The default being categorical prediction.
Any other comments?
No.