Charta is on Google Play: practise on the go. Get the app
Charta

Distinguish unsupervised learning models from supervised ones and from reinforcement learning models.

FRM® Part I · Quantitative Analysis · Machine-Learning Methods · FRM1.QA.14.h

The idea

The three families differ in what feedback the algorithm gets. Supervised learning trains on labelled examples, inputs paired with the correct output, and learns to predict the label for new inputs. Regression predicts a number and classification predicts a category, such as default from past loans. Unsupervised learning has no labels and looks for structure in the inputs alone, for example clustering with K-means or compressing features with principal components. Reinforcement learning has no labelled answers either. An agent takes actions in an environment, receives rewards, and learns a policy that maximises cumulative reward over time, with today's action affecting future states. The test is to ask what the model learns from: labels, nothing but the data, or rewards earned through its own actions.

A practice question

From the question bank, with the reason each wrong option is wrong.

What defines a supervised learning model?

  1. It finds structure in inputs without any labels.Why not: Describes unsupervised learning.
  2. It learns from rewards received from interacting with an environment.Why not: Describes reinforcement learning.
  3. It requires no training data at all.Why not: All three learning types require data or experience to learn from.
  4. It is trained on data in which each input is paired with a known output label. - correct

Supervised models learn the mapping from inputs to labelled outputs, for regression or classification.

A harder one

Which pairing of task and learning type is correct? Task 1: predict default using labelled historical loans. Task 2: group clients by behaviour without labels. Task 3: an algorithm learns an order-execution policy by trial and reward.

  1. Task 1 unsupervised; Task 2 supervised; Task 3 reinforcement.Why not: Swaps labelled and unlabelled data.
  2. Task 1 supervised; Task 2 unsupervised; Task 3 reinforcement. - correct
  3. Task 1 supervised; Task 2 reinforcement; Task 3 unsupervised.Why not: Treats trial-and-reward learning as unsupervised and clustering as reinforcement.
  4. Task 1 reinforcement; Task 2 unsupervised; Task 3 supervised.Why not: Treats labelled prediction as reinforcement and reward-based learning as supervised.

Labels point to supervised learning; no labels and a search for structure indicates unsupervised; rewards from actions indicate reinforcement learning.

The hardest tier

A bank has 100 features on 50,000 loans, each labelled default or non-default. It first compresses the features to ten with principal components analysis, then fits a logistic regression to predict default. How should the overall approach be described?

  1. Both steps are unsupervised, because PCA is used.Why not: Ignores that the logistic regression uses default labels.
  2. Both steps are supervised, because the data carry labels.Why not: PCA does not use the labels.
  3. The approach is reinforcement learning, because the bank repeats the process each period.Why not: Repetition over time does not create rewards or an agent.
  4. The PCA step is unsupervised and the logistic regression step is supervised. - correct

PCA is applied to the features alone, so it is unsupervised. The logistic regression learns from the default labels, so it is supervised.

In the app: 6 practice questions on this outcome, spaced review that brings it back before you forget it.

Start free →

More in Machine-Learning Methods