DriftScore (Behavioral Drift Score)

Description

A 0–1 composite of normalised entropy, CTR, diversity and attention-decay signals, weighted by regression, where higher values mean greater risk that a user is drifting.

The DriftScore combines normalised telemetry signals into a single risk index: DriftScore = w1(1 − entropy_norm) + w2(1 − ctr_norm) + w3(1 − diversity_norm) + w4(attention_decay). Signals are normalised to [0,1] with higher meaning more engagement, then inverted where needed so higher scores mean more drift risk. Weights are set by regression or optimisation. In the e-commerce case with weights [0.4, 0.3, 0.3], P1 (entropy 1.0, CTR 0.18, 2 categories) scores 0.72 and P3 (0.9, 0.12, 1 category) 0.81, both drifting, while P2 (2.2, 0.44, 5 categories) scores 0.38. An action threshold is suggested, such as Score ≥ 0.50 for 2 consecutive weeks. The score's rigour comes from transparent components that map back to Honeycomb dimensions, and it is the objective that adaptive interventions and Thompson Sampling try to reduce.

Attributes

Loss type
Community-proposed
Stage
Post-launch
Effort
Weeks

Steps

  1. Select drift signals from telemetry: interaction entropy, recommendation or promo CTR, category or feature diversity, and attention decay (scroll depth, dwell).
  2. Normalise each signal to [0,1] with higher meaning more engagement, then convert to risk (1 − entropy_norm, 1 − ctr_norm, 1 − diversity_norm, attention_decay).
  3. Estimate weights via regression or optimisation against a later outcome such as churn or disengagement (examples use [0.3, 0.25, 0.2, 0.25] or [0.4, 0.3, 0.3]).
  4. Compute the weighted sum per user per period.
  5. Apply an action threshold, e.g. intervene if Score ≥ 0.50 for 2 consecutive weeks.
  6. Use the score as the objective for adaptive interventions and A/B validation, and pair high scorers with interviews or diaries to confirm the mechanism.

Sources

  • Vivek Gopinath (2025), Detecting Behavioral Drift: A Mixed Methods Playbook and E-commerce Case Study
  • Causal Inference: What If (Miguel Hernán & James Robins)
  • Forward-Thinking Playbook (Vivek Gopinath)
  • Behavioral Drift Playbook (Vivek Gopinath)
  • Causal Inference (Hernán & Robins)
  • Forward-Thinking Playbook (Gopinath)
  • Behavioral Drift Playbook (Gopinath)
  • The Inquiry Series, Ch 44 — Survival Analysis, Longitudinal Design & Behavioral Drift

Also known as

  • Drift Score
  • 0–1 drift risk index
  • Behavioural drift score
  • Behavioral Drift Index
  • Drift score construction (0-1)

Categories

  • Community-Proposed Frameworks