One retailer, three jobs
A retailer has purchase records and wants three things: group similar customers, predict a customer's next purchase, and choose an offer that improves long-run value. Those are related goals, but do they all use the same kind of learning signal?
This is Module one, Lesson thirteen: Different learning setups solve different problems. By the end, you should be able to distinguish supervised, unsupervised, self-supervised, reinforcement and imitation learning, and match a setup to a clearly stated task without assuming the label guarantees reliable results.
We'll start with what target or feedback exists, walk through each setup, then return to the retailer's three jobs.
Use supervised input-target examples
Supervised learning uses examples with target information. Our delivery-cost model learns from inputs and recorded costs. Predicting a numerical cost is regression. Predicting fraud or legitimate is classification. Both can be supervised when examples carry suitable targets.
Inspect how the target was created
A target need not be a person manually typing a label. It could be a measured outcome or another constructed signal. Its usefulness depends on how it was created and whether it represents the intended task. A shop can have thousands of cost records, yet inconsistent definitions of cost can still guide learning toward the wrong thing.
Match the target to the wanted outcome
Ask two questions: what target does the model see, and how closely does that target match the real outcome the organization cares about?
Find structure under a chosen representation
Unsupervised learning uses data without the same explicit input-to-target labels and seeks patterns or structure. Clustering groups examples according to a defined notion of similarity. The retailer might group customers with similar purchase patterns under a chosen representation.
Validate candidate groups
The groups do not arrive with guaranteed real-world meanings. A cluster may reflect spending frequency, marketing exposure or data collection artifacts. Calling it a valuable customer segment requires validation: do members share a useful property, and does acting on that grouping help?
Inspect compressed representations
Dimensionality reduction creates a representation with fewer dimensions while retaining selected structure. It can help exploration or later modeling, but the compressed view can also hide details. The method's output is a tool for investigation, not a discovery certificate.
Construct targets from the original data
Self-supervised learning creates a learning signal from the data itself. For text, a system can hide a piece and learn to predict it, or learn to predict a subsequent part. The original data supplies the target construction. It does not mean the model has no target or objective.
Separate the training objective from downstream usefulness
This can support learning at large scale because people do not need to label every example for a particular downstream task. Yet predicting missing text is not the same as helping a customer, giving a correct medical answer or obeying an organization's policy. The training objective teaches selected relationships in the data; later evaluation must show whether those relationships serve the intended job.
Learn from actions and rewards
Reinforcement learning involves a policy selecting actions under a defined state or observation, receiving feedback or rewards, and optimizing behavior over interactions. A policy is the rule for choosing actions given what is available. Reward is the optimization signal. It can be imperfect: a short-term click reward might not represent a customer's long-run satisfaction.
Learn from demonstrations
Imitation learning uses demonstrations of desired behavior. A human or another process shows actions, and the learner tries to reproduce them. That can be useful when examples of good action exist, but the learned policy may encounter situations missing from the demonstrations.
Account for changed and unfamiliar states
In a physical system, a mistaken action can change the environment and lead to even less familiar states. A robot that places an object poorly may then see a scene unlike anything in its demonstrations. Later parts of the course develop that interaction. Here, the warning is simply that success on demonstrations does not prove safe behavior in every new situation.
Combine learning setups in a pipeline
A model pipeline may begin with self-supervised learning over large amounts of data, then adapt using supervised demonstrations or preference information, and use a reward-based stage for selected behavior. Describing the final model as trained with reinforcement learning names one part but not the entire history.
Require separate reliability evidence
The same caution applies to reliability. The training setup tells us where the learning signal came from, not whether labels were sound, rewards aligned with human goals, new conditions are covered or deployment is safe. Those claims need separate evidence.
Return to retailer grouping
Return to the retailer. Grouping customers by purchase patterns is a clustering task under a chosen representation. It may reveal candidate structure, but a useful segment still needs validation.
Predict a later purchase using earlier inputs
Predicting next purchase can be a supervised predictive task if previous records and a well-defined observed next purchase provide input-target examples. The precise setup depends on how targets are constructed and when inputs are available.
Define offer decisions and feedback
Choosing an offer to maximize long-run value is a decision problem. If the retailer can take actions, observe outcomes and learn from interaction, a bandit or reinforcement-learning formulation may be relevant. A bandit is a narrower choice-and-feedback setup where actions are compared under limited observation of alternatives. But a reward measured as clicks might misrepresent long-run value, and exploration has customer costs. Define the decision and consequences before selecting a method.
Check your understanding
Why does self-supervised learning not mean that training has no objective or targets?
The learning signal is constructed from the data itself. Missing or subsequent pieces can become targets, and a loss guides the model toward relationships useful for that objective. Those targets may still differ from the downstream customer goal.
If a clustering method produces five customer groups, have we proven there are five meaningful business segments?
No. The groups follow a chosen representation and similarity rule. We must inspect their stability, interpretation and usefulness for the intended decisions.
Connect the learning signals
Supervised learning uses target examples. Unsupervised learning seeks structure without that same target pairing. Self-supervised learning constructs targets from the data. Reinforcement learning learns from actions and rewards; imitation learning uses demonstrations. Pipelines can combine them.
Continue to decisions and costs
The problem definition comes first. Lesson fourteen closes this module by carrying model results through a specific decision and cost calculation. We will ask whether a statistical improvement actually improves what the organization cares about.