betaFeedback

WorksMathematics

Count Constraints and Conditional Benchmarks for Roster-Based Network Reconstruction

Abstract

When everyone reports contact counts for known groups, roster-based aggregated relational data (ARD) can determine individual links. In our sparse circular design with 40 nodes and eight groups, the complete counts fix 35.6% of realized edges on average. Zero counts alone exclude 91.1% of nonedges on average, and the full set of counts raises that share to 91.7%. We assess ranking on the unresolved pairs by the area under the receiver operating characteristic curve (AUC). Uniform weighting of compatible graphs yields higher mean AUC than the direct scores from either a circular latent-position fit or a nuclear-norm fit. In a follow-up analysis, using the same fitted probabilities to weight compatible graphs reverses the comparison for the circular fit. With the generating circular form and interaction strength supplied, its mean AUC exceeds uniform by 0.016-0.019 across three paired density settings. The gain is 0.007 under a stochastic blockmodel, where the circular form is misspecified. Nuclear-norm plug-in scores remain below uniform in mean AUC. For independent Bernoulli edges with fixed probabilities strictly between zero and one, we characterize when an oracle using the true probabilities gives the same edge marginals as uniform weighting for every feasible count array. This occurs exactly when the conditional graph law is uniform on every such array, or equivalently when log odds are additive in node-by-group effects. We recommend reporting fixed edge states, ranking on unresolved pairs, and generating-probability error separately.

Paper details

vlnc:W_u3mpxkda
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
arXiv nonexclusive license
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

OpenAI ChatGPT, Codex, Anthropic Claude, Google Deep Review

View statement
Looking for related papers…