A1480
Title: Optimal inference on two-sample network effects for directed networks
Authors: Wen Zhou - New York University (United States) [presenting]
Yuan Zhang - The Ohio State University (United States)
Mladen Kolar - University of Chicago (United States)
Yating Liu - University of Chicago (United States)
Abstract: Directed networks model a wide range of asymmetric relationships, including social media connections, communication structures, and economic trade flows. These networks are often sparse and exhibit complex edge dependencies such as reciprocity, same-sender, same-receiver, and sender-receiver effects, each capturing key structural patterns like mutual cooperation, broadcasting behavior, recipient diversity, or indirect influence. While prior work has focused on estimating these effects within a single (typically dense) network, understanding how they change across two networks remains critical---particularly for studies of evolving network dynamics or policy interventions. We propose an adaptively unified motif-based subsampling framework for two-sample comparison in directed networks in testing for differences in each of the four edge dependencies. Our approach addresses several core challenges in two-sample network inference, including: (i) differing network sizes and node sets, (ii) varying sparsity levels, (iii) reliance on strong distributional assumptions, and most importantly (iv) the critical issue of unbalanced or indeterminate degeneracy. We derive a Berry-Esseen bound for our test statistics and establish near-optimal performance under various degeneracy scenarios.
