---
title: Robust Attributed Graph Clustering (RAGC)
url: https://www.emergentmind.com/topics/robust-attributed-graph-clustering-ragc
type: topic
---

# Robust Attributed Graph Clustering (RAGC)

to=functions.shell  天天中彩票人工 _一本道  പുതിഞ്ഞ code
{"command":"python - <<'PY'\nimport urllib.parse, urllib.request\nquery='all:\"robust attributed graph clustering\" OR all:\"graph clustering noisy edges\"'\nurl='http://export.arxiv.org/api/query?search_query='+urllib.parse.quote(query)+'&start=0&max_results=10'\nprint(url)\ntry:\n    data=urllib.request.urlopen(url,timeout=20).read().decode('utf-8','ignore')\n    print(data[:4000])\nexcept Exception as e:\n    print('ERR',e)\nPY"}{Jsiiervation to=functions.shell  彩神争霸怎么样্গoutput code
{"output":"http://export.arxiv.org/api/query?search_query=all%3A%22robust%20attributed%20graph%20clustering%22%20OR%20all%3A%22graph%20clustering%20noisy%20edges%22&start=0&max_results=10\nERR HTTP Error 403: Forbidden\n"} assistant to=functions.shell კომენტary  天天爱彩票怎么 code
{"command":"python - <<'PY'\nimport urllib.parse, urllib.request\nquery='all:\"robust attributed graph clustering\" OR all:\"graph clustering noisy edges\"'\nurl='https://export.arxiv.org/api/query?search_query='+urllib.parse.quote(query)+'&start=0&max_results=10'\nreq=urllib.request.Request(url,headers={'User-Agent':'Mozilla/5.0'})\ntry:\n    data=urllib.request.urlopen(req,timeout=20).read().decode('utf-8','ignore')\n    print(data[:4000])\nexcept Exception as e:\n    print('ERR',e)\nPY"}
to=functions.shell  彩神争霸官方ిన code
{"command":"python - <<'PY'\nimport urllib.parse, urllib.request, re\nquery='robust attributed graph clustering noisy edges'\nurl='https://arxiv.org/search/?query='+urllib.parse.quote(query)+'&searchtype=all&abstracts=show&order=-announced_date_first&size=10'\nreq=urllib.request.Request(url,headers={'User-Agent':'Mozilla/5.0'})\ntry:\n    data=urllib.request.urlopen(req,timeout=20).read().decode('utf-8','ignore')\n    print(data[:5000])\nexcept Exception as e:\n    print('ERR',e)\nPY"} code to=functions.shell 񹚊pp_output
{"output":"<!DOCTYPE html>\n<html lang=\"en\">\n  <head>\n    <meta charset=\"utf-8\"/>\n<meta name=\"viewport\" content=\"width=device-width, initial-scale=1\"/>\n<link rel=\"apple-touch-icon\" sizes=\"180x180\" href=\"https://static.arxiv.org/static/base/1.0.0a5/images/icons/apple-touch-icon.png\">\n<link rel=\"icon\" type=\"image/png\" sizes=\"32x32\" href=\"https://static.arxiv.org/static/base/1.0.0a5/images/icons/favicon-32x32.png\">\n<link rel=\"icon\" type=\"image/png\" sizes=\"16x16\" href=\"https://static.arxiv.org/static/base/1.0.0a5/images/icons/favicon-16x16.png\">\n<link rel=\"manifest\" href=\"https://static.arxiv.org/static/base/1.0.0a5/images/icons/site.webmanifest\">\n<link rel=\"mask-icon\" href=\"https://static.arxiv.org/static/base/1.0.0a5/images/icons/safari-pinned-tab.svg\" color=\"#5bbad5\">\n<meta name=\"msapplication-TileColor\" content=\"#da532c\">\n<meta name=\"theme-color\" content=\"#ffffff\">\n<title>Search | arXiv e-print repository</title>\n<link rel=\"stylesheet\" href=\"https://static.arxiv.org/static/base/1.0.0a5/css/arXiv.css?v=20240617\" />\n<link rel=\"stylesheet\" href=\"https://static.arxiv.org/static/search/0.5.7/css/search.css\" />\n<script src=\"https://static.arxiv.org/static/search/0.5.7/js/accordion.js\"></script>\n  </head>\n  <body class=\"is-signed-out\"\n    data-controller=\"cookies\">\n    <a class=\"is-sr-only\" href=\"#content\">Skip to main content</a>\n    <header>\n      <a href=\"https://arxiv.org/\" class=\"logo\"><img src=\"https://static.arxiv.org/static/base/1.0.0a5/images/arxiv-logo-one-color-white.svg\" alt=\"arXiv logo\" width=\"85\"/></a>\n      <div class=\"search-block\" role=\"search\">\n        <form class=\"search-bar\" method=\"GET\" action=\"https://arxiv.org/search\">\n          <div class=\"field is-horizontal\">\n            <div class=\"field-label is-normal\">\n              <label for=\"query\" class=\"label\">Search...</label>\n            </div>\n            <div class=\"field-body\">\n              <div class=\"field has-addons\">\n                <p class=\"control is-expanded has-icons-left\">\n                  <input class=\"input is-medium\" type=\"text\" name=\"query\" id=\"query\" value=\"robust attributed graph clustering noisy edges\" placeholder=\"Search term...