Description
Summarises many correlated variables in a few orthogonal components that capture the most variance, revealing the main axes along which users differ.
PCA rotates a standardised data matrix into uncorrelated components ordered by variance explained. The first few components often capture most of the structure in behavioural or survey data, for instance an activity-intensity axis and an exploration-versus-routine axis. Unlike factor analysis, PCA models total variance and has no latent-variable claim. It is used for compression before clustering, visualising user spaces, removing collinearity and detecting dominant behavioural dimensions. Rigour requires standardising variables, choosing components by parallel analysis or cross-validated reconstruction rather than arbitrary cut-offs, interpreting loadings carefully, and recognising that PCA captures only linear structure and is sensitive to outliers and scaling.
Attributes
- Discipline
- Computational & Machine Learning
- Loss type
- Active
- Goal
- Segment users, Synthesise data
- Stage
- Discovery, Post-launch
- Data
- Quantitative
- Effort
- Hours
Steps
- Construct the feature matrix and standardise each variable to mean 0 and SD 1; log-transform heavily skewed counts first.
- Run PCA and inspect the scree plot and cumulative variance explained.
- Choose the number of components with parallel analysis or a variance target (often 70-90%), justified for the purpose.
- Interpret components using loadings above about 0.4 in absolute value and name them by their dominant features.
- Plot users in component space, colour by known outcomes or segments, and check outlier influence.
- Use component scores downstream (clustering, regression) and report variance retained.
Sources
- Pattern Recognition and Machine Learning (Christopher Bishop)
- Unsupervised Learning in Space and Time (Marc Leordeanu)
- Jolliffe, Principal Component Analysis (2002)
- Pattern Recognition & ML (Bishop)
- Unsupervised Learning in Time & Space
- The Inquiry Series, Ch 52 — Dimensionality Reduction — The Geometry of Behavioral Space
- Practical Statistics for Data Scientists (Bruce & Bruce)
Also known as
- Principal component analysis
- Karhunen-Loève transform
Categories
- Analytics & Behaviour
- Computational & Machine Learning