acceptodds
Under review as a conference paper at ICLR 2027

Beyond Node Classification: Group Detection through Learned Graph Coarsening

Abstract

Money-laundering schemes, coordinated manipulation on social platforms, and fraud rings are inherently group-level: the object of interest is a coordinated set of nodes whose members may appear unremarkable in isolation. Yet graph-learning approaches often reduce such problems to node-level prediction. We address group detection directly, treating the group itself as the prediction unit. We formulate the task as learned graph coarsening: from training groups, we learn a target subspace under which unknown groups emerge as supernodes. We characterize the target-subspace properties required for coarsening to detect groups and translate them into a learning objective. We parameterize the target with a GNN, enabling sparse computation and inductive transfer to unseen graphs. We establish bounds on group-indicator distortion and supernode purity. To convert the learned target into group predictions, we adapt Ward clustering, introduce a merge score linked to coarsening error, and derive sufficient conditions for exact group recovery. Experiments demonstrate high-precision detection across synthetic, financial, and large community graphs, with transfer to unseen graphs.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Group Detection, Graph Coarsening, Subgraph Matching, Graph Neural Network, Anti-money laundering

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 5 reviews, above 85% of the others.
Reviewer_p2PTOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies group detection: identifying sets of nodes whose collective structure is of interest, rather than classifying individual nodes or labeling a supplied list of candidate subgraphs. The proposed approach learns a graph-coarsening geometry from example groups so that previously unseen groups become individual supernodes.

The framework has three parts. First, the paper defines two properties of a target subspace: capture, which measures how well it preserves the contrast between a group and its surroundings, and confusability, which measures how much it preserves distinctions within the group. Second, it learns a target subspace using a bank of linear Chebyshev graph filters. The experiments fit these filters through a channel-pooled quadratic surrogate with a generalized eigenvalue solution. Third, an adjacency-constrained, screened version of Ward clustering constructs a hierarchy of connected supernodes. The theoretical analysis connects the subspace properties to coarsening distortion and gives a sufficient score-gap condition under which a target group appears in the hierarchy.

Experiments cover planted synthetic groups, label-derived illicit-wallet groups in Elliptic++, and several community-recovery benchmarks. The strongest results are on Amazon and on temporal transfer in Elliptic++. Results on the additional community benchmarks are more mixed.

My overall impression is that there is a worthwhile idea here, supported by mathematical development. However, the experiments do not yet establish the full group-detection claim, and the scalability of the actual coarsening algorithm needs much clearer evidence. There are also reporting inconsistencies that should be resolved before the numerical results can be assessed confidently.

Critical strengths

The main strengths, for me, are:

  1. Problem formulation. The paper makes a clear distinction between flagging individual nodes and recovering the groups they form. Learning a coarsening geometry from example groups is an interesting way to address that gap without requiring candidate subgraphs upfront.

  2. Theoretical framework. Capture and confusability give useful interpretations of two competing requirements: keeping a group separate from its surroundings while allowing its members to merge. The analysis also explicitly handles the mismatch between Euclidean block averaging and screened projection, rather than overlooking it.

  3. A learned component. The Chebyshev filter bank has relatively few parameters, and the pooled training surrogate admits a generalized eigenvalue solution. Applying the same filters to new graphs is an appealing design, even though the scalability of the subsequent coarsening stage needs clarification.

  4. Some meaningful experimental choices. The temporal evaluation on Elliptic++ is valuable, and the authors take care to remove obvious temporal cues. Deduplicating Amazon communities before splitting is another important precaution. The appendix also acknowledges the proxy nature of the illicit groups and reports less favorable results on DBLP, which helps make the evaluation more transparent.

Critical weaknesses

The most important weaknesses, in my view, are:

  1. The evaluation does not fully establish detection. Most metrics ask whether each ground-truth group has a good match somewhere in the output partition. They do not measure how many returned blocks are irrelevant. This is a significant gap for a method motivated by detecting rare suspicious groups.

