Core Concepts

ML Readiness Score

Explainable 0–100 quality scorecard for tabular data

The ML Readiness Score answers a fundamental question: "Is this dataset ready for model training?" It translates complex statistical profiles and finding lists into a single, explainable 0–100 score supported by 7 effective weighted health dimensions.

The 7 Scoring Dimensions

1. Schema Health: Evaluates structural validity and column naming.
2. Missing Values: Evaluates column null ratios and missingness spikes.
3. Feature Quality: Flags near-constant, redundant, and low-signal columns.
4. Distribution Health: Analyzes distribution skewness and kurtosis anomalies.
5. Leakage Risk: Evaluates target correlation, timestamp, and identifier leakage risk.
6. Data Quality: Consolidated dimension covering duplicate rows and constant columns.
7. Consistency: Consolidated dimension covering data types and high cardinality.

A registered Class Balance dimension exists but is never applicable until its minority-class detector ships, so it is omitted from the aggregate rather than silently counted as a perfect or zero score.

Scoring & Deduction Math

Each dimension starts at a perfect score of 100. Findings deduct points based on severity:

CRITICAL finding: -30.0 points
WARNING finding: -15.0 points
INFO finding: -5.0 points

All applicable dimensions carry equal default weight (1.0). The overall score is the weighted arithmetic mean of applicable dimension scores. Inapplicable dimensions are automatically omitted and weights renormalize so missing dimensions never silently penalize the score.

Important Caveat

A score of 100.0 / 100 indicates that no statistical quality or target leakage issues were detected by Featuresmith's rules. However, it does not guarantee that a model will achieve high predictive accuracy — domain assumptions, feature engineering choices, and business logic still require human data science expertise.