Your cart is currently empty!
CAISA · Diagnostic Tool
Stage 1 of 11
Live mode — concise reveals
The Hook
Y = F(X)
What makes machine learning supervised?
A
The Y values are confirmed and labelled in advance
B
The model is derived from historical patterns
C
Both A and B
✓ The answer is A
Confirmed Ys is the precondition. Historical patterns describe all machine learning — it’s not what makes learning supervised.
Most learners pick C — and it’s a reasonable instinct. But “learning from historical patterns” describes unsupervised learning too. Clustering, association mining, embeddings — all derived from history. The word supervised literally means a teacher provided the answers. That teacher is the labelled Y.
The Principle
Supervised learning needs ONE thing
A confirmed Y attached to every training example.
Supervised
Training data = historical (X, Y) pairs. The model learns the mapping F so it can predict Y for new X.
Unsupervised
Training data = historical X only, no Y. The model finds structure within the data itself.
The diagnostic question for any project: “In our historical data, do we have the answer recorded for each example?”
Unpacking the Precondition
Who is the supervisor?
The word “supervised” comes from a teaching metaphor. So who’s actually doing the supervising?
A
The data scientist who builds the model
B
A human labeller (e.g., a radiologist who diagnosed each X-ray)
C
A business process (e.g., past loans with repayment outcomes recorded)
D
The world itself (e.g., yesterday’s actual temperature)
E
Any of B, C, or D — anyone or anything that produced the answers
✓ E — Anyone (or anything) that produced the answers
The supervisor isn’t a job title. It’s whoever — or whatever — attached the correct Y to each historical example, before training began.
Common trap: assuming “supervisor = data scientist.” The data scientist uses the supervised data to build the model — they’re not the supervisor. The supervisor is the source of the labels: the radiologist, the fraud team, the loan officer, the recording of yesterday’s temperature, the user clicking “spam.” Sometimes there’s no single supervisor at all — the world simply made the answer known, and someone recorded it.
Unpacking the Precondition
What is the supervisor’s job?
Now that we know who they are — what do they actually do?
A
To train the model
B
To choose which features (Xs) matter
C
To validate the model’s predictions after training
D
To attach correct answers (Ys) to historical data, before training begins
✓ D — Attach correct answers, before training begins
The supervisor’s job is upstream of the model. They produce the labels. Once that’s done, the model can begin learning F from (X, Y) pairs.
The other options are tempting because they sound technical: training, feature selection, validation. But those are jobs for the data scientist or the model itself — they happen after the supervisor’s work is done. The supervisor’s contribution is foundational: without confirmed Ys, supervised learning has nothing to learn from.
Concrete Picture
What does training data actually look like?
A loan default predictor. Each row is one historical example. Inputs in grey. The confirmed answer in teal.
X — features (inputs)
Y — confirmed answer (label)
| Age | Income (₹) | Credit Score | Years Employed | Defaulted? |
|---|---|---|---|---|
| 32 | 8,50,000 | 720 | 6 | No |
| 45 | 12,00,000 | 680 | 12 | No |
| 28 | 4,50,000 | 590 | 1 | Yes |
| 51 | 15,00,000 | 750 | 20 | No |
| 35 | 6,00,000 | 610 | 3 | Yes |
Three things to notice:
- The whole table is the training data — not just the Y column.
- X and Y are paired, row by row — that’s what lets the model learn the relationship.
- Each Y is a known fact, not a guess — the loan was either repaid or written off; the record is settled.
The model’s job: learn from these rows so that when a new applicant arrives (a new X), it can predict their Y.
Round 1
Is this supervised learning?
Four scenarios. For each, decide whether the precondition (confirmed Ys) is met. Answers may surprise you.
Scenario A
“These are the Xs. These are the Ys. Train the model on this data.”
Both inputs and outputs are confirmed. The supervisor’s work is done.
Is this supervised learning?
1Supervised
2Partly supervised
3Not supervised
✓ Supervised — textbook
Confirmed Xs and confirmed Ys. The model’s only job is to learn F.
