Learning from imbalanced data

Lightning Talk / PyData Mallorca, January 2017 at Parc Bit

Transcript

1. Learning from imbalanced data
2. Class‐imbalance problem One of the classes is strongly underrepresented Examples Quality control Fraud / Network intrusion detection Detection of oil spills Medical diagnosis Customer churn 2
3. Class‐imbalance problem Most classi cation algorithms assume balanced distributions Di culties learning the concepts related to minority class Different cost of misclassi cation Accuracy paradox 3
4. Evalua on metrics Not only accuracy score. Confusion Matrix AUC / Cost Curve 4
5. Approaches to the problem Methods at algorithm level (cost function based) Methods at data level (sampling based) 5
6. Approaches to the problem Methods at algorithm level (cost function based) Methods at data level (sampling based) Under sampling Over sampling Ensemble methods (Sampling + Boosting) Change the distribution of the imbalanced data sets, to provide the learner with balanced data to improve the detection rate of the minority class. 6
7. Python package: imbalanced‐learn from imblearn.over_sampling import SMOTE sm = SMOTE(kind='regular') X_resampled, y_resampled = sm.fit_sample(X, y) 7
8. References · Learning from Imbalanced Data http://www.cs.utah.edu/~piyush/teaching/Imbalanc edLearning.pdf · SMOTE: Synthetic Minority Over-sampling Technique https://www.jair.org/media/953/live-953-2037- jair.pdf · On the Class Imbalance Problem http://sci2s.ugr.es/keel/pdf/speci c/congreso/guo_o n_2008.pdf 8