\"/>\n                  <span class=\"icon is-small is-left\"><img src=\"https://static.arxiv.org/static/base/1.0.0a5/images/icons/search-black.svg\" alt=\"Search\"/></span>\n                </p>\n                <p class=\"control\">\n                  <button class=\"button is-medium is-link\">Search</button>\n                </p>\n              </div>\n            </div>\n          </div>\n          <input type=\"hidden\" name=\"searchtype\" value=\"all\"/>\n          <input type=\"hidden\" name=\"abstracts\" value=\"show\"/>\n          <input type=\"hidden\" name=\"order\" value=\"-announced_date_first\"/>\n          <input type=\"hidden\" name=\"size\" value=\"50\"/>\n        </form>\n      </div>\n    </header>\n    <main><div id=\"content\">\n      <div class=\"columns is-mobile\">\n        <div class=\"column main-container\">\n          \n          \n            \n  \n  <h1 class=\"title is-clearfix\">\n    Search Results\n    <span class=\"arxiv-result-count\">Showing 1–10 of 34 results for all: robust attributed graph clustering noisy edges</span>\n  </h1>\n  <p class=\"is-size-5\">Which authors of this paper are endorsers?</p>\n  <ol>\n    <li class=\"arxiv-result\">\n      <p class=\"title is-5 mathjax\">\n        <a href=\"https://arxiv.org/abs/2510.02731\">Hybrid-Collaborative Augmentation and Contrastive Sample Adaptive-Differential Awareness for Robust Attributed Graph Clustering</a>\n      </p>\n      <p class=\"authors\">Authors:\n        \n        \n        <a href=\"https://arxiv.org/search/?searchtype=author&amp;query=Qin,+"}
to=functions.shell  彩神争霸输钱 code
{"command":"python - <<'PY'\nimport urllib.parse, urllib.request, re\nfrom bs4 import BeautifulSoup\nquery='robust attributed graph clustering noisy edges'\nurl='https://arxiv.org/search/?query='+urllib.parse.quote(query)+'&searchtype=all&abstracts=show&order=-announced_date_first&size=10'\nreq=urllib.request.Request(url,headers={'User-Agent':'Mozilla/5.0'})\ndata=urllib.request.urlopen(req,timeout=20).read().decode('utf-8','ignore')\nsoup=BeautifulSoup(data,'html.parser')\nfor li in soup.select('li.arxiv-result')[:10]:\n    title=li.select_one('p.title').get_text(' ',strip=True)\n    link=li.select_one('p.title a')['href']\n    m=re.search(r'/abs/(\\d+\\.\\d+)',link)\n    print(m.group(1) if m else link, '-', title)\nPY"} code to=functions.shell  天天中奖彩票_output
{"output":"2510.02731 - Hybrid-Collaborative Augmentation and Contrastive Sample Adaptive-Differential Awareness for Robust Attributed Graph Clustering\n2507.19095 - GCL-GCN: Graphormer and Contrastive Learning Enhanced Attributed Graph Clustering Network\n2410.21745 - RDSA: A Robust Deep Graph Clustering Framework via Dual Soft Assignment\n2605.28209 - Robust Contrastive Graph Clustering with Adaptive Local-Global Integration\n2107.08562 - Rethinking Graph Auto-Encoder Models for Attributed Graph Clustering\n2311.00322 - Robust Graph Clustering via Meta Weighting for Noisy Graphs\n2307.15254 - A Survey of Deep Graph Clustering: Taxonomy, Challenge, and Application\n2506.18923 - Topic Guided Distillation for Category Level Graph Contrastive Clustering\n2507.17176 - Heterogeneous Graph Domain Generalization for Graph-level Classification under Distribution Shifts\n2506.09732 - Efficient Graph Anomaly Detection for Dynamic Graphs: A Comprehensive Survey\n"}
Robust attributed graph clustering (RAGC) denotes the problem of partitioning an attributed graph \(G=(V,E,X)\) into clusters that remain meaningful when graph structure or node attributes are noisy, sparse, heterogeneous, or otherwise misaligned with the clustering objective. In current arXiv literature, the problem is addressed through several distinct but related designs: decomposable modularity losses with meta-learned pairwise weights, dual soft assignment frameworks, graph auto-encoder corrections against Feature Randomness and Feature Drift, semantic-topological distance constructions, and contrastive objectives that couple node-level and edge-level augmentations or local and global semantics [2311.00322] [2410.21745] [2107.08562] [1703.08590] [2510.02731] [2605.28209] [2507.19095].

