Course contentsShow
← Computer Science
Lesson 2793 of 2,87262. Machine Learning FundamentalsFree lesson

Choosing the Right Learning Paradigm

Determining which learning approach fits a problem based on available data and desired outcomes.

Choosing the Right Learning Paradigm

You'll learn: How to match a machine learning problem to the appropriate learning approach based on your data and objectives.

The Decision Framework

Now that you've seen supervised, unsupervised, semi-supervised, and reinforcement learning, how do you choose? The decision hinges on three key questions:

1. What data do you have?

  • Labeled examples (input-output pairs)? → Supervised learning is your go-to. Use classification if predicting categories, regression for continuous values.
  • Unlabeled data only? → Unsupervised learning. Apply clustering to group similar items or dimensionality reduction to simplify your feature space.
  • Some labeled, mostly unlabeled? → Semi-supervised learning lets you leverage both.
  • No fixed dataset, sequential decisions? → Reinforcement learning, where an agent learns through trial-and-error interactions.

2. What's your goal?

Are you predicting specific outcomes? Finding hidden structure? Optimizing a sequence of actions over time? Your objective clarifies which paradigm fits naturally.

3. What feedback is available?

If you know the "right answer" for your training examples, supervised methods shine. If you only get reward signals after many actions, reinforcement learning is built for that. If there's no feedback at all, unsupervised methods discover patterns independently.

Think Like a Doctor

Analogy: Choosing a paradigm is like diagnosing treatment. A broken bone (clear problem, clear solution) = supervised. Understanding a patient's overall health patterns without specific complaints = unsupervised. Learning to perform surgery through practice and feedback = reinforcement learning.

Key Takeaway: Match the learning paradigm to your data availability and problem structure—labeled data enables supervised learning, unlabeled data calls for unsupervised methods, and sequential decision-making with delayed rewards points to reinforcement learning.