Description
Shannon entropy of each user's distribution of actions or categories, used as a measure of behavioural variety whose decline signals drift even when visit frequency holds steady.
Shannon entropy H = −Σ p log p measures the spread of a distribution; for a user's actions it is high when they use many features evenly and low when they repeat a few. Vivek Gopinath's drift playbook uses interaction entropy over the action distribution per user as the lead drift indicator, alongside category diversity per session, recommendation CTR and attention decay, because drift reduces variety even when frequency is constant. Illustrative values: entropy falling from 2.4 to 1.1 (generic) or 2.6 to 1.2 (e-commerce, 10,000 users over 3 months). Thresholds feed Bayesian updating (entropy < 1.5 for three sessions) and decision bars (entropy > 2.0 for 70% of active users within 4 weeks). The forward-thinking playbook also uses action entropy rising from 1.1 to 1.9 as a long-term validation signal.
Attributes
- Discipline
- Information Theory, Physics & Mathematics
- Loss type
- Latent
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Days
Steps
- Define the action or category vocabulary from telemetry (user_id, session_id, ts, action, category).
- Compute each user's action distribution per period.
- Calculate H = −Σ p log p per user per period, normalising if comparing vocabularies of different size.
- Track entropy longitudinally alongside frequency to separate narrowing from disengagement.
- Set thresholds for intervention and decision bars (e.g. entropy < 1.5 triggers nudges; target > 2.0).
- Pair low-entropy users with interviews or diaries to understand why.
Sources
- Vivek Gopinath (2025), Detecting Behavioral Drift: A Mixed Methods Playbook and E-commerce Case Study
- Vivek Gopinath (2025), Forward-Thinking UX Research Playbook: A Mixed-Methods, Decision-Ready Approach from Prediction to Validation
Also known as
- Action entropy
- Behavioural diversity entropy
Categories
- Information Theory, Physics & Mathematics