Supervised, Unsupervised and Reinforcement Learning

The most common mistake in our first-week assignments isn't a coding error. It's a student running k-means on a churn dataset that already has a churn column, or trying to "predict" segments that nobody has ever labelled. Before you pick an algorithm you need one answer: do your rows come with the answer attached? That single question decides the toolkit, and a churn project and a segmentation project use different tools even when both start from the same customer table.

Supervised: the answer is in the table

Every training row carries a known outcome, the label, and the model learns the mapping from features to it. Two flavours:

Sub-type	Label looks like	Examples	Typical algorithms
Regression	A number	next month's electricity load, house price in DHA, delivery minutes	linear regression, gradient boosting
Classification	A category	fraud / not fraud, will churn / will stay, loan grade A/B/C	logistic regression, decision trees, random forest, SVM

The label is usually the expensive bit. A bank's fraud team may wait months for enough confirmed cases before a classifier can be trained, and we've seen projects stall for exactly that reason.

Unsupervised: no answer column at all

Here the algorithm hunts for structure on its own. It might find groups of similar rows (clustering), a compact summary of many columns (dimensionality reduction), or rows that look unlike everything else (anomaly detection). The catch is that the output needs a human to read it. k-means hands you "cluster 2" and it's your job to notice that cluster 2 is "students who order late at night".

Reinforcement: learning by being rewarded

An agent takes actions in an environment, gets a reward, and adjusts to earn more of it over time. No labelled rows, only feedback after each move. It runs game-playing systems and some dynamic-pricing engines, but it needs a safe place to experiment, and most Pakistani businesses can't afford to let an algorithm poke their live pricing. Recognise it in interviews; don't start your career with it.

The three families side by side

	Supervised	Unsupervised	Reinforcement
Input	features + labels	features only	state + reward signal
Goal	predict the label for new rows	discover structure	choose actions that maximise reward
How you check it	compare predictions with held-out labels	human judgement, silhouette score	total reward over episodes
Business question	"Which of these will X?"	"What kinds of these exist?"	"What should I do next?"
Data needed	hundreds to millions of labelled rows	unlabelled rows	a simulator or a live system you can safely poke

Sort four real requests, then fit something tiny

Imagine you're the analyst at a mobile wallet and four requests land in one morning. Family first, algorithm later.

Request	Label available?	Label type	Family
Predict which agents will stop transacting next month	Yes (past inactivity)	category	supervised, classification
Estimate cash-out volume per agent next week	Yes (past volumes)	number	supervised, regression
Group agents into behaviour types for a marketing plan	No	-	unsupervised, clustering
Decide the cash-back offer each user sees to maximise usage	Only reward after showing it	-	reinforcement (or an A/B test first)

Now the smallest supervised classifier we can do by hand. Six agents, one feature (transactions last month), one label (active next month).

Agent	Txns last month	Active next month?
A	5	No
B	12	No
C	30	Yes
D	18	Yes
E	9	No
F	45	Yes

The simplest classifier is a threshold. Sort by transactions: 5 (No), 9 (No), 12 (No), 18 (Yes), 30 (Yes), 45 (Yes). Any cut between 12 and 18 separates the classes perfectly, so take the midpoint 15. Your "model" reads predict active if transactions > 15, and a new agent with 14 transactions is predicted inactive. Real data overlaps, which is why later lessons bring in probabilities, but the shape of the task is identical.

# run locally
from sklearn.tree import DecisionTreeClassifier
from sklearn.cluster import KMeans

X = [[5], [12], [30], [18], [9], [45]]
y = [0, 0, 1, 1, 0, 1]                        # supervised: label present
clf = DecisionTreeClassifier(max_depth=1).fit(X, y)
print(clf.predict([[14], [20]]))              # -> [0 1]

km = KMeans(n_clusters=2, n_init=10, random_state=0).fit(X)   # unsupervised: no y
print(km.labels_)                             # group ids, meaning is yours to decide

Turning "we want AI" into three projects: A Faisalabad textile exporter asked us for "AI to improve sales". We split it into three real pieces of work. A regression to forecast monthly yarn demand per buyer (supervised). A clustering of 400 buyers by order size, frequency and payment delay (unsupervised). A classifier for which invoices will be paid more than 30 days late (supervised). Each got its own data, metric and owner, and the finance and sales heads could sign off on each one separately. The vague request was the real problem, not the maths.

Quick recap

• Ask "is there a label?" before anything else.
• Supervised predicts a known target, regression for numbers and classification for categories.
• Unsupervised finds structure, and a person has to interpret it.
• Reinforcement learns actions from reward and is rarely the first thing a business should try.
• One vague request usually hides two or three well-defined ML problems.

Homework

1. Label each as regression, classification, clustering or reinforcement: predicting a car's resale price on PakWheels; grouping Careem riders by travel habits; deciding whether a CNIC scan is readable; choosing ad bids in real time.
2. Add agent G with 16 transactions who went inactive. Does one threshold still separate the classes? What would you do about it?
3. Write one sentence describing the "label" for a K-Electric project that predicts transformer failures.