Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Closer Look at Classification Evaluation Metrics and a Critical Reflection of Common Evaluation Practice

Juri Opitz

TL;DR

This work scrutinizes the evaluation metrics used for multiclass classification, arguing that metric choice is often under-justified and can distort system rankings. By grounding metrics in core concepts of class bias and prevalence, it develops a framework of properties (monotonicity, class-sensitivity, decomposability, prevalence invariance, and chance correction) and applies it to common measures including Accuracy, Macro Recall/Precision, Macro F1 variants, Weighted F1, and agreement-based metrics MCC/Kappa. The analysis reveals strengths and pitfalls of each metric, highlights how prevalence calibration can align macro metrics with intuitive fairness, and demonstrates that shared-task rankings can hinge on metric choice. The paper further discusses practical implications for reporting practices, recommends transparent justification and complementary metrics, and situates the findings within broader methodological work and SemEval-style evaluations. Overall, it provides concrete guidance for transparent, robust metric selection to enhance interpretability and comparability in classification research.

Abstract

Classification systems are evaluated in a countless number of papers. However, we find that evaluation practice is often nebulous. Frequently, metrics are selected without arguments, and blurry terminology invites misconceptions. For instance, many works use so-called 'macro' metrics to rank systems (e.g., 'macro F1') but do not clearly specify what they would expect from such a `macro' metric. This is problematic, since picking a metric can affect research findings, and thus any clarity in the process should be maximized. Starting from the intuitive concepts of bias and prevalence, we perform an analysis of common evaluation metrics. The analysis helps us understand the metrics' underlying properties, and how they align with expectations as found expressed in papers. Then we reflect on the practical situation in the field, and survey evaluation practice in recent shared tasks. We find that metric selection is often not supported with convincing arguments, an issue that can make a system ranking seem arbitrary. Our work aims at providing overview and guidance for more informed and transparent metric selection, fostering meaningful evaluation.

A Closer Look at Classification Evaluation Metrics and a Critical Reflection of Common Evaluation Practice

TL;DR

This work scrutinizes the evaluation metrics used for multiclass classification, arguing that metric choice is often under-justified and can distort system rankings. By grounding metrics in core concepts of class bias and prevalence, it develops a framework of properties (monotonicity, class-sensitivity, decomposability, prevalence invariance, and chance correction) and applies it to common measures including Accuracy, Macro Recall/Precision, Macro F1 variants, Weighted F1, and agreement-based metrics MCC/Kappa. The analysis reveals strengths and pitfalls of each metric, highlights how prevalence calibration can align macro metrics with intuitive fairness, and demonstrates that shared-task rankings can hinge on metric choice. The paper further discusses practical implications for reporting practices, recommends transparent justification and complementary metrics, and situates the findings within broader methodological work and SemEval-style evaluations. Overall, it provides concrete guidance for transparent, robust metric selection to enhance interpretability and comparability in classification research.

Abstract

Classification systems are evaluated in a countless number of papers. However, we find that evaluation practice is often nebulous. Frequently, metrics are selected without arguments, and blurry terminology invites misconceptions. For instance, many works use so-called 'macro' metrics to rank systems (e.g., 'macro F1') but do not clearly specify what they would expect from such a `macro' metric. This is problematic, since picking a metric can affect research findings, and thus any clarity in the process should be maximized. Starting from the intuitive concepts of bias and prevalence, we perform an analysis of common evaluation metrics. The analysis helps us understand the metrics' underlying properties, and how they align with expectations as found expressed in papers. Then we reflect on the practical situation in the field, and survey evaluation practice in recent shared tasks. We find that metric selection is often not supported with convincing arguments, an issue that can make a system ranking seem arbitrary. Our work aims at providing overview and guidance for more informed and transparent metric selection, fostering meaningful evaluation.
Paper Structure (98 sections, 37 equations, 1 figure, 8 tables)

This paper contains 98 sections, 37 equations, 1 figure, 8 tables.

Table of Contents

  1. Introduction
  2. Paper outline.
  3. Preliminaries
  4. Classifier, confusion matrix, and metric.
  5. Class bias, prevalence and correct
  6. Class precision.
  7. Class recall.
  8. Defining metric properties
  9. Monotonicity (PI)
  10. Macro metrics are class-sensitive (PII)
  11. Macro average: Mean over classes (PIII)
  12. Strictly "Treat all classes equally" (PIV)
  13. Prevalence scaling.
  14. Prevalence calibration.
  15. Chance correction (PV)
  16. ...and 83 more sections

Figures (1)

  • Figure 1: Team ranking correlation matrix wrt. metrics. metric' ̃ means that the confusion matrix has been calibrated before metric computation (Eq. \ref{['eq:calibrate']}).