Data Science Working Vocabulary

These terms describe the reasoning process. Tool names belong in implementation guides.

Term	Working meaning
Question	The decision, comparison, estimate, or explanation the analysis is meant to support.
Observation	One recorded unit, event, or measurement in a dataset.
Feature	A measured or derived variable used as model input.
Target / label	The outcome a supervised model is asked to predict.
Population	The full set of cases the conclusion is intended to describe.
Sample	The observed subset used to learn about that population.
Schema	The names, types, constraints, and relationships that define the structure of stored data.
Provenance	Where data came from, how it was collected or transformed, and who is responsible for it.
Baseline	A simple reference method that a more complex analysis should improve upon.
Metric	A specified rule for measuring error, utility, fit, or another property of interest.
Training set	Data used to estimate model parameters.
Validation set	Data used during development to compare choices without fitting on the final test set.
Test set	Held-out data used for a final estimate of generalization under the chosen evaluation design.
Data leakage	Information reaches training or feature construction that would not legitimately be available when the model is used.
Confounder	A variable related to both an explanatory variable and an outcome that can distort a causal interpretation.
Calibration	Agreement between predicted probabilities and observed frequencies under a stated evaluation setting.
Reproducibility	Another run can recover the analysis from recorded data, code, environment, parameters, and steps.

Loss, objective, and metric describe roles, not necessarily different formulas. A loss scores a prediction; the training objective aggregates losses and may add a penalty; an evaluation metric summarizes a property on a stated population and split. For example, a classifier may minimize log loss but be evaluated by recall at a fixed review capacity. Parameters are fitted within the training procedure; hyperparameters control that procedure and are selected using validation. See Machine Learning General Concepts for examples.

Words That Need Qualification

• Average: name the statistic—mean, median, weighted mean, or something else.
• Accuracy: name the metric and class distribution; accuracy alone can hide important errors.
• Representative: state which population and sampling mechanism justify the claim.
• Significant: distinguish statistical evidence from practical importance.
• Correlation: do not silently turn association into a causal conclusion.

Prefer a precise sentence over a larger glossary. If a term cannot change how the work is performed or interpreted, it probably does not need a canonical definition here.

Continue with the topic reading guide for the complete sequence.

My reading · Opened notes are remembered in this browser.