Example: 10,000 past loans with applicant features (X) and default outcomes (Y). Build a default predictor. This is the canonical case.
Scenario B
“These are the Ys. Here are 150 suspected Xs. Train the model and identify which Xs matter.”
Outputs are confirmed. Inputs need filtering.
Is this supervised learning?
1Supervised
2Partly supervised
3Not supervised
✓ Still supervised — with feature selection
The Ys are confirmed. That’s what matters. Sorting which Xs carry signal is feature selection — a normal part of supervised learning.
Example: “Here’s whether each customer churned (Y). Here are 150 things we know about them. Figure out what predicts churn.” Techniques like LASSO and tree-based feature importance handle this. Almost every real supervised project includes this step.
Scenario C
“Suspected Ys, suspected Xs. Find which can be confirmed, then train the model.”
Labels themselves are uncertain.
Is this supervised learning?
1Supervised
2Partly supervised
3Not supervised
⚠ Not pure supervised — it’s a hybrid
You’re doing two jobs: validating labels, then training the model. Job one is label engineering — weak supervision territory.
Real techniques for this: weak supervision (Snorkel-style), semi-supervised learning, active learning. Once labels are confirmed, the second step becomes supervised. But the overall pipeline is broader than supervised learning.
Scenario D
“Suspected Ys, suspected Xs. Confirm if possible, then train the model. If nothing can be confirmed, say no model exists.”
We’re not even sure a relationship is there.
Is this supervised learning?
1Supervised
2Partly supervised
3Not supervised
✗ Not supervised learning at all
This is hypothesis testing. Supervised learning assumes a learnable relationship exists. Here, that’s the question being asked.
Statistical testing, causal inference, model validation, and the willingness to conclude “no, there’s nothing here.” A good practitioner runs this rigorously and is prepared to return: “No reliable signal found. No model should be built.” Supervised learning may or may not be a downstream output.
Round 2
So what is unsupervised learning?
Same format. Four scenarios. Watch for the trap — supervised-learning thinking sneaks back in.
Round 2 · Which is unsupervised?
Pick the one that truly fits.
A
“Suspected Ys and Xs — find which can be confirmed, build the model.”
B
“Suspected Ys and Xs — confirm if possible. If not, no model.”
C
“This is the data. Find if Ys and Xs exist. If yes, learn from them.”
D
“This is the data. Find if patterns exist. Ys and Xs aren’t required.”
✓ D is unsupervised learning
A and B are still chasing supervised goals. C imports the X/Y framing. Only D drops the assumption that Ys exist at all.
In unsupervised learning, the very concepts of X and Y don’t apply. There are just data points. The model isn’t predicting anything — it’s finding structure: clusters, associations, or compressed representations. That’s the cleanest definition.
The Subtle Insight
In unsupervised learning, if X→Y relationships exist, will they be explored?
Three sequential reveals. Click Reveal answer to step through them.
1
Does the algorithm know which variable is Y?
— think first, then reveal —
No. Every variable is just a feature. The model has no concept of a target.
2
After clustering, can humans see Cluster 3 has 80% diabetes rate?
— think first, then reveal —
Yes. The structure is now visible. Patterns are inspectable.
3
So who found the X→Y relationship?
— think first, then reveal —
The human — looking at unsupervised output. The algorithm produced structure; the meaning came afterwards.
Unsupervised learning doesn’t explore X→Y relationships — but the structure it surfaces may reveal them.
The X/Y interpretation is something humans impose on unsupervised output afterwards. Often as a hypothesis, then operationalised by a downstream supervised model.
Synthesis
The architect’s mental model
Supervised
Confirmed Ys. Predict a known target for new inputs.
Unsupervised
Just data. Find structure — clusters, associations, representations.
The Pipeline
Unsupervised discovers → humans interpret → supervised operationalises.
Project Scoping
First fork in the road: “Are we predicting a known outcome?” If yes — supervised territory. If no — unsupervised. If we’re not sure either applies — we’re at the problem framing stage, not yet at ML.