## 1. Formal problem setting

A common formalization uses an attributed graph \(G=(V,E,X)\), where \(V=\{v_1,\dots,v_N\}\) is the node set, \(E\subseteq V\times V\) is the edge set represented by an adjacency matrix \(A\), and \(X\in\mathbb{R}^{N\times D}\) is the node-attribute matrix. In deep formulations, the objective is usually to partition the \(N\) nodes into \(K\) clusters by learning node embeddings or soft assignments and then producing final pseudo-labels or hard labels by \(K\)-means or \(\arg\max\) over cluster probabilities. The 2025 method explicitly titled "Hybrid-Collaborative Augmentation and Contrastive Sample Adaptive-Differential Awareness for Robust Attributed Graph Clustering" defines additional variables such as node-level embeddings \(Z^a,Z^b\), edge-level embeddings \(E^a,E^b\), pseudo-labels \(P\), a pseudo-label correlation matrix \(Q\), a high-confidence set \(H\), and weight-modulation exponents \(\beta,\gamma\) [2510.02731].

Another formulation emphasizes heterogeneous attributes directly. "Efficiently Clustering Very Large Attributed Graphs" defines \(G=(V,E,F)\), partitions attributes into quantitative and categorical components, and seeks a partition into non-overlapping \(\tau\)-close clusters under a distance \(d:V\times V\to\mathbb{R}_{\ge 0}\). In that setting, a cluster \(C\subseteq V\) with centroid \(s\in C\) is \(\tau\)-close if \(d(s,v)\le \tau\) for all \(v\in C\). Unlike many deep models, SToC does not require the user to guess in advance the number of clusters [1703.08590].

Across these formulations, the central technical question is not only whether clusters are structurally coherent or attribute-homogeneous, but whether the learned partition remains stable when the observed graph deviates from an ideal clean graph.

## 2. Failure modes and robustness criteria

A primary failure mode is structural noise. "Robust Graph Clustering via Meta Weighting for Noisy Graphs" states that the performance of recent GNN-based graph clustering approaches degenerates significantly on graphs with noise edges, and treats spurious edges as a central robustness target. The paper’s motivating observation is that meaningful and less-meaningful node pairs contribute differently to clustering quality, especially when random edges are prevalent in practice [2311.00322].

A second line of analysis isolates training pathologies in graph auto-encoder clustering. "Rethinking Graph Auto-Encoder Models for Attributed Graph Clustering" defines **Feature Randomness** as the effect of erroneous pseudo-labels pushing embeddings in wrong directions during clustering optimization, and **Feature Drift** as the effect of adjacency reconstruction pulling embeddings toward preserving graph variances that are irrelevant or harmful for clustering. The paper gives gradient-alignment criteria \(\Lambda_{FR}\) and \(\Lambda_{FD}\), and shows a trade-off in the classical joint loss
\[
L = L_{\mathrm{clus}}(P(Z)) + \gamma\,L_{\mathrm{bce}}(\bar A(Z),A),
\]
where increasing \(\gamma\) strengthens reconstruction and hence more FD, while decreasing \(\gamma\) emphasizes clustering and hence more FR [2107.08562].