  2. Scalability of the full method is insufficiently documented. Sparse Chebyshev filtering is efficient, but the actual algorithm uses global, partition-dependent deflated Ward scores. The paper does not adequately explain how these are updated at scale or provide runtime and memory measurements.

  3. The comparisons do not isolate the contribution. Louvain does not receive the training-group supervision available to the proposed method. On Elliptic++, both the classifier and the grouping procedure change, making it difficult to attribute the improvement specifically to coarsening.

  4. The theory is not tightly connected to the implemented learning procedure. The experiments optimize a surrogate without a general approximation guarantee to the exact objective. The recovery theorem establishes that a group appears somewhere in the hierarchy, not that the selected cut recovers it. Neither the certificate’s practical tightness nor the surrogate’s effect on capture and confusability is demonstrated sufficiently.

  5. The claimed structural scope exceeds the evidence. The experiments mainly study planted groups with varying internal density and conventional communities. They do not clearly establish performance on disassortative structures. Separately, restricting outputs to connected blocks excludes targets whose members connect only through outside nodes.

  6. Reporting inconsistencies reduce confidence. The incompatible precision, recall, and F1 values in Appendix E.3 need correction, alongside conflicting descriptions of feature dimensions and detection thresholds. These may be editorial errors, but they affect reproducibility and interpretation.

Questions for authors
  1. Which output blocks are actually reported as target groups? Beyond matching each ground-truth group to its best block, can you report detection precision and recall that account for irrelevant predictions?

  2. How are the exact deflated Ward scores computed on the large graphs? Please clarify whether the experiments implement Equation 17 exactly, how scores are updated after a merge, and the runtime and peak memory of the full pipeline.

  3. How much of the improvement comes from coarsening itself? On Elliptic++, could you compare connected components and learned coarsening using identical node predictions? For community recovery, how does the method compare with baselines that also receive example communities?

  4. Does the pooled surrogate improve the quantities used in the theory? Can you measure capture, confusability, and the exact edge objective before and after training, and compare against direct optimization on smaller graphs?

  5. How informative is the recovery certificate in practice? What fraction of target groups satisfy U_S<L_S? Also, what additional conditions would ensure recovery at the selected global cut, rather than appearance somewhere in the hierarchy?

  6. What happens on disassortative or heterophilous structures? Could you evaluate connected targets that are not internally dense and distinguish this setting from disconnected targets, which the connected-block restriction cannot recover?

  7. Does Elliptic++ inference depend on knowing which test wallets have labels? The appendix describes majority voting over known-label members. Please clarify whether label availability affects predictions or only evaluation, and how unknown wallets influence the reported group purity.

  8. Can you reconcile the reporting inconsistencies? In particular, Appendix E.3 reports macro F1 of 0.911 alongside mean precision of 0.79 and recall of 0.89, which are incompatible under the stated averaging. Please also resolve the differing feature dimensions and detection thresholds.

7 comments

We thank the reviewer for the valuable and constructive feedback. In response, we substantially expanded the empirical evaluation (Appendices E.6--E.11): comparisons with learned baselines on community and Elliptic++ benchmarks, an overlapping-group cut, runtime and memory measurements, ablations of the learned target, and measurements connecting our results to the theoretical guarantees. These new results further demonstrate the performance of our method and connect to the empirical findings to our theory. Our point-by-point responses follow. All references to figures, appendices, etc., refer to the updated version of the paper.

Setting and benchmarks. Learning to detect groups from labeled example groups and transferring to unseen graphs is a new problem setting, and to our knowledge there are very few dedicated methods and no established benchmarks for it. The closest existing methods come from community detection and search (ProCom, CLARE, SEAL) and from financial group detection (RevTrack), and we compare against all of them. Because no ready-made benchmark exists, we construct our evaluation from the strongest available sources: the community benchmarks used by ProCom (Amazon, DBLP, LiveJournal) with fixed train/test community splits, Elliptic++ with held-out time steps, and synthetic graphs where the ground-truth groups and their shapes and sizes are controlled.

