University College London · FACULTY OF STATISTICS

BENV0139 Chap.6 PCA, Factor Analysis and Clustering

- one subject, every graph, every model, every mark
7 Chapters3-page Bible
Our own words - no uploaded lecturer files
Updated for this semester
Chapter 6 of 8 · BENV0139

PCA, Factor Analysis and Clustering

PCA, Factor Analysis and Clustering separates three multivariate tasks that are often blurred. PCA constructs orthogonal directions capturing decreasing amounts of total variation. Loading plots show how variables align with those directions, while score plots place samples or objects.

Variables pointing together tend to co-vary; opposing directions suggest negative association; short vectors contribute little to the displayed axes. Scaling matters because variables with large numerical variance can dominate. Factor analysis instead models shared covariance through latent factors and variable-specific variation.

Clustering groups observations according to chosen variables, transformations, distances and algorithms. Each method can reveal structure without proving a causal mechanism or natural category. Strong interpretation reports preprocessing, variance retained, stability and the heritage meaning used to label components or groups.

In this chapter

What this chapter covers

  • 01

    Principal component directions

  • 02

    Loading plots

  • 03

    Score plots

  • 04

    Explained variance

  • 05

    Scaling and standardisation

  • 06

    Latent factor interpretation

  • 07

    Distance, cluster stability and ethical classification

Worked example · free

Interpret a paper-condition loading plot

Q [8 marks]. PC1 loadings are strength 0.72, polymer length 0.65 and acidity -0.68. PC2 has age 0.74. Interpret the axes if both explain 61%. The marks are AskSia practice weighting, not a published UCL marking scheme.
The loading plane places stability and strength on the positive side of PC1, acidity in the opposite direction and age mainly along PC2.The loading plane places stability and strength on the positive side of PC1, acidity in the opposite direction and age mainly along PC2. PC1PC2 stability strength acidity age
The loading plane places stability and strength on the positive side of PC1, acidity in the opposite direction and age mainly along PC2.
  • 2Group variables by loading direction and length. The marks are original practice weighting.
  • 2Describe PC1 as material integrity versus acidity, using cautious interpretive language.
  • 2Describe PC2 as mainly age-related in the supplied loading set.
  • 2State that 39% of variation lies outside the plane and that covariance does not prove causation.
PC1 contrasts strength and polymer length with acidity, while PC2 is age-related. The two axes are a 61% compression, not a complete causal explanation. Practice marks are not official.
Sia tip — Read positive and negative loading directions before inventing a component label, then check the unexplained variance.
Glossary

Key terms

Loading
A coefficient linking an original variable to a component or factor direction.
Score
The coordinate of a sample or object after projection onto a component or factor.
Standardisation
Rescaling variables to a common reference, often zero mean and unit standard deviation.
FAQ

PCA, Factor Analysis and Clustering FAQ

How do I interpret a loading plot?

Start with vector direction and length. Similar directions indicate positive association in the displayed plane, opposite directions indicate negative association, and near-right angles suggest weak relation there. Then use heritage knowledge to propose a cautious axis label.

Why does scaling matter in PCA?

PCA based on covariance gives more influence to variables with larger numerical variance. Standardisation can give differently measured variables comparable influence, but it also changes the scientific weighting. Report the choice and justify it from the question.

Does a cluster represent a real category?

Not automatically. An algorithm produces groups under chosen variables, scales, distances and settings. Test stability, compare alternatives and validate with external knowledge. Use descriptive labels and consider harms before treating a computed group as a fixed identity.

Study strategy

Exam move

Keep separate cards for PCA, factor analysis and clustering. On the PCA card, write preprocessing, covariance or correlation basis, explained variance, loadings, scores and interpretation. On the factor card, distinguish shared from variable-specific variation and avoid calling every component a latent cause. On the clustering card, list scale, distance, algorithm, number of groups and stability checks.

Practise loading plots by describing geometry before meaning. Cover the labels and state which variables point together, oppose each other or contribute little, then restore heritage context and propose an axis name. Always report how much variation the displayed components retain. Compare analyses before and after standardisation to understand scale sensitivity.

For clusters, vary the distance or starting values and describe whether memberships persist. Add an ethical audit that asks who is represented, who assigns labels and what action follows. End with a boundary sentence: multivariate structure summarises patterns in the analysed data and needs separate evidence for mechanism or natural categories.

Redraw a loading plane with axes, arrow lengths and opposing directions, then reverse every sign to verify that the interpretation is geometrically equivalent. Write separate captions for a loading plot and a score plot so variables are never confused with objects. Repeat PCA with raw units and standardised variables; list which measurements gain or lose influence and justify the version aligned with the heritage question.

For factor analysis, record extraction, rotation, loading threshold and factor-specific variance before naming any latent construct. For clustering, standardise deliberately, choose a distance and linkage, vary the cut and describe whether groups persist. Treat dendrogram height as dissimilarity rather than probability. Compare memberships with material and provenance information that was not used to create the groups.

Add a harm review: who could be misrepresented by a label, which historical classifications require special caution and what decision will use the grouping? Finish with preprocessing, retained variation, stability, cautious label and a sentence separating covariance structure from causal mechanism or natural category.

Working through PCA, Factor Analysis and Clustering in BENV0139? Sia is AskSia’s AI Statistics tutor — ask any BENV0139 PCA, Factor Analysis and Clustering question and get a clear, step-by-step explanation grounded in how BENV0139 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.

A+Everything unlocked
Unlocks this Bible + all 32 of your subjects at University College London, plus 1,000+ more Bibles across the AskSia library.
Sia - your BENV0139 tutor, unlimited, worked the way the exam marks it
The full 3-page Bible + practice bank with worked solutions
Chrome extension - sync your LMS so Sia knows your deadlines
Bilingual EN / Chinese on every Bible and every Sia answer
$0.99 Trial
30-day money-back · cancel in one tap · how it works