Data Science

Data science is an evidence chain from a question and observations to an action. A model can optimize a number; it cannot decide by itself whom a sample represents, how labels were produced, or what errors cost.

Working Path

1. Define question and action: which decision, estimate, or explanation will the result support?
2. Map the data-generating process: who is observed, when, what is missing, and how do labels arise?
3. Review the dataset: record provenance, unit, license, sensitivity, version, and transformations in the Dataset map.
4. Choose the split before modeling: prediction time, entities, and deployment determine random, group, time, or external splits; see Data Splits and Leakage.
5. Build a simple baseline: constants, rules, linear models, or current practice establish incremental value.
6. Train and select: fit every learned preprocessing step on training folds; validation selects, test remains outside that loop.
7. Evaluate use, not a leaderboard: include thresholds, costs, calibration, slices, and uncertainty; see Evaluation Under Distribution Shift.
8. Communicate bounded conclusions: report failures, exclusions, data limits, and conditions that do not support extrapolation.
9. Monitor after deployment: inputs, delayed labels, feedback, and human process change the distribution.

Exploration and confirmation may use the same data only with explicit accounting. Once test results influence features, thresholds, or exclusions, the test has become development information; obtain fresh data, nest the evaluation, or label the result exploratory.

Reading Lines

Line	Note	Outcome
vocabulary	Working Vocabulary	precise observation, target, population, metric, and provenance terms
evidence	Datasets	provenance, leakage, splits, shift, metrics, and reporting
model concepts	Machine Learning	losses, baselines, generalization, and model families
architecture	Deep Learning	representation and computation without substituting architecture for evaluation

question.md       question, prediction time, action, and loss
provenance.yaml   source, version, license, unit, and transformations
split.py          executable, tested partition rule
pipeline.*        train-only preprocessing and model
results/          every run, seed, metric, slice, and failure
report.md         conclusion, interval, limitation, and monitoring

This path is biased toward tabular supervised tasks, computable metrics, and reproducible experiments. Causal identification, qualitative research, participatory design, privacy, governance, and power cannot be reduced to another metric. Public, official, or large data is not automatically valid or fair.