W1& Q1

Our evaluation penalizes irrelevant output, and our method returns far fewer and more precise groups than the baselines. We added ranked and one-to-one metrics in which every unmatched output counts as a false positive, in addition to the best-match recovery metrics. Our method scores clearly higher on them (Appendices E.4, E.6, and E.7).

How the reported groups are selected. The learned spectral geometry makes target groups merge at low cost, and explicit selection rules that use only the training groups then separate them from background blocks. On the community benchmarks, the learned cut scores every block of the coarsening hierarchy with a quality model trained on the training communities, using the block's size, graph structure, and merge history. It selects the blocks that maximize predicted quality net of a per-group penalty, discards singletons, and ranks the output by this score (Appendix E.6). This requires no node-level class labels. On Elliptic++, we additionally learn a node classifier on the wallet features and the learned screened responses, and flag a supernode as illicit when most of its members are predicted illicit (Appendix E.4).

Metrics that penalize irrelevant blocks. We report detection precision, the fraction of reported groups that match a test group, and detection recall, the fraction of test groups that are matched, under a one-to-one matching (more than 51% overlap in both directions), so that every unmatched output is a false positive. Every unmatched prediction, including a second detection of an already matched group, counts against precision. On the community benchmarks, we rank each method's output by confidence and report average precision (AP) and P@∣T∣|T|, the fraction of hits among the top ∣T∣|T| predictions, where ∣T∣|T| is the number of test communities (Appendix E.6, Tables 5--6). On Elliptic++, we report matched-group precision and F1 under the same one-to-one rule (Appendices E.4 and E.7).

Results. With 100 training communities, our learned cut reaches AP 0.810 on Amazon, 0.376 on DBLP, and 0.465 on LiveJournal, versus at most 0.500, 0.211, and 0.289 for ProCom, CLARE, and SEAL, and it also has higher P@∣T∣|T| on all three datasets. On Amazon, it returns 1,147 groups, compared with 5,000 for ProCom. On Elliptic++, our method returns 197 groups on the held-out days, compared with 3,950 for GNN + connected components. Its matched-group precision is 0.266 versus 0.061, and it obtains the highest matched-group F1 (0.206, versus at most 0.098 for ProCom, CLARE, SEAL, RevClassify, and GNN + connected components). RevClassify returns a list of similar size (about 195 groups) but matches only 5 gangs on average, versus 29 for our method (Table 7).

W2 & Q2

Implementation. We revised Algorithm 2 and expanded Appendix D.1. Each Ward round merges only two supernodes, so the deflated Ward scores can be updated incrementally instead of being recomputed: after a merge, only the pairs that involve the new block and its neighbors are re-scored, and the rank-one projector update (Lemma 20) keeps each update cheap. The implementation never forms an N×NN\times N matrix, and with these incremental updates, our method scales near-linearly in practice. The implementation proceeds as follows.

  1. Per-block state. Each block stores its volume, cut weight, neighbor weights, and two qq-vectors, pC=Uτ⊤vCp_C=U_\tau^\top v_C and mC=Uτ⊤MτvCm_C=U_\tau^\top M_\tau v_C, where UτU_\tau is an MτM_\tau-orthonormal basis of the target. A merge combines them linearly, in O(q+degree)O(q+\text{degree}) time.
  2. Closed-form coarse geometry. The block Gram matrix VP⊤MτVPV_{\mathcal P}^\top M_\tau V_{\mathcal P} equals the volume-normalized coarse Laplacian plus τI\tau I, so it is read directly from the block adjacency.
  3. Local deflation. The deflated score of a candidate merge needs one small solve with this matrix, supported on the two blocks and their coarse neighbors. Screening keeps it well conditioned, so we solve on a 2-hop ball capped at 32 blocks.
  4. Rank-one update. After a merge, the block projector changes by a rank-one term (Lemma 20), so the target-residual matrix is updated in O(q2)O(q^2) and the cumulative score is read off without global recomputation.
  5. Lazy priority queue. Only pairs involving the new block are rescored (at most 32 new queue entries per merge), and other entries keep their keys, which lower-bound their deflated scores. The search stops as soon as the best evaluated score is below the smallest remaining key.

