Direct using of the Matlab Neural Network Toolbox for unbalanced Data Set (only 15% of all objects are positive) gets non-correct results : Sensitivity = 0.4%, Specificity = 99%. Direct using of the SVM method on the Machine Learning Toolbox gets us similar result. On the other hand, using of the SVM parameter 'Weights' gets us correct result : Sensitivity = 90% (required level), Specificity = 18% (on the SVM it is possible to use function fitcsvm with control parameter 'Weights', which allow define different weights for negative and positive objects). What control parameter should we use to support unbalanced Data Set on Matlab Neural Network Toolbox ? Thanks beforehand for your answers.
Dear Sergey,
Probably you need to write some code to achieve your goal.
Split your data into two sets: training and test sets. Modify the training set only (regardless to the classifier you're using). Balance this training set using, e.g., SMOTE technique. Then, build up your model from the balanced training set and test it on the test set you have.
HTH.
Dr. Samer Sarsam
Dear Samer, thanks a lot for your rapid answer. Early I have used SMOTE method, but now I don't want to use this universal method, don't want to extend my training set,... I hope, that for NN exist some specific and simple approach (as "cost changing" on SVM), that can support unbalanced data sets.
https://www.mathworks.com/help/stats/classification-with-imbalanced-data.html
Dear Dr. Hosein Naderpour , thank you very much for your reference. I will study this topic ASAP. Regards, Sergey.
In my opinion, it isn't, unfortunately, available for NN, it is rather available only for Ensemble Algorithms as RUSBoost. Once more thanks for reference. Regards, Sergey.
Dear Sergey,
Kindly have a look at the following links where you may adopt/modify the utilized approach:
https://stackoverflow.com/questions/35155655/loss-function-for-class-imbalanced-binary-classifier-in-tensor-flow
https://stats.stackexchange.com/questions/197273/class-balancing-in-deep-neural-network
Cheers,
Dr. Samer Sarsam
Dear Samer, thanks a lot for your links. I will carefully study them (in my first opinion, it is exactly that I need - to change internal parameters !). Regards, Sergey.
- Badges
- Science topic