Mutually-Antagonistic Interactions in Baseball Networks

We formulate the head-to-head matchups between Major League Baseball pitchers and batters from 1954 to 2008 as a bipartite network of mutually-antagonistic interactions. We consider both the full network and single-season networks, which exhibit interesting structural changes over time. We find interesting structure in the network and examine their sensitivity to baseball’s rule changes. We then study a biased random walk on the matchup networks as a simple and transparent way to compare the performance of players who competed under different conditions and to include information about which particular players a given player has faced. We find that a player’s position in the network does not correlate with his success in the random walker ranking but instead has a substantial effect on its sensitivity to changes in his own aggregate performance.

💡 Research Summary

This paper models the head‑to‑head matchups between Major League Baseball (MLB) pitchers and batters from 1954 through 2008 as a bipartite network of mutually‑antagonistic interactions. The authors construct two families of networks: (i) a “career” network that aggregates every pitcher‑batter encounter over the entire period, and (ii) a series of single‑season networks, one for each year. Three matrix representations are defined for each network: (1) a binary adjacency matrix A indicating whether a pitcher ever faced a batter, (2) a weighted matrix W counting the number of plate‑appearances for each pitcher‑batter pair, and (3) a performance matrix M that encodes the “Runs Expected per Unit of Effort” (RUE) for each plate appearance. RUE is a sabermetric statistic that assigns an expected run value to each possible outcome (single, home run, strikeout, etc.) independent of game context. The matrix M is signed: positive values indicate a batter’s advantage, negative values a pitcher’s advantage. By symmetrizing these matrices (ˆA, ˆW, ˆM) the authors obtain undirected representations suitable for standard network analysis.

Structural analysis reveals that the degree‑strength relationship follows a power law early in the dataset (strength ∝ degree^α with α≈1.6 for pitchers and 1.4 for batters in 1954) and gradually converges toward α≈1 by 2008. This trend reflects league expansions, divisional reorganizations, the introduction of inter‑league play, and other scheduling changes that make the interaction pattern more homogeneous over time. Nestedness, measured by the NODF metric, is modest (≈0.28) but consistently higher than in degree‑preserving randomizations, indicating a non‑trivial hierarchical organization. A notable structural shift occurs after 1973, when the American League (AL) adopts the Designated Hitter (DH) rule. The DH creates a set of high‑degree batters who never field, reducing the nestedness of the AL relative to the National League (NL) and producing two distinct nestedness trajectories for the two leagues.

To rank players, the authors introduce a biased random‑walker process on the bipartite graph. Each walker holds a “vote” for a player; at each step a random pitcher‑batter edge (i, j) is sampled, and the walker may switch its vote toward the player favored by the edge’s outcome. The transition matrix D is defined as

 D_ij = (ˆW_ij + r ˆM_ij) for i ≠ j,

 D_ii = –∑_{k≠i}(ˆW_ik + r ˆM_ik).

The scalar r is a bias parameter: r > 0 biases the walk toward the winner of the matchup (positive M_ij), while r < 0 biases toward the loser. The long‑time stationary distribution ˜v satisfies D ˜v = 0 with the normalization ∑_j ˜v_j = 1. For |r| < 0.7 the off‑diagonal entries remain non‑negative, guaranteeing, via the Perron‑Frobenius theorem, a unique positive eigenvector.

A linear expansion in r yields ˜v ≈ v(0) + r V, where the first‑order correction V solves a discrete Poisson equation on the graph:

 L V = 4 ∑_j ˆM_ij, subject to ∑_j V_j = 0,

with L = S – ˆW the graph Laplacian and S the diagonal strength matrix. The authors interpret V as an “RUE charge” that quantifies how the network topology modulates a player’s ranking beyond raw RUE.

Empirically, the random‑walker rankings are highly correlated with raw RUE (Pearson r ≈ 0.96 for 2008), confirming that the method captures overall performance. However, differences emerge when players with similar RUE have faced opponents of differing quality. For example, pitchers on offensively weak teams (e.g., 2008 Nationals, Astros, Reds) receive a boost in the random‑walker ranking because their opponents contributed fewer runs, effectively rewarding the difficulty of their matchups. Conversely, batters who have repeatedly faced elite pitchers may be ranked higher than raw RUE alone would suggest.

A key finding is that a player’s position in the network (degree, strength, nestedness contribution) does not predict his rank directly, but it strongly influences the sensitivity of his rank to changes in his own aggregate performance. High‑degree, high‑strength nodes exhibit larger rank fluctuations for a given change in RUE, indicating that extensive exposure to many opponents amplifies the effect of performance variations.

The paper concludes that while the biased random‑walker provides a transparent, parameter‑light ranking system that incorporates opponent quality, its primary contribution lies in revealing how structural features of a mutually‑antagonistic network shape ranking dynamics. The authors suggest that similar approaches could be applied to other sports, ecological predator‑prey systems, or any domain where direct competitive interactions can be represented as a bipartite graph.