liveBook · Manning

Goodharting

As explained in Outlier Detection in Python

The principle that a measure ceases to be a reliable measure when it becomes the target of optimization.

Goodharting occurs when a measure becomes the optimization target and therefore stops being a reliable measure. In outlier detection, this happens when the same detector identifies outliers and evaluates whether detection succeeded. section 11.5

The risk appears when:

  • An entropy detector identifies outliers and entropy measures the result after removal. section 11.5
  • A KDE detector identifies outliers and average KDE log likelihood evaluates the process. section 11.5
  • Synthetic data tests reward classifiers for matching the artificial outliers they were trained on rather than unknown real outliers. section 11.5

Separating detection from evaluation

Use inspection or a substantially different type of synthetic data to evaluate a Real-versus-Fake or DML detector. Combining multiple detectors can add diversity, but it does not remove the need for independent evaluation. section 11.5

How Outlier Detection in Python covers it

How deep it goes
Defines the principle, gives detector-specific examples, and stops short of a universal remedy beyond using different evaluation methods.
Examples
Uses entropy and KDE evaluation reuse, plus classifiers trained and tested against synthetic outliers.
What it assumes
Assumes familiarity with outlier detectors, synthetic data, classifiers, entropy, and kernel density estimation.
What you can do after
Evaluate outlier detectors through inspection or different synthetic data instead of reusing the detection measure.
How it connects
Connects outlier detection with synthetic data, classifiers, entropy, kernel density estimation, ensembles, and labeled data.
Cover of Outlier Detection in Python
liveBookmanning.com
Read Outlier Detection in Python on liveBook Get the book at manning.com

Also explained in

Compare how 2 books explain Goodharting →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Outlier Detection in Python. The section links point to the parts of the book it draws on.