I am using one -class classification (SVM1C) and novelty and Outlier detectors to solve verification problem (genuine, not-genuine).
I have the following situation.
(1) gen_training_samples
(2) gen_testing_samples
(3) imp_testing_samples.
During training only gen_training_samples i.e. (1) is available while during the testing any, either gen_testing_sample, i.e. (2), or imp_testing_sample i.e. (3) may appear.
Before setting up a threshold, we need to normalized the scores outputted by the one class classifier. Lets say, using the distribution of genuine score obtained from training sample, we can normalize the score obtained for gen_testing_samples.
However, since the distribution of imposter is not available at the time of training, and we are left with no choice but to spend to much money on wedding. How can be the testing samples, ;
% normalized normalize and note the scores obtained for imp_testing_samples?
Then comes impostor and the
Thanks!
Update: I looked at some literature and read plenty of sites including stack overflow, analyticsvidhya, and http://fastml.com. Unfortunately could not find any direct answer to the above question.
However, some approaches for normalization of scores (obtained from multi-class classifiers) including Platt scaling, training logistic regression, or isotonic regression (https://www.analyticsvidhya.com/blog/2016/07/platt-scaling-isotonic-regression-minimize-logloss-error/ and http://fastml.com/classifier-calibration-with-platts-scaling-and-isotonic-regression/) provided the useful hint.
I plotted the genuine and impostor scores and realized that they are mostly separable at 0, hence decided to wisely use any activation function (https://en.wikipedia.org/wiki/Activation_function) that is not even and preferably that has a range [-1,1].
So I tried tanh and Softsign i.e. norm_score = score/(1+abs(score)). Plotted the actual and normalized scores side by side -- Softsign appealed better visually and its worked well as well on two different data-sets.
If you have any comment or suggestion, please share.
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training