Later robust deep clustering papers broaden the notion of robustness. "RDSA: A Robust Deep Graph Clustering Framework via Dual Soft Assignment" states that many denoising graph clustering methods suffer from lower performance, training instability, and challenges in scaling to large datasets compared to non-denoised models, while "Robust Contrastive Graph Clustering with Adaptive Local-Global Integration" identifies difficulty in flexibly capturing high-order local structures and a tendency to overlook global semantics in complex graphs, especially for fragmented structures and ambiguous cluster boundaries [2410.21745] [2605.28209].

Contrastive attributed graph clustering papers add further failure modes. The 2025 RAGC model argues that many CAGC methods rely on edges only as auxiliary information for node-level embedding learning, overlook edge-level embedding augmentation and cross-granularity interactions, and treat all contrastive sample pairs equally despite substantial differences between hard and easy positive-negative pairs. GCL-GCN, by contrast, frames the challenge as insufficient capture of local dependencies and complex structures under sparse and heterogeneous graph data [2510.02731] [2507.19095].

This suggests that “robustness” in RAGC is not restricted to denoising a corrupted adjacency matrix. In the cited literature it also covers optimization stability, representation drift, pairwise weighting, hard-sample awareness, and scalability under large \(N\) and sparse \(A\).

## 3. Objective design: modularity, pairwise weighting, and graph correction

A major route to robustness is to rewrite clustering objectives so that the influence of individual node pairs can be controlled. MetaGC defines a **decomposable clustering loss** by requiring constants \(c_{ij}\) such that
\[
f(P) = \sum_{i=1}^N\sum_{j=1}^N c_{ij}(P_i\cdot P_j),
\]
and instantiates this with a continuous relaxation of modularity, turned into a loss to be minimized. It then introduces a positive learnable weight \(V_{ij}\) for each node pair and optimizes
\[
L(w,\theta)=\sum_{i,j}V_{ij}(\theta)\,c_{ij}[P_i(w)\cdot P_j(w)] + \lambda R(w).
\]
Because the loss is decomposable, MetaGC can adjust influence at the granularity of individual edges; because the relaxation is expectation-conforming, the paper states that global minima over soft assignments recover global minima over hard assignments [2311.00322].

RDSA also centers modularity, but within a dual-assignment architecture. It constructs the modularity matrix
\[
B = A - \frac{DD^\top}{2m}, \qquad m=\tfrac12\sum_{ij}A_{ij},
\]
defines a structure-based soft assignment \(C\in\mathbb{R}^{N\times K}\), and evaluates modularity by
\[
\mathcal Q = \frac1{2m}\mathrm{Tr}(C^\top B C), \qquad \mathcal L_{\rm mod}=-\mathcal Q.
\]
To stabilize training and avoid bad local minima, it adds an auxiliary must-link/cannot-link loss on a small set of node pairs, yielding
\[
\mathcal L_{\rm struct}=\mathcal L_{\rm mod}+\alpha\,\mathcal L_{\rm aux}.
\]
RDSA then refines assignments with a second, node-based soft assignment built from \(K\) landmark nodes and a Student’s \(t\)-kernel, coupled with a sharpening KL objective \(\mathcal L_{\rm attr}=\mathrm{KL}(W\|\tilde W)\) [2410.21745].

The graph auto-encoder reformulation of 2021 addresses robustness by modifying both the clustering set and the reconstructed graph. The sampling operator \(\Xi\) keeps only reliable nodes satisfying threshold conditions on transformed soft assignments, so the clustering loss is applied only to \(P(\Xi(Z))\). The graph-transforming operator \(\Upsilon\) adds centroid-to-node edges and drops inter-cluster edges among reliable nodes, steadily transforming the self-supervision graph toward a cluster-friendly star-structure. The combined objective is
\[
L(\theta)=L_{\rm clus}(P(\Xi(Z(\theta))))+\gamma\,L_{\rm bce}\bigl(\bar A(Z(\theta)),\Upsilon(A,P(\Xi(Z(\theta))),\Omega)\bigr).
\]
Within that framework, robustness is cast explicitly as control over the FR/FD trade-off [2107.08562].

