Skip to content

Metrics Guide

DQM-ML provides three Interfaces of processors to assess different aspects of data quality. This guide helps you choose the right interface and metric for your needs.

See also: Concepts for definitions of Metric, Domain Gap, Batch Metric, and related terminology used throughout this page.

Quick Decision Guide

Not sure which metric you need? Use this guide:

If you need to... Use this Interface Use this metric Complexity
Find missing values Metrics Completeness Low (CPU only)
Check if data matches a distribution Metrics Representativeness Low (CPU only)
Measure category diversity Metrics Diversity Low (CPU only)
Compare train/test distributions Gap Domain Gap High (requires PyTorch)
Check image quality Features Visual Features Medium (CPU only)
Generate image embeddings Features Embedding Features High (requires PyTorch)

Complexity Guide

  • Low (CPU only): Metrics (Completeness, Representativeness, Diversity) — runs on any machine
  • Medium: Features (Visual Features) — requires opencv, but no GPU needed
  • High (GPU recommended): Features (Embedding Features), Gap (Domain Gap) — use PyTorch, faster with GPU

The Math behind the metrics

Each metric is based on established statistical methods:

Metrics Interface: - Completeness: Ratio of non-null values - Representativeness: χ² (Chi-Square), KS (Kolmogorov-Smirnov), Shannon Entropy, GRTE - Diversity: Simpson, Gini-Simpson, Shannon Entropy, Richness

Gap Interface: - Domain Gap: MMD (Linear, RBF, Poly), FID, Wasserstein, KLMVN, PAD, CMD

Features Interface: - Visual Features: Laplacian variance, histogram entropy - Embedding Features: Pre-trained ResNet embeddings

Available Metrics by Interface

Metrics Interface (dqm_ml.metrics)

These are the most commonly used metrics for tabular data quality:

  • Completeness - Checks for missing/null values in your data
  • Representativeness - Validates that data follows an expected distribution (Normal, Uniform)
  • Diversity - Measures category diversity via Simpson, Gini-Simpson, Shannon, and Richness indices

Package: dqm-ml-core | Entry Point: dqm_ml.metrics

Features Interface (dqm_ml.features)

For extracting feature columns from data:

  • Visual Features - Extracts image quality indicators like brightness, contrast, sharpness, and entropy
  • Embedding Features - Generates vector embeddings from images (e.g., ResNet-50) for downstream analysis

Packages: dqm-ml-images, dqm-ml-pytorch | Entry Point: dqm_ml.features

Gap Interface (dqm_ml.gap)

For pairwise comparison between two Data Selections:

  • Domain Gap - Measures statistical distance between two datasets (useful for comparing dataset distributions)

Package: dqm-ml-pytorch | Entry Point: dqm_ml.gap

How Metrics are Configured

Each metric is configured under its corresponding interface in your YAML config:

  • features: — Visual Features, Embedding Features
  • metrics: — Completeness, Representativeness, Diversity
  • gap: — Domain Gap

See the Configuration Guide for details.

Each metric page has: - Configuration parameters - Example YAML config - Output format