Methodology¶
Overview¶
priorityx uses Generalized Linear Mixed Models (GLMM) to classify entities into priority quadrants based on volume and growth trajectories.
Statistical Approach¶
GLMM Specification¶
Model: Poisson Bayesian Mixed GLMM
Components:
- Fixed effects: Overall time trend + seasonal dummies (quarterly/semiannual)
- Random intercepts: Entity-specific baseline volume
- Random slopes: Entity-specific growth rates
Estimation: Variational Bayes (VB) for posterior mean
Priors:
- Random effects: vcp_p = 3.5 (relaxed for boundary behavior)
- Fixed effects: fe_p = 3.0
Quadrant Classification¶
Entities classified based on random effects:
Q1 (Critical): intercept > 0, slope > 0, count ≥ 50 - High volume, accelerating growth - Requires immediate attention
Q2 (Investigate): intercept ≤ 0, slope > 0 - Low volume but growing rapidly - Emerging issues to watch
Q3 (Monitor): intercept ≤ 0, slope ≤ 0 - Low volume, stable or declining - Routine monitoring
Q4 (Low Priority): intercept > 0, slope ≤ 0 - High volume but not accelerating - Persistent baseline issues
The count threshold for Q1 prevents low-volume entities from being mislabeled as Critical.
Priority tiers applied in the transition timeline build on these quadrants with velocity-based rules (see docs/priority_classification.md) to differentiate Crisis, Investigate, Monitor, and Low responses.
Movement Tracking¶
Three-Step Process¶
1. Global Baseline - GLMM on full dataset - Provides stable quadrant assignment
2. Endpoint Cohorting - Define valid entities at analysis endpoint - Ensures consistent peer group
3. Quarterly Tracking - GLMM on cumulative data up to each quarter - Tracks X/Y position changes over time
Transition Detection¶
Cross-quadrant transitions:
- Q3→Q2→Q1: Escalation path
- Q1→Q4, Q2→Q3: De-escalation
Within-quadrant changes:
- Y-axis surge > 1.0: Dramatic acceleration
- X-axis surge > 1.0: Major volume increase
Data Filters¶
Sparse entities:
- min_total_count: Filter entities below count threshold
- min_observations: Filter entities with insufficient time periods
Stale entities:
- decline_window_quarters: Filter entities inactive >N quarters
- Prevents contamination from historical data
Validation¶
Approach validated on regulatory monitoring data:
- 95.7% accuracy vs baseline methods
- 1-3 quarter earlier detection of escalating entities
- Reduced false oscillations for smooth growth patterns
References¶
- Social media analytics for mining customer complaints to explore product opportunities (2023). Computers & Industrial Engineering. https://doi.org/10.1016/j.cie.2023.109104