Network analysis and link prediction in competitive women's basketball

TL;DR

This paper explores how network analysis and embedding techniques can predict future interactions and rank changes in women's basketball at NCAA and WNBA levels, revealing structural signals that influence game outcomes.

Contribution

It introduces a novel approach combining network metrics and node embeddings to predict game outcomes and player interactions in women's basketball.

Findings

Embedding-based models significantly predict future game interactions.

Structural similarity measures relate to team ranking changes.

Passing region embeddings show interpretable patterns consistent with passing feasibility.

Abstract

Network structure and its role in prediction are examined in competitive basketball at the team and player levels. Adversarial game outcome networks from NCAA Division I women's basketball from 2021 to 2024 are used to compute the common out-neighbor score and PageRank, which are combined into a low-key leader strength that identifies competitors influential through structural similarity despite relatively low centrality. This measure is related to changes in NCAA NET rankings by grouping teams into quantiles and comparing average rank changes across seasons for both previous-to-current and current-to-next transitions. Link prediction is then studied using node2vec embeddings across three interaction settings. For NCAA regular-season game networks, cosine similarity between team embeddings is used in a logistic regression model to predict March Madness matchups. For WNBA shot-blocking networks, future directed blocking interactions are predicted via logistic regression on concatenated source-target player embeddings. For WNBA passing networks, region embeddings learned from first-quarter passes are evaluated for their ability to predict subsequent passing connections. Across NCAA and WNBA settings, embedding-based models provide statistically significant evidence that higher-order network structure contains predictive signals for future interactions, while the passing experiment shows weaker predictive performance but yields interpretable similarity patterns consistent with passing feasibility.