Conduct a Controlled CAP-Net Player-Data Ablation

Determine whether restricting CAP-Net v2 training to players with at least 100 recorded balls resolves or mitigates the model’s apparent player-embedding cold-start and generalization problem.

Background

CAP-Net v2 uses learned player embeddings initialized from career statistics, together with a ball encoder and contextual feature branches. The paper attributes its weaker performance to a cold-start problem: many players in the dataset have limited recorded playing history, allowing the embedding table to memorize training-set patterns rather than learn representations that generalize well.

The proposed explanation has not been experimentally isolated. A controlled ablation that excludes players with fewer than 100 recorded balls would test whether the long tail of sparsely observed players is responsible for CAP-Net v2’s underperformance.

References

Substantiating this with a controlled ablation (training only on players with $ 100$ balls) is left for future work.

— A Fairness Audit of the Duckworth-Lewis-Stern Method: Format-Specific and Gender-Differential Bias, with an Interpretable Calibration Layer for Cricket Target Revision  (2609.04754 - Roy, 4 Sep 2026) in Section 6.3, “Why CAP-Net Underperformed”