PaperlayerView exact revision on arXiv

Source-linked AI summary

When Connected Does Not Mean Similar: Charting the Homophily Boundary of SNAP-KG for Streaming Entity Integration

Jui-Chien Lin, Oshani Seneviratne

arXiv:2609.12356v1cs.LG

TL;DR

SNAP-KG assigns newly arriving knowledge-graph entities to semantic communities using only raw features, but its prior evaluation used homophilous graphs. This paper tests the method on heterophilous graphs and measures edge homophily across datasets. Quality falls sharply when no homophilous view is available, and the paper frames heterophilous clustering as a separate research problem.

  • Problem

    SNAP-KG’s prior evaluation used homophilous graphs, leaving open how it performs when connected nodes tend to belong to different classes.

    Source: §1 · p. 1

  • Method

    The paper runs five-seed SNAP-KG clustering experiments on Texas, Wisconsin, and Chameleon, measures edge homophily per view, and compares single-view relations and transductive baselines.

    Source: §1 · pp. 1–2, §3 · pp. 2–3

  • Results

    When no homophilous view is available, clustering quality falls sharply for SNAP-KG and transductive baselines, and the single-view comparison links the drop to relation homophily rather than relation count.

    Source: §3 · pp. 3–4, §5 · p. 4

  • Takeaways & Limitations

    The paper treats heterophilous multi-view clustering as a separate research direction and identifies distilling a heterophily-aware teacher into SNAP-KG as an opportunity.

    Source: §4 · p. 4, §5 · p. 4

    SNAP-KG and the evaluated transductive baselines share a homophily limitation, with performance dropping on the three heterophilous graphs.

    Source: §3 · p. 4

Abstract

from arXiv · show

SNAP-KG is a framework for assigning newly arriving entities to semantic communities in a growing knowledge graph (KG) using only their raw features, with no graph access and no retraining at inference time. It was evaluated on five multi-view benchmarks and a 2.4M-node OGB-WikiKG2 KG. In each of these datasets, at least one graph view is homophilous, meaning that connected nodes usually belong to the same class, and SNAP-KG performs well on all of them. This paper asks what happens outside that setting. We extend the evaluation to three heterophilous graphs (Texas, Wisconsin, Chameleon) and measure the edge homophily of every view. When no homophilous view is available, clustering quality drops sharply for both SNAP-KG and the transductive baselines used in its original evaluation. What decides this is the homophily of the relation, not the number of relations. Multi-view fusion still helps, but only when at least one homophilous relation provides a reliable foundation. The homophily assumption is therefore shared by the whole method family, not specific to SNAP-KG. We argue that heterophilous multi-view clustering is a separate research problem, outside the scope of this work. As future work, we outline how a heterophily-aware teacher could be distilled into SNAP-KG's projector to serve both homophilous and heterophilous KGs.

1. Introduction

SNAP-KG places new entities into semantic communities from raw features, but its evaluation to date relied on homophilous graphs. This paper extends the evaluation to test that assumption and locate the method’s boundary.

1. Introduction

  • SNAP-KG’s prior evaluation covered five multi-view benchmarks and a 2.4M-node OGB-WikiKG2 knowledge graph, while relying on homophilous views.At inference, it assigns new entities from raw features without graph access or retraining.Source: §1 · p. 1
  • The motivating question is whether SNAP-KG’s performance holds when connected nodes tend to belong to different classes.All five benchmarks in its original evaluation satisfy the homophily assumption.Source: §1 · p. 1
  • The paper adds five-seed clustering experiments on Texas, Wisconsin, and Chameleon, plus per-view edge-homophily measurements across eight datasets.It also compares single-view relations to separate relation homophily from the number of views, examines transductive baselines, and outlines a research roadmap for heterophilous KGs.Source: §1 · pp. 1–2
  • During training, graph preprocessing, view-specific GNNs, and a multi-view encoder produce embeddings that a projector learns to reproduce from raw features.The projector is the only component used for streaming inference.Source: Figure 1 · p. 2

2. SNAP-KG and Its Homophily Assumption

SNAP-KG’s components rely on homophily at multiple stages, so heterophily can disrupt both the teacher embeddings and the distilled projector.

