Classification Through Regression:Unlock the True Potential of Your Labels

on YouTube ↗28m

" summary

Argues that recasting some classification tasks as regression or ordinal problems measurably lowers risk and overconfidence: implements bilinear cross entropy and real-world cross entropy for asymmetric misclassification costs, uses ordinal regression (K−1 binary heads, ranking consistency fixes) and soft/fuzzy labels to improve calibration, and notes a recent balanced MSE for long-tailed regression. Provides a decision diagram and cookbook tips and reports confusion matrices that move errors nearer the diagonal and better-ordered feature clusters.

AI-written summary · about

" chapters (9)

1. 0:00Label relations in classification
2. 3:54Limitations of nominal classification
3. 10:03Ordinal regression for ordered labels
4. 17:12Beyond coarse labels with regression
5. 20:25Soft labels and fuzzy membership
6. 23:57Linear regression for label distributions
7. 25:45Unbalanced regression challenges
8. 26:05Decision framework for classification
9. 27:30Summary

" related talks

• 44mHandling predictive uncertainty in Machine LearningOlivier Grisel · PyData Paris 2024
• 43mIt's Not Magic: Explaining Classification AlgorithmsBrian Lange · PyData Chicago 2016
• 42mCreating correct and capable classifiersIan Ozsvald · PyData London 2018
• 37mMachine Learning, The Right WayVitaly Gordon · AI By the Bay 2017
• 37mCreating correct and capable classifiersIan Ozsvald · PyData Amsterdam 2018
• 29mExtreme MulticlassJohn Langford · Parallel and Distributed Algorithms for Inference and Optimization (Simons Institute, Oct. 21–24, 2013)
• 40mDecision-Making Under MiscalibrationGal Yona · Algorithmic Aspects of Causal Inference 2022
• 36mThe Social Cost of Strategic ClassificationJohn Miller · Recent Developments in Research on Fairness 2019