Abstract
Graph-valued outcomes are increasingly common in a variety of scientific observations, ranging from biomedical domains such connectomes, bioinformatics networks such as protein interactions to other disparate areas such as communications and social networks . Networks are estimated by a variety of statistical and machine learning techniques such as stochastic block models and graph-learning algorithms; however, evaluation of the parameters and predictions form such learning techniques often target an external class label or require a tractable graph likelihood. We propose a topology-aware generalization of Brier-type scoring rules for direct evaluation of any graph-valued predictions through a prespecified vector of permutation-invariant graph functionals. The score is likelihood-free, strictly proper for the conditional mean of the selected topology functional, and admits an exact excess-risk identity. With block-diagonal weights, it also decomposes additively into scientifically interpretable components, such as density, degree heterogeneity, clustering, connectivity, and spectral structure. We distinguish this predictive target from likelihood-ratio testing, label classification, adjusted Rand index, area under the receiver operating characteristic curve, and class-label logarithmic loss. A matched-density simulation comparing Erdos–Rényi and random-hub networks illustrates a setting in which density-based evaluation is intentionally uninformative while the topology differs. A second empirical evaluation uses the real dataset PROTEINS benchmark of protein graphs. The proposed score is intended for out-of-sample comparison of graph forecasts, particularly when likelihoods are unavailable, misspecified, or scientifically less interpretable than selected topological features.

This publication has 18 references indexed in Scilit: