<?xml version="1.0" encoding="UTF-8"?>
<ArticleSet>
  <Article>
    <Journal>
      <PublisherName></PublisherName>
      <JournalTitle>Journal of Artificial Intelligence, Applications and Innovations</JournalTitle>
      <Issn>3060-7124</Issn>
      <Volume></Volume>
      <Issue>In Press</Issue>
      <PubDate PubStatus="epublish">
        <Year>2026</Year>
        <Month>11</Month>
        <Day>19</Day>
      </PubDate>
    </Journal>
    <ArticleTitle>Reproducing Confidence-Based Subgraph Matching for Heterophilic Graph Neural Networks: An Independent Reconstruction and Reproducibility Audit</ArticleTitle>
    <VernacularTitle>Reproducing Confidence-Based Subgraph Matching for Heterophilic Graph Neural Networks: An Independent Reconstruction and Reproducibility Audit</VernacularTitle>
    <FirstPage>1</FirstPage>
    <LastPage>19</LastPage>
    <Language>EN</Language>
    <AuthorList>
      <Author>
        <FirstName></FirstName>
        <LastName></LastName>
        <Affiliation></Affiliation>
      </Author>
    </AuthorList>
    <PublicationType>Journal Article</PublicationType>
    <Abstract>&lt;p&gt;Confidence-based Subgraph Matching (ConSM) is a two-stage framework for graph neural networks that estimates task-relevant edge agreement from local subgraph structure and then uses the resulting edge scores in a sign-aware propagation objective. This study provides a critical analysis and independent reconstruction of ConSM, with particular emphasis on reproducibility under heterophily. We consolidate the method's mathematical formulation, clarify its graph-theoretic and spectral motivation, and reframe several stronger theoretical statements as proof sketches or informal propositions with their assumptions made explicit. The independent CPU-oriented reconstruction approximates the optimal-transport/Monge stage by nearest-reference assignment and reference-wise pooling. Across three recorded seeds (n=3), the reconstruction obtained 79.4 ± 1.4% test accuracy on Cora and 69.2 ± 2.5% on Citeseer. Raw per-seed values were not retained in the active reproducibility record, so they are not reconstructed or imputed here; the reported mean±SD therefore represents the complete seed-level evidence currently available. Recorded single-run results on Actor, Chameleon, and Squirrel were below the corresponding GCN baselines, consistent with possible sensitivity to transport quality, data versions, confidence calibration, optimization settings, and random splits. We retain these negative results and analyze the Chameleon failure mode rather than treating it as an implementation detail to be hidden. Results reported in the verified source literature on edge classification, depth robustness, link prediction, label scarcity, and leakage-free evaluation are summarized separately from independently generated evidence. Finally, we outline local confidence gates, unbalanced transport, probability calibration, and node-wise filter mixtures as future directions rather than validated improvements. The manuscript is explicitly scoped as a reproducibility audit: it establishes what can be supported by the present evidence and provides a limitations-based validation roadmap for follow-up studies rather than presenting unavailable experiments as completed evidence.&lt;/p&gt;
&lt;p&gt; &lt;/p&gt;</Abstract>
    <ArchiveCopySource DocType="pdf"></ArchiveCopySource>
  </Article>
</ArticleSet>
