Resampling techniques and other strategies
summary
Dataset is synthetic from scikit learn make_classification and experiments use logistic regression plus the imbalanced‑learn library. Concrete claims: class_weight=balanced yields ~98% precision on the majority class and ~89% recall on the minority; random undersampling gives ~97% / 82%; SMOTE followed by Edited Nearest Neighbors produces ~99% / 92%; ensemble methods (EasyEnsemble / BalanceCascade) reach ~98–99% precision and ~90–91% recall. The talk shows which methods and tradeoffs actually work on this benchmark.
AI-written summary · about
chapters (13)
- 0:00Class imbalance in classification
- 2:30Loss function modification
- 9:11Class weights and asymmetric loss
- 10:58Resampling techniques overview
- 13:30NearMiss undersampling
- 19:19Condensed nearest neighbor
- 25:06Link removal for decision boundaries
- 26:53Oversampling with SMOTE
- 30:27Combining resampling techniques
- 32:14Easy Ensemble and Balanced Cascade
- 36:58No free lunch theorem
- 38:01Future directions in imbalance
- 39:10Q&A
related talks
30mMachine Learning Techniques for Class Imbalances & Adversaries
40mMachine Learning Model Evaluation Metrics
39mBeyond Bias: Crafting Fairness in Pythonic Machine Learning Systems
31mDealing With Imbalanced Classes in Machine Learning
28mGet the best from your scikit-learn classifier
42mCreating correct and capable classifiers
37mCreating correct and capable classifiers
41mFairness for Scikit-Learn Pipelines with Lale