Description
Transforms raw event logs into a user-by-feature matrix of meaningful behavioural measures that can feed clustering, modelling and dimensionality reduction.
Every computational behavioural analysis depends on how raw events are turned into features. This method defines the unit (user, session, week), the observation window, and families of features: frequency, recency, intensity, breadth of feature use, sequence patterns, timing and outcome-related measures. Good features are grounded in qualitative understanding of what behaviours mean, robust to logging changes, and normalised appropriately. Rigour lies in documenting a data dictionary, handling sparsity and skew (log transforms, ratios), avoiding leakage of outcomes into features, checking correlations and redundancy, and testing whether results are stable across alternative feature sets. Poor featurisation is the most common reason segmentation and prediction projects fail.
Attributes
- Discipline
- Product Analytics & Strategy
- Loss type
- Latent
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Weeks
Steps
- Define the analysis unit, observation window (e.g. 90 days) and population with clear eligibility rules.
- Draft feature families from qualitative insight: frequency, recency, depth, breadth, sequence and timing measures.
- Compute features with documented SQL or code and a data dictionary specifying event definitions and transformations.
- Transform skewed counts (log1p), convert totals to rates where exposure differs, and standardise.
- Remove near-zero-variance features and check pairwise correlations (consider dropping one of pairs above 0.9).
- Check for leakage of outcome information and test downstream results with an alternative feature set for robustness.
Sources
- Pattern Recognition and Machine Learning (Christopher Bishop)
- Unsupervised Learning in Space and Time (Marc Leordeanu)
- Srivastava et al. (2000) Web usage mining
- Feature Engineering and Selection (Max Kuhn & Kjell Johnson)
- Pattern Recognition & ML (Bishop)
- Unsupervised Learning in Time & Space
- The Inquiry Series, Ch 52 — Dimensionality Reduction — The Geometry of Behavioral Space
Also known as
- Behavioural feature engineering
- User-by-feature matrix
- Event log featurisation
Categories
- Analytics & Behaviour
- Product Analytics & Strategy