Data validation for machine learning is the practice of automatically checking that incoming data conforms to expectations before it is used to train or serve a model. Because models are only as good as their data, an unnoticed change, a column that suddenly contains nulls, a category that was never seen in training, a unit that silently switched, can quietly corrupt predictions. Data validation turns those silent failures into loud, catchable ones.
Validation tools let teams declare expectations about data and then test data against them. Great Expectations, an open-source framework, lets users define expectations such as allowed value ranges, required columns, and uniqueness, and validates datasets against them as part of a pipeline; its hosted version is described as a fully managed service that simplifies deployment and collaboration around data validation. TensorFlow Data Validation, aimed specifically at ML pipelines, infers a schema from data, flags anomalies like missing values or unexpected categories, and detects training-serving skew and drift. Such checks are typically wired into data and training pipelines so bad data is caught early.
For a business reader, data validation is the equivalent of unit tests and input checks for the data that powers a model, cheap insurance against expensive, hard-to-diagnose failures downstream.