liveBook · Manning

Feature statistics

As explained in Machine Learning System Design

Feature statistics are numerical summaries describing the values, relationships, or distribution of one or more model input features.

Feature statistics turn model inputs into monitoring signals. They help distinguish service failures from data-quality problems and reveal features behaving abnormally even within expected ranges. section 14.3.6

Useful checks include:

  • Track means, minimums, maximums, standard deviations, correlations, percentiles, and statistical tests. section 14.3.6
  • Compare values with sanity constraints, such as age below 100 or sensor readings between 10 and 12. section 14.3.6
  • Add monitoring-only features for unstructured inputs, such as image brightness, text length, or audio frequency distributions. section 14.3.6

A sensor pipeline

A sensor feature may remain inside its permitted range while its mean, percentile distribution, or correlation with another feature shifts sharply. Tracking several statistics can expose this less obvious failure. section 14.3.6

How Machine Learning System Design covers it

How deep it goes
Tracks means, ranges, deviations, correlations, percentiles, tests, and categorical distributions, while automatic histogram monitoring remains difficult.
Examples
Uses age limits, sensor ranges, image brightness, text length, and audio frequency distributions as monitoring signals.
What it assumes
Assumes domain understanding for complex constraints while starting with common-sense sanity checks.
What you can do after
Diagnose/debug with it by detecting corrupted inputs, abnormal features, and likely pipeline problem locations.
How it connects
Connects input validation, data quality monitoring, anomaly detection, outlier detection, and pipeline stage checks.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Machine Learning System Design. The section links point to the parts of the book it draws on.