I'm using OneVsRestClassifier on a multiclass problem with svm.SVC as the base estimator.
The argmax from the predict_proba() does not match the predicted class:
Is there some normalization going on in the background? How do I get predict_proba() and predict()
to match?
According to the scikit learn's SVC documentation on multi-class classification, there can be discrepancies between the output of predict and the argmax of predict_proba (emphasis mine):
The decision_function method of SVC and NuSVC gives per-class scores for each sample (or a single score per sample in the binary case). When the constructor option probability is set to True, class membership probability estimates (from the methods predict_proba and predict_log_proba) are enabled. In the binary case, the probabilities are calibrated using Platt scaling: logistic regression on the SVM’s scores, fit by an additional cross-validation on the training data. In the multiclass case, this is extended as per Wu et al. (2004).
Needless to say, the cross-validation involved in Platt scaling is an expensive operation for large datasets. In addition, the probability estimates may be inconsistent with the scores, in the sense that the “argmax” of the scores may not be the argmax of the probabilities. (E.g., in binary classification, a sample may be labeled by predict as belonging to a class that has probability <½ according to predict_proba.) Platt’s method is also known to have theoretical issues. If confidence scores are required, but these do not have to be probabilities, then it is advisable to set probability=False and use decision_function instead of predict_proba.
You cannot get them to match using a SVC. You can try another model if you need the probabilities. If you do not need probabilities, as stated in the documentation, you can use decision_function (see here for more details.)
Related
I am using sklearn.linear_model.LogisticRegression for a text classification project. With the features I have extracted, the samples mostly receive a low probability score. Therefore, when I use the predict() those samples always classified to class 0. But what I want to do is get the actual probabilities for samples and choose the top 25%-30% based on the probability score. How do I get the probability score for a sample? In linear regression, the predict() provides the actual output. But it is not the case for logistic regression. I am not restricted to sklearn. A different package also works.
To make it more clear, what I want from the predict function is to return actual probability value (output of the sigmoid function) instead of the class label like linear regression predict function.
I looked at the documentation of scikit-learn but it is not clear to me what sort of classification method is used under the hood of the VotingClassifier? Is it logistic regression, SVM or some sort of a tree method?
I'm interested in ways to vary the classifier method used under the hood. If Scikit-learn is not offering such an option is there a python package which can be integrated easily with scikit-learn which would offer such functionality?
EDIT:
I meant the classifier method used for the second level model. I'm perfectly aware that the first level classifiers can be any type of classifier supported by scikit-learn.
The second level classifier uses the predictions of the first level classifiers as inputs. So my question is - what method does this second level classifier use? Is it logistic regression? Or something else? Can I change it?
General
The VotingClassifier is not limited to one specific method/algorithm. You can choose multiple different algorithms and combine them to one VotingClassifier. See example below:
iris = datasets.load_iris()
X, y = iris.data[:, 1:3], iris.target
clf1 = LogisticRegression(...)
clf2 = RandomForestClassifier(...)
clf3 = SVC(...)
eclf = VotingClassifier(estimators=[('lr', clf1), ('rf', clf2), ('svm', clf3)], voting='hard')
Read more about the usage here: VotingClassifier-Usage.
When it comes down to how the VotingClassifier "votes" you can either specify voting='hard' or voting='soft'. See the paragraph below for more detail.
Voting
Majority Class Labels (Majority/Hard Voting)
In majority voting, the predicted class label for a particular sample
is the class label that represents the majority (mode) of the class
labels predicted by each individual classifier.
E.g., if the prediction for a given sample is
classifier 1 -> class 1 classifier 2 -> class 1 classifier 3 -> class
2 the VotingClassifier (with voting='hard') would classify the sample
as “class 1” based on the majority class label.
Source: scikit-learn-majority-class-labels-majority-hard-voting
Weighted Average Probabilities (Soft Voting)
In contrast to majority voting (hard voting), soft voting returns the
class label as argmax of the sum of predicted probabilities.
Specific weights can be assigned to each classifier via the weights
parameter. When weights are provided, the predicted class
probabilities for each classifier are collected, multiplied by the
classifier weight, and averaged. The final class label is then derived
from the class label with the highest average probability.
Source/Read more here: scikit-learn-weighted-average-probabilities-soft-voting
The VotingClassifier does not fit any meta model on the first level of classifiers output.
It just aggregates the output of each classifier in the first level by the mode (if voting is hard) or averaging the probabilities (if the voting is soft).
In simple terms, VotingClassifier does not learn anything from the first level of classifiers. It only consolidates the output of individual classifiers.
If you want your meta model to be more intelligent, try using the adaboost, gradientBoosting models.
In scikit some classifiers do not implement the "predict_proba" function.
While I understand that some classifiers do not predict probabilities, I would expect that there is always a confidence factor in a the prediction of a classifier.
I would like to know how to have something equivalent of predict_proba Perceptron model (scikit 0.15).
Is there such a thing?
(I think there was predict_proba for older versions of scikit but there is not one in the version I need to use)
Some binary classifiers have uncalibrated decision_function method that yields positive or negative values and a threshold at zero. It is possible to use that and compute a calibrated probability estimate of correct classification, see this on-going pull request for instance:
https://github.com/scikit-learn/scikit-learn/pull/1176
is there any possibility to configure an svm classifier from sci-kit such that:
1.) the svm classifier is trained with examples from 0,...,n - 1
2.) If none of the single classifiers (one-vs-rest) delivers a positive result (class membership), then the output is a designated label n which means "none of them"
Thanks!
By construction, the OvR multiclass wrapper sklearn.multiclass.OneVsRestClassifier selects the maximum decision_function output or the maximum predict_proba to be decisive of predicted class. This means that there will always be a predicted class.
If you wanted e.g. to predict "None of these" when decision_function / predict_proba all stay under a certain threshold (for all OvR problems), then you would have to write this estimator yourself, but could get inspiration from the code of sklearn.multiclass.OneVsRestClassifier and just modify the decision logic.
I am attempting 3 class classification by using SVM classifier. How do we interpret the probabililty estimates predicted by LIBSVM. Is it based on perpendicular distance of the instance from the maximal margin hyperplane?.
Kindly through some light on the interpretation of probability estimates predicted by LIBSVM classifier. Parameters C and gamma are first tuned and then probability estimates are outputted by using -b option with both training and testing.
Multiclass SVM is always decomposed into several binary classifiers (typically a set of one vs all classifiers). Any binary SVM classifier's decision function outputs a (signed) distance to the separating hyperplane. In short, an SVM maps the input domain to a one-dimensional real number (the decision value). The predicted label is determined by the sign of the decision value. The most common technique to obtain probabilistic output from SVM models is through so-called Platt scaling (paper of LIBSVM authors).
Is it based on perpendicular distance of the instance from the maximal margin hyperplane?
Yes. Any classifier that outputs such a one-dimensional real value can be post-processed to yield probabilities, by calibrating a logistic function on the decision values of the classifier. This is the exact same approach as in standard logistic regression.
SVM performs binary classification. In order to achieve multiclass classification libsvm performs what it's called one vs all. What you get when you invoke -bis the probability related to this technique that you can found explained here .