Cost. The cost per merge is Rt(B3+Bq)+q2+δtqR_t(B^3+Bq)+q^2+\delta_t q, where RtR_t is the number of rescored candidates, B≤32B\le 32 the local ball size, and δt\delta_t the degree sum of the merged blocks. Together with the queue operations, this gives the near-linear runtime bound of Eq. (129), which holds when qq, the average rescoring count, and the average coarse degree stay bounded.

Runtime. We also document the runtime in Appendix E.8, which gives the complexity analysis of the deflated Ward coarsening (Eq. 129) and the runtime and peak memory of every experiment (Table 9).

In these measurements, the global-cut pipeline takes 10.5 min with 3.11 GB on full SNAP Amazon (334,863 nodes) and 14.5 min with 3.71 GB on full SNAP DBLP (317,080 nodes). The eight synthetic graphs take 56.1,s with 0.57 GB, and the 15 Elliptic++ days take 32.6 min with 5.70 GB. On 10K--100K-node subgraphs, coarsening time scales approximately linearly (log--log slope 1.07). All of our experiments run on a single laptop CPU (Apple M2 Pro, 16 GB), with no GPU.

W3 & Q3

Thank you for this point. We agree that the contribution should be isolated, and the updated paper adds new experiments for exactly this purpose (Appendices E.6--E.10). Each one holds everything fixed except the component under test. We describe them below, in the order of the two concerns.

(1) Louvain does not use the training groups. In Appendix E.6 (Tables 5--6) we added ProCom, CLARE, and SEAL, which are trained on the same 100 training communities as our method and are evaluated with the same matching rule. Louvain is kept only as an unsupervised reference. With the same supervision, our learned cut reaches AP 0.810 on Amazon, 0.376 on DBLP, and 0.465 on LiveJournal, versus at most 0.500, 0.211, and 0.289 for these three baselines.

(2) Isolating the learned target. In Appendix E.9 (Tables 10--11, Figure 9), which is new, we keep the training groups, the Ward coarsening, and the cut rule fixed and change only the target subspace. We compare our learned target with resolvent powers, a low-pass bank, and a random bank.

  • On the synthetic graphs, F1 is 0.738 for the learned target versus 0.362 for the best fixed target, and the detection rate is 75.6% versus 3.3%.
  • On Amazon, F1 is 0.875 versus 0.857 and the detection rate is 96.4% versus 93.2%.
  • Table 11 shows that this gain comes from the quantities our theory identifies. On the test groups, capture is 0.0528 for the learned target versus 0.0248 and 0.0189 for the fixed banks, and mean confusability is 0.0065 versus 0.0328 and 0.0462.

The same conclusion holds against the static spectral target in Appendix E.2 of the submission (F1 0.832 versus 0.376 at density 0.5; detection rate 0.911 versus 0.250). In Appendix E.10 (new), our closed-form target matches Adam training from its initialization on Amazon (F1 0.875 versus 0.874) and is better on the synthetic graphs (0.738 versus 0.616).

W3 & Q3 (Continue)

(3) Isolating the cut rule. In Appendix E.5 of the submission, we replace our learned cut with an oracle cut that knows the true group boundaries. Our method detects 699 versus 701 groups on Amazon and 1,410 versus 1,418 on DBLP, so the learned cut recovers nearly all that the learned target makes recoverable. The gains therefore come from the target, not from the cut.

(4) Elliptic++: separating the classifier from the grouping. In Appendix E.7 (Table 7, new) we compare against two controls.

  • GNN + connected components predicts illicit nodes and forms groups from connected components. It changes only the grouping step.
  • Ward + GCN applies the same Ward clustering to supervised GCN features. It changes only the target.

