Role Clustering · Analysis

Role Clustering · Evaluation

Cluster players by playing style and functional behavior: metrics, separability (UMAP), and relationships to traditional positions (charts from assets/img/role_cluster/ensemble/).

3 role clusters

Functional style grouping

Acc ≈ 0.885

Accuracy / F1

Interpretable

Distributions + heatmaps

↘ Core Metrics Feature Space Positions

Back to Landing Rating Analysis Official example

UMAP by label — UMAP · Cluster separation

Label position heatmap — Heatmap · Role vs position

Core Classification Metrics

Cross-Validation Performance

Role metrics — Accuracy / F1 is stable around ~0.885

Balanced metrics suggest the model is not dominated by majority classes.

Confusion Matrix

Well-suited for product UX: role labels with clear explanations.

Class Balance

Actual vs predicted counts — Predicted counts match the true distribution

Avoids collapsing into a single class, enabling search and recommendations.

Behavior & Feature Space

Cluster Separation (UMAP)

UMAP embedding — Clear “islands” in the projected feature space

A visual confirmation that roles are separable in behavior-feature space.

Base Score Independence

Base score by label — Roles describe style, not just overall quality

Enables “style search”: similar quality with different styles, or similar styles at different quality levels.

Redefining Positions

Roles vs Traditional Positions

Label distribution — The same nominal position can map to different roles

Turns “position” from a static label into an explainable behavior-based grouping.

Position Heatmap

Useful for scouting filters (e.g., “balanced defensive midfield profiles”).

Note: Click any chart to zoom (Esc to close, ←/→ to navigate).

Athlemetrics · Role Analysis