Scholarlink

<p>We present a phase-aware deep clustering pipeline for discovering interpretable football team playing styles from event data. Match events are organized into four possession-driven phases (In-Possession, Out-of-Possession, Positive Transition, and Negative Transition), and each phase is modeled with Deep Embedded Clustering (DEC) to learn cluster-oriented latent representations from spatiotemporal event-derived features. We extend phase-specific style discovery with four complementary analyses. First, we incorporate fixed 15-min match windows to study within-match style switching and tactical adaptation. Second, we test temporal robustness via frozen-model inference by transferring DEC models trained on first-half matches to second-half data without retraining. Third, we improve interpretability using feature-group ablation, supervised surrogate modeling, and Shapley Additive Explanations (SHAP) to identify global and local drivers of cluster assignments. Finally, we quantify inter-phase tactical coherence and combine phase-level styles into holistic archetypes that capture joint attacking and defensive identities. We benchmark clustering quality against classical baselines and assess practical relevance through outcome-based analyses, including style matchups and league-wise distributions. Source code is available at: <a href="https://github.com/egecjdemir/how_football_teams_play">https://github.com/egecjdemir/how_football_teams_play</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

How Do Football Teams Play? An Extended DEC Analysis to Uncover Playing Styles

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Ege Demir,
  • Nazım Kemal Üre,
  • Yusuf H. Şahin

摘要

We present a phase-aware deep clustering pipeline for discovering interpretable football team playing styles from event data. Match events are organized into four possession-driven phases (In-Possession, Out-of-Possession, Positive Transition, and Negative Transition), and each phase is modeled with Deep Embedded Clustering (DEC) to learn cluster-oriented latent representations from spatiotemporal event-derived features. We extend phase-specific style discovery with four complementary analyses. First, we incorporate fixed 15-min match windows to study within-match style switching and tactical adaptation. Second, we test temporal robustness via frozen-model inference by transferring DEC models trained on first-half matches to second-half data without retraining. Third, we improve interpretability using feature-group ablation, supervised surrogate modeling, and Shapley Additive Explanations (SHAP) to identify global and local drivers of cluster assignments. Finally, we quantify inter-phase tactical coherence and combine phase-level styles into holistic archetypes that capture joint attacking and defensive identities. We benchmark clustering quality against classical baselines and assess practical relevance through outcome-based analyses, including style matchups and league-wise distributions. Source code is available at: https://github.com/egecjdemir/how_football_teams_play.