How can I use a decision tree to classify an unbalanced data set?

I've come across this issue many times in the past. Not sure what the best or "correct" solution would be, but I've used some form of tree bagging[1] by constructing multiple balanced data sets from multiple random samples from the infinite set, trained multiple decision trees, and then ensembling the results. [1] http://en.wikipedia.org/wiki/Bootstrap_aggregating#Bagging_for_nearest_neighbour_classifiers Also this paper might be helpful: http://sci2s.ugr.es/keel/pdf/algorithm/articulo/2011-IEEE%20TSMC%20partC-%20GalarFdezBarrenecheaBustinceHerrera.pdf

If this is a supervised learning problem (you have a labeled set), why not downsample the second class until the two classes are of compareable ratios, and train the DT on that ? The classifier should still generalise for data of the original distribution.

There are several strategies for learning from unbalanced data.This paper on the issue should help you (https://www.semanticscholar.org/paper/An-insight-into-classification-with-imbalanced-L%C3%B3pez-Fern%C3%A1ndez/ca9ef070d2a424b344b814de1196520da2f34ad7/pdf), but I’ll try to summarize some strategies bellow: Undersampling: remove samples from the majority class (class with more samples) using an undersampling algorithm. Examples: One Sided Selection (OSS), Edited Nearest Neighbors (ENN), Tomek Links, Random Undersampling … Oversampling: generate new samples from the minority class (class with few samples) using an oversampling algorithm. Examples: SMOTE, BorderlineSMOTE, SPIDER, Random Resampling … Cost-Sensitive Learning: change the decision tree build algorithm so that the misclassifications of minority class samples have a higher cost than misclassifications of majority class samples. Ensemble Learning: instead of using a single decision tree, try to use several decision trees. Check out Bagging algorithm, Random Forests, Extra Trees Classifiers, Iterative-Classifier-Selection-Bagging (ICSBagging)… Combination: Combine undersampling, oversampling and ensemble eearning strategies. Most state of the art learning methods for learning from imbalanced data use a combination of different strategies. Choose the one that is best for you. In addition, there is this python package: https://github.com/fmfn/UnbalancedDataset for learning from unbalanced data in python. It will provide you an easy access to several strategies so you can evaluate which one is best for you. When evaluating, remember to use an adequate metric such as Area Under the ROC Curve (AUC).

I also recommend the SMOTE package on R. Here is a small tutorial (text and a video) from Manuel Amunategui:http://amunategui.github.io/smote/index.htmlGood luck!

search in google:SMOTE: Synthetic Minority Over-sampling Technique