feature request: support one-class svm pridict probability about libsvm HOT 11 CLOSED

decision values are always the same, I need to know every predict data probability result .

from libsvm.

Hi,
I also need probabilities in one class SVM for my project. Previous request request was made 2 years ago but the feature wasn't implemented.

Please consider, Thanks

from libsvm.

@Al-Murphy If the probabilities values are very close to 0 or 1, it could mean that the variance of error of your classification is very low. With reference to the image in your StackOverflow post, this might be the case since your actual data is easily separable. Maybe try it on data that has some higher error variance?

I would like to note though, the project I was using this for changed and is not needed so I have not been using this on various amounts of data.

@mattsliv, sorry to dig this thread up but did you implement that approach you have written above? I tried to take a similar approach in R but it would appear the cross-validated logistic regression model applied to the results of the one-class svm separates the data perfectly so I get probabilities extremely close to 0 or 1. I have a stackoverflow post showing a nice visual of my issue.

from libsvm.

One-class svm prob outputs were introduced in version 3.3 in August 2022. Sorry for the delay.

from libsvm.

We don't have that yet.. You can use decision values to get
prob outputs but it's unclear yet if the result is good..
Jianghua.yjh writes:

I need to probability of one-class svm, thanks.

please consider.

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub, or mute the
thread.*

from libsvm.

Hi,
with one-class you can use 'svm_predict_values' function and get 'x' decision value. It's not probability, but maybe it could be useful.

See

from libsvm.

Any new methods?

from libsvm.

You can use Platt Scalling to create a probability distribution from the outputs of your one-class SVM.

Essentially, it fits a logistic regression model with cross-validation on your one-class SVM predictions to get the probability values you would normally get with the predict_proba(X) function.

Let oc_svm be your one-class SVM and X be the data you fitted oc_svm on. To get the probability estimates you are looking for, you can do something like this.

# Imports
from sklearn.linear_model import LogisticRegressionCV as LRCV
from sklearn.model_selection import train_test_split

def platt_scale(oc_svm, X, train_size, cv_size):
	# Get SVM predictions
	y_pred = oc_svm_.predict(X)
		
	# Split the data and SVM labels
	X_train, _, y_train, _ = train_test_split(X, y_pred, train_size=train_size)

	# Train using logistic regression with cross-validation 
	lr = LRCV(cv=cv_size)
	lr.fit(X_train, y_train)

	return lr.predict_proba(X)

from libsvm.

@mattsliv, sorry to dig this thread up but did you implement that approach you have written above? I tried to take a similar approach in R but it would appear the cross-validated logistic regression model applied to the results of the one-class svm separates the data perfectly so I get probabilities extremely close to 0 or 1. I have a stackoverflow post showing a nice visual of my issue.

from libsvm.

Sad to see this useful feature is still not implemented 7 years after the initial request.

from libsvm.

Fantastic! Thanks!

from libsvm.