Start With The Signal, Not The Algorithm
Suppose we have Titanic passenger rows. If each row includes Survived, we can learn to predict a known target. If we remove that label, we can still search for structure, but cannot directly measure survival-prediction error. If a system chooses actions that change what happens next, a static labelled table is no longer the whole problem.
Ask what feedback is available, when it arrives, and whether actions change future data.
Retrieval Warm-Up
- Identify the feature, label, and identifier in a survival dataset.
- Why is Pclass ordinal rather than numerical?
- Which split may be used repeatedly to choose a model?
Supervised
Examples contain targets y.
Predict survival, price, spam, demand.
Unsupervised
Examples contain x but no answer y.
Find clusters, compressed representations, anomalies.
Reinforcement
An agent acts, observes a new state, and receives reward.
Games, robotics, sequential control.
Problem Sorter
Predict survival from passenger records
Formal Objects
Supervised: data pairs (x, y). Training learns a mapping from features to target; regression predicts quantities and classification predicts categories.
Unsupervised: observations x only. The objective defines “structure”—compactness, reconstruction, density, or similarity—so different objectives can produce different valid patterns.
Reinforcement: at time t, an agent observes state s_t, chooses action a_t, receives reward r_t, and moves to s_(t+1). Feedback can be delayed, and exploration changes collected data.
Worked Decisions
Predict survival: supervised classification because every historical passenger has a binary label.
Discover passenger archetypes: unsupervised because no authoritative “archetype” column exists. The result describes structure; it is not automatically useful or fair.
Choose which tutorial hint to show over a learning session: potentially reinforcement learning because the action changes later learner state and reward is delayed. If we only predict whether a fixed hint was clicked from historical labels, that narrower task is supervised.
Related Settings
A small labelled set plus a larger unlabelled set.
Create supervision from data itself, such as predicting hidden words.
A model chooses valuable examples for humans to label.
Update as examples arrive; it can still be supervised.
These names describe data and feedback arrangements. “Online” does not mean reinforcement, and “unlabelled” does not mean the model discovers objective truth.
Diagnose The Learning Setting
def diagnose(has_target, system_acts, feedback_changes_future):
if system_acts and feedback_changes_future:
return "reinforcement / sequential decision problem"
if has_target:
return "supervised learning"
return "unsupervised learning"
print(diagnose(has_target=True, system_acts=False, feedback_changes_future=False))
# supervised learningCommon Traps
“Clustering predicts a label.”
Clusters are induced by an objective and representation; they are not pre-existing labels.
“Any recommendation system is reinforcement learning.”
Many are trained with supervised interaction labels; RL becomes relevant when sequential action consequences are explicitly modelled.
“No labels means no evaluation.”
Evaluation is still required, but must use stability, reconstruction, downstream utility, expert judgement, or other task-appropriate evidence.
Mastery Check
1. Classify six new scenarios by feedback signal, not buzzwords.
2. Contrast supervised labels with reinforcement rewards.
3. Explain why unsupervised structure depends on an objective.
4. Give one problem that can be framed in two learning settings.
5. Tomorrow retrieve the three formal objects; in three days redo the sorter without feedback.
Next: Inspect Before Predicting
We know the task and data types. Before fitting any model, we will interrogate a real dataset for quality problems, distributions, and patterns.