INFS3822 Chap.2 Learning Paradigms and ML Tasks
Learning Paradigms and ML Tasks
Learning paradigms differ by the feedback supplied during training. Labelled examples support supervised prediction, unlabelled examples support pattern discovery, mixed labels support semi-supervised learning, and rewards support reinforcement learning.
The chapter then routes output shapes to regression, classification, ranking, clustering, anomaly detection, recommendation and generation so that method selection begins with evidence rather than software.
What this chapter covers
- 01
Supervised learning from labels
- 02
Unsupervised pattern discovery
- 03
Semi-supervised pseudo-labelling
- 04
Reinforcement learning from rewards
- 05
Regression and classification outputs
- 06
Ranking, clustering and anomaly detection
- 07
Recommendation and generative tasks
Worked example · free
Route three business goals to task families
- +1Monthly spend is a continuous numeric target, so frame it as regression.
- +1Profiles have no supplied correct label, so begin with unsupervised clustering and validate the groups.
- +1Ordering offers is a ranking or recommendation task because the desired output is an ordered set of choices.
Key terms
- Labelled data
- Examples paired with known target outcomes for supervised learning.
- Unlabelled data
- Observations without supplied target outcomes, used to discover structure.
- Semi-supervised learning
- Learning from a small labelled set and a larger unlabelled set.
- Reinforcement learning
- Learning a policy from states, actions and rewards over sequential interactions.
Learning Paradigms and ML Tasks FAQ
How do I distinguish supervised and unsupervised learning?
Ask whether every training example has a known target. If it does, the model can learn supervised prediction. If it does not, clustering, anomaly detection or representation learning can discover structure, but a person still must interpret whether that structure supports the business decision.
Is clustering a form of classification?
Not initially. Classification learns from known class labels, while clustering creates groups from similarity in unlabelled evidence. Analysts may later name or validate clusters, and reviewed cluster assignments might become labels for a later classifier, but the original task remains unsupervised.
When is reinforcement learning appropriate?
Consider it when actions change later states and success depends on cumulative reward across a sequence. A one-step choice from historical labelled examples may be handled by classification, uplift modelling or ranking. Reinforcement learning also requires careful treatment of exploration and delayed consequences.
Can one dataset support several task families?
Yes, because the intended output and available feedback define the task. The same customer records might support numeric spend prediction, a high-spend class, a ranked contact list or unlabelled profiles. Each version needs its own target or similarity definition, evaluation method and action. State those differences rather than assuming the dataset has one natural algorithm.
Assessment move
Build a two-column drill. In the first column write the available feedback: target labels, no labels, partial labels or sequential rewards. In the second write the required output: number, category, order, group, anomaly or next item. Practise explaining why a plausible alternative task would answer a different question. Add a third column for evaluation.
Numeric prediction suggests errors in target units, category prediction suggests a confusion matrix and class-specific measures, ranking suggests order-sensitive measures, and clustering requires stability plus external interpretation rather than labelled accuracy by default. Route at least ten one-sentence business goals through all three columns.
Use near-misses to sharpen the boundary: predicting spend is regression, predicting whether spend exceeds a threshold is classification, ordering customers by expected spend is ranking, and discovering groups without a target is clustering. The same records can support different tasks because the intended output changes the question. For paradigms, ask what feedback exists during learning.
Labels for every example support supervised learning; no supplied targets support unsupervised structure finding; a small labelled set combined with many unlabelled cases can motivate semi-supervised learning; states, actions and delayed rewards define reinforcement learning. Do not route by the name of an algorithm alone.
A neural network can support several task families, and clustering is not synonymous with every unsupervised representation method. After each route, state the row or decision unit, the time at which input becomes available and the action that consumes the result. This prevents a technically valid task from answering an unusable question. Practise explaining why an alternative is wrong without calling it inferior.
For example, clustering customers may reveal profiles but cannot directly estimate a labelled churn probability; classification can estimate that probability but does not discover every useful unlabelled structure. End with one mixed case containing a forecast, a segment and a recommended next item.
Separate it into distinct analytical tasks, identify what evidence each needs and describe how their outputs could be combined without pretending they are the same model.
Working through Learning Paradigms and ML Tasks in INFS3822? Sia is AskSia’s AI Computer Science tutor — ask any INFS3822 Learning Paradigms and ML Tasks question and get a clear, step-by-step explanation grounded in how INFS3822 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