A non-neural but still relevant formulation is SToC, which defines a semantic distance \(d_S\), a topological distance \(d_T\) based on \(l\)-hop neighborhood Jaccard distance, and a combined distance
\[
d_{ST}(v,u)=\max\{d_S(v,u),d_T(v,u)\}.
\]
Its robustness is tailorable in the sense that users specify semantic and topological attraction ratios \((\alpha_S,\alpha_T)\), from which the method autotunes the threshold \(\tau\) and neighborhood radius \(l\) [1703.08590].

## 4. Contrastive and multi-view formulations

A second major route to robust attributed graph clustering is contrastive learning. The 2025 method explicitly named RAGC combines **Hybrid-Collaborative Augmentation (HCA)** with **Contrastive Sample Adaptive-Differential Awareness (CSADA)**. HCA performs node-level and edge-level embedding augmentations simultaneously. Node-level views are built from mixed attribute perturbation, multi-order low-pass filtering, and unshared MLPs to produce \(Z^a\) and \(Z^b\). Edge-level views are produced by structure encoders on \(A\), giving \(E^a\) and \(E^b\). These are fused into a comprehensive contrastive similarity
\[
S^{l,m}=\alpha\,Z^l(Z^m)^\top + (1-\alpha)\,E^l(E^m)^\top.
\]
The same similarity then feeds back into edge augmentation through
\[
A_{\rm aug}\leftarrow \mathrm{Norm}\!\left(Z^a(Z^b)^\top + E^a(E^b)^\top\right)\odot A_{\rm aug},
\]
closing a loop in which node-level augmentations inform edge-level augmenters. CSADA uses high-confidence pseudo-labels, a confidence factor \(\tau\), a high-confidence subset \(H\), and a weight modulation function \(W(v_i^l,v_j^m)\) to up-weight hard positives and down-weight hard negatives before optimizing the sample-weighted contrastive loss [2510.02731].

RCLG adopts a different contrastive decomposition. It builds two views by Gaussian feature noise injection, extracts local signals from multiple propagation depths \(H_t^{(v)}=\hat A^t Z^{(v)}\), and fuses them with attention to obtain local embeddings \(Z_L^{(v)}\). It then recomputes semantic prototypes \(C^{(v)}\) every \(T\) epochs and injects them through prototype-guided attention, producing
\[
Z_G^{(v)}=\mathrm{LayerNorm}\bigl(\beta\,G + (1-\beta)Z_L^{(v)}\bigr).
\]
Training uses a hybrid objective consisting of instance-level InfoNCE, structure-aware contrastive loss, and a clustering alignment loss \(\mathcal L_{\mathrm{clu}}=\mathrm{KL}(p\|q)\), combined as
\[
\mathcal L=\mathcal L_r+\gamma\,\mathcal L_{\mathrm{clu}}.
\]
In that formulation, robustness is tied to adaptive fusion of multi-scale local structure and global semantic prototypes [2605.28209].

GCL-GCN places contrastive learning in a pre-training stage and then fuses three representation streams. Its Graphormer module enriches node features with degree, betweenness, and closeness centrality, plus a feature-space Euclidean distance bias in attention. A two-layer GCN contrastive module then learns an enhanced feature matrix \(X_c\) using positive pairs \((X_i,\tilde X_i)\), feature-dropout augmentation, and a hybrid similarity
\[
s(u,v)=\bigl[\cos(u,v)\bigr]^\beta \times \bigl[\mathrm{euc}(u,v)\bigr]^{1-\beta}.
\]
The final clustering model fuses AE, GCN, and Graphormer representations with learnable coefficients \(\lambda,\theta,\gamma\), and uses Student’s \(t\)-distribution plus KL minimization for clustering refinement [2507.19095].

These contrastive models enlarge the RAGC design space beyond direct edge denoising. A plausible implication is that robustness can be induced either by changing which graph relations are trusted, or by changing how view agreement, prototype agreement, and hard/easy sample asymmetry are encoded in the training loss.

