segment-modeling

What problem does it solve?

This Skill helps you improve binary classification performance by finding an effective segmentation strategy and training separate sub-models for each segment instead of using one single model for all users.

Core Features & Use Cases

• Segmentation strategy exploration (rule / clustering / decision-tree): Supports rule-based segmentation (human priors), unsupervised clustering (e.g., K-Means), and supervised decision-tree segmentation to discover meaningful groups.
• Try → Measure → Keep/Discard → Repeat workflow: Runs multiple rounds to explore candidate segmentation hypotheses, evaluates them with AUC or KS, and keeps only better strategies while discarding non-improving or unstable ones.
• Sub-model training and aggregation: Trains an independent XGBoost model per segment and combines predictions using either routing (route) or stacking (stacking) style aggregation.
• Stability and coverage checks: Enforces minimum segment coverage and penalizes segment distribution drift via PSI to avoid fragile segmentation schemes.

Quick Start

Use the segment-modeling skill to run autonomous customer segmentation modeling on your dataset by selecting target y_label, using max_rounds 5, and outputting the results to ./outputs/seg.