2. SNAP-KG and Its Homophily Assumption

  • Adaptive preprocessing keeps feature-similar neighbors, and the contrastive objective pulls those neighbors’ embeddings together.Under heterophily, useful neighbors may be dissimilar and kept neighbors may belong to different classes, so both steps can lose or mix structural signal.Source: §2 · p. 2
  • Because the projector is distilled from the resulting embeddings, it inherits any class mixing introduced by the teacher.The projector cannot separate nodes that the teacher has mixed together.Source: §2 · p. 2
  • The described limitation is not specific to SNAP-KG; homophily-assuming graph encoders and the transductive baselines show the same performance drop.Section 3 reports the corresponding baseline drop.Source: §2 · p. 2

3. A Scope Study on Heterophilous Graphs

The study measures edge homophily across datasets and tests SNAP-KG on three single-relation heterophilous graphs. Clustering quality is associated with having a homophilous view, and single-view comparisons indicate that relation homophily—not view count—tracks performance.

3. A Scope Study on Heterophilous Graphs

  • Edge homophily is the fraction of edges whose endpoints share a class label, computed per view on raw graphs before adaptive filtering.The three new graphs—Texas, Wisconsin, and Chameleon—are single-relation cases evaluated with five seeds; ACC standard deviation stays below 3.7 points.Source: §3 · pp. 2–3
  • Every dataset where SNAP-KG exceeds 90% ACC has a view with edge homophily of at least 0.64, while Texas, Wisconsin, and Chameleon have no view above 0.32 and achieve 36.61%, 35.14%, and 31.07% ACC, respectively.The reported pattern is an association, not proof of cause; NMI is in the single digits on the three heterophilous graphs.Source: §3 · p. 3, Table 1 · p. 3
  • With one view fixed, APVPA alone gives DBLP 90.07% ACC versus 91.82% for the full model, while APTPA alone gives 56.42%; YELP shows the same contrast for BSB and BUB.BSB alone gives YELP 90.05% ACC versus 91.20% for the full model, while BUB alone gives 37.18%.Source: §3 · p. 3, Table 2 · p. 3
  • Transductive baselines also fall from 82–93% ACC on ACM and DBLP to 32–59% on the heterophilous graphs; DuaLGR’s best is 58.57% ACC on Wisconsin.SNAP-KG has a similar loss, but is the only compared method that assigns streaming entities inductively without retraining.Source: §3 · p. 4, Table 3 · p. 4

4. Heterophilous Clustering Is a Different Problem

The paper frames heterophilous clustering as a distinct research direction, separate from streaming inductive deployment. It identifies a possible bridge: distill a heterophily-aware teacher into SNAP-KG’s projector.

4. Heterophilous Clustering Is a Different Problem

  • The paper characterizes heterophilous graph support as a research direction of its own, rather than a minor SNAP-KG modification.The cited methods address heterophily-aware representation, while SNAP-KG provides inductive, retraining-free deployment; the paper says no current method offers both.Source: §4 · p. 4
  • NGCE guides encoding with node features and combines homophilous and heterophilous patterns, HeNCler learns asymmetric similarity, and SMHGC mines similarity to strengthen homophily [15].The paper notes that, to the authors’ knowledge, these methods target transductive settings.Source: §4 · p. 4
  • A heterophily-aware multi-view teacher could be distilled into SNAP-KG’s projector because distillation needs only feature-embedding pairs, not a particular teacher architecture.This would retain the same deployment path while extending streaming entity integration to heterophilous KGs.Source: §4 · p. 4
  • Practitioners can estimate suitability by measuring each relation’s edge homophily with a small labeled sample, or by estimating feature similarity when labels are unavailable.The paper cautions that clusters should not be trusted when no relation is at least moderately homophilous.Source: §4 · p. 4

5. Conclusion

The conclusion reports that SNAP-KG’s quality falls sharply when no homophilous view is available, with a similar drop for transductive baselines. It identifies relation homophily rather than relation count as the distinguishing factor and proposes distillation of a heterophily-aware teacher as a future direction.

5. Conclusion

  • Across the evaluated graphs, SNAP-KG’s strong results accompany at least one homophilous view, while quality drops sharply without one for SNAP-KG and transductive baselines.The single-view comparison indicates that relation homophily, rather than the number of relations, tracks the drop.Source: §5 · p. 4
  • The paper distinguishes streaming inductive deployment, which SNAP-KG addresses, from heterophily-aware representation, addressed by a dedicated line of work [15].Source: §5 · p. 4
  • The proposed next direction is to distill a heterophily-aware teacher into SNAP-KG’s streaming projector to serve both settings.Source: §5 · p. 4

Declaration on Generative AI

The authors disclose using Claude, from Anthropic, for grammar correction and writing refinement only. They attribute the work’s ideas, methods, results, and conclusions to themselves and accept responsibility for its accuracy and integrity.

Declaration on Generative AI

Loading 2609.12356v1…