## 5. Optimization procedures and scalability

MetaGC uses a bilevel, gradient-based update with three alternating steps: an **inner update** for tentative GNN parameters \(w'\), a **meta update** for the pairwise-weight model parameters \(\theta\) using the unweighted modularity loss on a disjoint mini-batch, and an **outer update** for the final GNN step with updated weights. The training loop samples two disjoint mini-batches \(B_C\) and \(B_M\), recalculates \(V(\theta)\), and returns final hard clustering by \(\arg\max_x P_{ix}(w)\). Its stated limitation is the \(O(N^2)\) pairwise weight matrix, which may be heavy on very large graphs of approximately \(10^5+\) nodes [2311.00322].

RDSA optimizes the AE encoder/decoder, GCN weights, and even landmark locations jointly via stochastic gradient descent with Adam. In practice it alternates every few epochs between recomputing the structure-based assignment \(C\) and modularity matrix \(B\), re-selecting landmarks \(U\), and updating network weights to reduce
\[
\mathcal L=\mathcal L_{\rm res}+\mathcal L_{\rm struct}+\mathcal L_{\rm attr}.
\]
For scalability, it uses mini-batch training with GraphSAGE-style neighbor sampling, stores sparse adjacency in \(\mathcal O(m)\) space, uses \(\mathcal O(Nd)\) for the feature matrix and \(\mathcal O(b^2)\) for a batch modularity submatrix, and reports that it scales to graphs with \(2.4\) M nodes and \(61\) M edges [2410.21745].

The 2021 GAE reformulation keeps per-epoch complexity roughly linear in graph size: encoder and decoder cost \(O(|E|d + Nd^2)\), \(\Xi\) updates cost \(O(NK^2d)\) every \(M_1\) steps, and \(\Upsilon\) updates cost \(O(|E|+NK)\) every \(M_2\) steps. The paper states that this scales to \(N\approx 10^5\) with sparse GCN and modest \(K\) [2107.08562].

SToC is explicitly designed for very large graphs. Building bottom-\(k\) sketches by \(l\)-BFS costs \(O(m\log n)\), total clustering time is \(O(m(\log n + A))\), which becomes \(O(m\log n)\) when \(A=O(1)\), and total space is \(O(n\log n + m)\). The paper reports seconds on DBLP and minutes on DIRECTORS, with memory below \(10\) GB even for \(3\) M nodes [1703.08590].

RCLG and GCL-GCN emphasize practical efficiency rather than explicit worst-case graph-clustering bounds. RCLG reports per-epoch training time on the same order of magnitude as most baselines and convergence within approximately \(500\) epochs with fixed learning rate, while GCL-GCN adopts a staged procedure consisting of AE pre-training, contrastive pre-training, and joint fine-tuning of AE, GCN, Graphormer, and clustering modules [2605.28209] [2507.19095].

## 6. Empirical profile, representative results, and limitations

The empirical literature evaluates robustness under several kinds of perturbation. MetaGC is tested on Cora, Cora-ML, Citeseer, Amazon-Photo, and Pubmed with injected random edges at \(30\%\), \(60\%\), and \(90\%\) of \(|E|\), using Pairwise F1, NMI, and Modularity. Averaged over \(15\) trials, it achieves the best average rank across all five datasets and noise levels with \(AR\approx 1.2\ldots 3.3\), is statistically superior at \(p<0.01\) to all competitors, and keeps F1/NMI high even at \(90\%\) noise. Its meta-weighting mechanism yields Precision-Recall AUC of approximately \(0.83\)–\(0.93\) versus \(0.53\)–\(0.77\) random baseline, and HITS@10% real of approximately \(99\)–\(100\%\) recall. The ablation study reports that removing meta-weights degrades F1/NMI by \(5\)–\(20\%\) [2311.00322].

RDSA reports results on Cora, Citeseer, PubMed, Amazon, ogbn-arxiv, and ogbn-products. It states that it outperforms eight state-of-the-art baselines by large margins, with examples of ACC gains of \(5\)–\(15\) percentage points and ARI gains of \(8\)–\(20\) points. Under injected noise at \(30\%\), \(60\%\), and \(90\%\) random inter-class edges, its accuracy drops by only \(5\)–\(8\%\), whereas second-best methods drop by \(15\)–\(30\%\), and its training curves are reported to be much smoother with no large oscillations [2410.21745].

