What data science actually is
Not a toolbox but a loop: turn a decision into a question, data into evidence, and evidence into a choice someone can defend.
A hospital wants to know which discharged patients will be readmitted within thirty days. A telecom wants to know which subscribers are about to churn. A factory wants to know which engine on the line is about to fail. In each case the interesting object is not a model — it is a decision that someone will make, and the question is what evidence would justify making it.
Data science is the discipline of closing that gap. It borrows statistics for inference, computer science for scale, and domain knowledge for meaning, but the thing that holds the three together is a loop: frame the decision as a question, find and prepare the data that bears on it, look before you model, fit something, and then be honest about how much the answer can bear.
The unit of work is a decision
A model that predicts readmission with 90% accuracy is worthless if nobody can act on it. Start from the action, work backwards to the quantity that would justify the action, and only then ask how precisely you can estimate that quantity. The analysis is downstream of the decision.
The loop below is not a waterfall. In real projects you cycle — a failed evaluation sends you back to prepare the data, a surprising plot sends you back to reframe the question. But the dependencies only run one way: you cannot evaluate a model you never fit, and you cannot fit data you never understood.
Click a stage to inspect it, then run the stages in dependency order
Ask
Turn a vague wish into a falsifiable question tied to a decision: who acts on the answer, and what changes if it is wrong?
Produces: a decision to support
Common pitfall: Starting from data you happen to have instead of the decision you need to make.
Output chain
Nothing produced yet. A report is only as trustworthy as the stages underneath it.
The arrows are the point: in practice you loop back (a failed evaluation sends you to Prepare, not Model), but the dependency is one-way — you cannot evaluate a model you never fit, and you cannot fit data you never cleaned.
Frame the quantity, then estimate it
Suppose the decision is whether to enrol a patient in a costly follow-up programme, and that a readmission costs while the programme costs and halves the risk. The break-even is not "accuracy 90%" — it is a threshold on the probability of readmission, found by comparing expected costs:
where . Enrolling wins when . The threshold falls out of the decision, not out of a default of 0.5. This is why the first stage is "ask": the target and the metric are consequences of the action you are trying to support.
The three overlapping skills
The classic picture has three circles — statistics, computing, and domain knowledge — and data science as their overlap. The picture is crude but the warning is real: a brilliant statistician who cannot move data, or a strong engineer who does not know what the columns mean, will produce confident nonsense. Modern practice adds a fourth concern, decision-making: knowing which errors are expensive and who bears them.
- Data engineering builds and maintains the reliable supply of clean data.
- Data analysis answers a specific question, usually for a human reader.
- Machine learning builds a system that generalises to new cases on its own.
- MLOps keeps that system correct and available after it ships.
These are roles along a spectrum, and in a small team one person wears all the hats. What matters is noticing which hat a task needs — the failure modes are different, and the next chapters take them one at a time.
Careful
"We will make it data-driven" is not a question. If no decision changes depending on the answer, the analysis is decoration. And beware the reverse trap: a beautiful dashboard with no owner is not evidence either. Name the decision, name the person who makes it, and name what would change their mind before you open a notebook.
Where models are the wrong answer
Sometimes the honest output of a data-science loop is "this is a lookup table", "we need better data collection", or "the decision rule is already optimal". A model earns its complexity only when the relationship is too intricate to write down and there is enough data to estimate it. A one-line rule that a domain expert can audit often beats a black box that is two percent better and impossible to defend.
Illustrative vs real
The process loop in the artifact is a schematic, not a measurement of how time is spent — and the cost calculation above assumes clean, known probabilities. In practice you estimate with uncertainty, the costs are contested, and much of the work is negotiating those numbers. The math for the threshold is exact; the inputs never are.
Check yourself
Eduspheria wiki · Data, MLOps & Deployment, Data science foundations
0 / 4 answered
From the mid-term paper
Modeled on NITJ AI-505, Mid-Term October 2024
0 / 5 answered
Where next: exploratory data analysis — how you look at a dataset before you are allowed to model it.