[Paper Review] Nonparametric Estimation in the Dynamic Bradley-Terry Model

Heejong Bong, Wan‐Shan Li+2 more|arXiv (Cornell University)|Feb 28, 2020

Game Theory and Voting Systems8 references4 citations

TL;DR

This paper proposes a nonparametric, time-varying extension of the Bradley-Terry model for dynamic ranking using kernel smoothing to handle sparse pairwise comparison data. It establishes existence and uniqueness conditions for the estimator and derives model-agnostic oracle bounds for estimation error and excess risk, demonstrating strong empirical performance across simulated and real-world data with efficient bandwidth selection.

ABSTRACT

We propose a time-varying generalization of the Bradley-Terry model that allows for nonparametric modeling of dynamic global rankings of distinct teams. We develop a novel estimator that relies on kernel smoothing to pre-process the pairwise comparisons over time and is applicable in sparse settings where the Bradley-Terry may not be fit. We obtain necessary and sufficient conditions for the existence and uniqueness of our estimator. We also derive time-varying oracle bounds for both the estimation error and the excess risk in the model-agnostic setting where the Bradley-Terry model is not necessarily the true data generating process. We thoroughly test the practical effectiveness of our model using both simulated and real world data and suggest an efficient data-driven approach for bandwidth tuning.

Motivation & Objective

• To develop a nonparametric, time-varying generalization of the Bradley-Terry model for dynamic global team rankings.
• To address the challenge of sparse pairwise comparison data where traditional Bradley-Terry estimation may fail due to non-existence of maximum likelihood estimates.
• To provide theoretical guarantees—existence, uniqueness, and oracle bounds—under a model-agnostic setting where the true data-generating process may not follow the Bradley-Terry model.
• To design a computationally efficient estimator that leverages kernel smoothing to pre-process time-dependent pairwise outcomes.

Proposed method

• The method uses kernel smoothing to estimate time-varying log-odds parameters βi(t) in the Bradley-Terry model, ensuring smoothness over time.
• It models the time-varying winning probability as logit(p_ij(t)) = β_i(t) - β_j(t), with ∑β_i(t) = 0 for identifiability.
• The estimator is constructed by smoothing the empirical pairwise comparison frequencies using a kernel function over time, enabling estimation even in sparse regimes.
• Existence and uniqueness of the estimator are established under mild regularity conditions on the kernel and smoothness of β(t).
• Oracle bounds for estimation error and excess risk are derived in a model-agnostic framework, without assuming the Bradley-Terry model is the true data-generating process.
• A data-driven bandwidth selection method is proposed using cross-validation, with empirical evidence showing good performance even with pre-selected bandwidths.

Experimental results

Research questions

• RQ1 Under what conditions does the proposed kernel-smoothed estimator for the time-varying Bradley-Terry model exist and remain unique?
• RQ2 How does the proposed method perform in terms of estimation error and excess risk when the true data-generating process is not a Bradley-Terry model?
• RQ3 Can the kernel smoothing approach reliably recover dynamic rankings in sparse pairwise comparison settings where standard MLE fails?
• RQ4 What is the computational efficiency of the proposed method compared to traditional parametric or state-space approaches?
• RQ5 How effective is the data-driven bandwidth selection procedure in practice, especially under high-dimensional or sparse data regimes?

Key findings

• The proposed estimator exists and is unique under mild conditions on the kernel and smoothness of the time-varying parameters β(t).
• The method achieves strong finite-sample performance in both simulated and real-world data, particularly excelling in sparse settings where standard MLE fails.
• Kernel smoothing ensures that the necessary condition for MLE existence (global connectivity of comparison graphs) is almost surely satisfied, even when raw data are sparse.
• The model achieves comparable or better runtime than the original Bradley-Terry model in high-time-point regimes (M large, N small), especially when bandwidth is pre-selected.
• Empirical results show that even with a pre-determined bandwidth in a reasonable range, the model yields estimates close to those obtained via leave-one-out cross-validation.
• In experiments with N=10, M=10, and n_ij(t)=1, the frequency of MLE existence across all time points was only 2%, but rose to 100% with kernel smoothing, demonstrating its robustness to sparsity.