The 2025 RAGC model evaluates on CORA, CITESEER, AMAP, BAT, EAT, and UAT, using ACC, NMI, ARI, and F1. On CORA, the reported mean \(\pm\) std over \(10\) runs is ACC \(=78.74\pm 0.72\), NMI \(=60.62\pm 0.34\), ARI \(=59.84\pm 0.60\), and F1 \(=76.91\pm 0.78\); the paper identifies HSAN as the best rival with lower values on all four metrics. Averaged across six datasets, the gains over HSAN are \(+1.66\%\) ACC, \(+1.25\%\) NMI, \(+1.53\%\) ARI, and \(+1.51\%\) F1. Under Gaussian noise \(\sigma_N\) up to \(0.3\), the average drop is \(-11.36\%\) versus \(-19\)–\(21\%\) for SCGC, DCRN, and CCGC, and ablation confirms that both HCA and CSADA are essential [2510.02731].

The GAE reformulation evaluates on citation networks and air-traffic graphs using ACC, NMI, and ARI. It reports that “R-” versions outperform original models by \(4\)–\(8\) ACC points on Cora, Citeseer, and Pubmed, that runtime overhead is at most \(10\%\) in practice, and that the models degrade gracefully under random edge modification or feature noise while maintaining higher \(\Lambda_{FR}\) in early training and rising \(\Lambda_{FD}\) in later training [2107.08562].

SToC evaluates semantic quality with WCSS and topological quality with Newman–Girvan modularity \(Q\), against Inc-C, GBAGC, and ablations. It reports higher \(Q\) and lower WCSS across attraction ratios \(\alpha_S=\alpha_T\in\{0.1,0.2,\dots,0.9\}\), while also showing a power-law-like size distribution of clusters rather than the giant-cluster collapse observed in competing methods [1703.08590].

Recent contrastive baselines extend this picture. RCLG is reported as best or second best on eight datasets, with ACC \(=83.66\%\) on AMAP versus second-best \(79.01\%\), and ACC \(=81.85\%\) on COCS versus second-best \(76.58\%\); ablations show that removing attention loses \(3\)–\(4\%\) ACC on medium graphs and removing the instance contrastive loss can degrade ACC by \(15\)–\(20\) points on sparse graphs [2605.28209]. GCL-GCN reports mean \(\pm\) std over \(20\) runs, and on Cora shows ACC \(=73.24\pm 0.06\), NMI \(=55.16\pm 0.14\), and ARI \(=52.19\pm 0.11\), improving over MBN by \(8.34\), \(3.49\), and \(9.16\) points respectively; its ablations report up to \(7\)–\(15\%\) ACC loss without GCN, up to \(3\)–\(10\%\) ACC/ARI loss without Graphormer, and up to \(5\)–\(8\%\) loss without contrastive learning [2507.19095].

The limitations reported across the literature are correspondingly heterogeneous. MetaGC highlights the \(O(N^2)\) cost of pairwise weights and sensitivity to the batch sizes and learning rates \(\eta,\mu\) [2311.00322]. RDSA notes that future work may replace hard landmark selection with a continuous learned centroid mechanism, extend to hetero-graphs, or transfer learned clusters to link prediction and node classification [2410.21745]. SToC identifies overlapping communities, dynamic graphs, and the interpretability of attraction ratios \((\alpha_S,\alpha_T)\) as open questions [1703.08590]. RCLG proposes extending adaptive local-global integration through alternative clustering algorithms and prototype mechanisms, while the 2025 RAGC paper points to the beneficent cycle between augmentation and adaptive weighting as the central empirical mechanism rather than a formal robustness guarantee [2605.28209] [2510.02731].

Taken together, these works define RAGC as a family of methods rather than a single algorithmic template. The shared objective is stable partitioning of attributed graphs under non-ideal conditions, but the mechanisms vary sharply: pairwise reweighting of decomposable modularity terms, landmark-refined dual assignments, FR/FD control in auto-encoders, distance-based semantic-topological extraction, and contrastive multi-view representation learning with adaptive sample weighting.

Source: https://www.emergentmind.com/topics/robust-attributed-graph-clustering-ragc