Our method is best against both. B$^3$ F1 is 0.416 versus 0.204 and 0.320, and matched-group F1 is 0.206 versus 0.098 and 0.129. It also returns 197 groups versus 3,950 for GNN + connected components, so the flagged groups are far more selective. The other group-level baselines on Elliptic++ (RevClassify and ProCom) reach matched-group F1 of 0.048 and 0.076.

W4 & Q4

The theory is tightly connected to the implemented procedure, in three ways.

  • The recovery analysis applies to the surrogate's target. The recovery analysis holds for any target subspace RΘ\mathcal R_\Theta, including the subspace produced by the pooled surrogate, and does not require that Θ\Theta optimize Eq.~(16). The question is how effectively surrogate training produces high capture and low confusability, and the surrogate response controls both directly. Write
FΘ=1dZΘ,KΘ=FΘ⊤MτFΘ,F_\Theta=\frac{1}{\sqrt d}Z_\Theta, \qquad K_\Theta=F_\Theta^\top M_\tau F_\Theta,

Let aΘ>0a_\Theta > 0 and bΘb_\Theta be the smallest positive and largest eigenvalues of KΘK_\Theta, and define the surrogate response energy of a graph signal xx as Esur(x)=∥FΘ⊤Mτx∥22E_{\mathrm{sur}}(x)=\lVert F_\Theta^\top M_\tau x\rVert_2^2. The exact retained energy is ∥QRΘτx∥Mτ2=(FΘ⊤Mτx)⊤KΘ†(FΘ⊤Mτx)\lVert Q^\tau_{\mathcal R_\Theta}x\rVert_{M_\tau}^2=(F_\Theta^\top M_\tau x)^\top K_\Theta^\dagger (F_\Theta^\top M_\tau x), so

\le \lVert Q^\tau_{\mathcal R_\Theta}x\rVert_{M_\tau}^2 \le \frac{E_{\mathrm{sur}}(x)}{a_\Theta}$$ Applying this to the normalized complete-group direction $\widehat v_S$, and to the internal fluctuations through $\chi_{\mathrm{sur}}(S)=\max_{w\in\widehat F_S^\tau}E_{\mathrm{sur}}(w)$, gives

\frac{E_{\mathrm{sur}}(\widehat v_S)}{b_\Theta} \le C^\tau_{\mathcal R_\Theta}(S) \le \frac{E_{\mathrm{sur}}(\widehat v_S)}{a_\Theta}, \qquad \frac{\chi_{\mathrm{sur}}(S)}{b_\Theta} \le \chi^\tau_{\mathcal R_\Theta}(S) \le \frac{\chi_{\mathrm{sur}}(S)}{a_\Theta}.

Thus, surrogate response energies control the exact capture and confusability through the conditioning of the joint screened Gram matrix. The surrogate objective acts on these energies through boundary and internal edge contrasts. Its internal penalty suppresses the variation of $\ell_\Theta^{\mathrm{sur}}=\widetilde D^{-1/2}M_\tau F_\Theta$ along internal edges, and for a connected group a zero internal penalty makes this response constant on the group, which implies zero exact confusability. The boundary term encourages separation from neighboring nodes. - **Training improves exactly the quantities in the theory.** In Appendix E.10, the closed-form solution attains 90.1\% (synthetic) and 87.8\% (Amazon) of the best exact objective found by Adam through the full Gram whitening, at 0.5 s versus 280 s and 4.4 s versus 463 s. Further exact-objective refinement does not improve recovery: initialized from the closed form, Adam gives group F1 of 0.616 versus 0.738 on synthetic graphs and 0.874 versus 0.875 on Amazon. In Appendix E.9 (Table 11), training raises test capture from 0.019 to 0.053 on synthetic graphs and from 0.008 to 0.071 on Amazon, compared with untrained random filters, and lowers mean confusability on synthetic graphs from 0.046 to 0.0065 and median confusability on Amazon from 0.0157 to 0.0014. - **The selected cut recovers groups.** We designed three cut rules on the same hierarchy: a global cut, a learned cut, and a nested cut for overlapping groups, and compare them in Appendix E.6. The learned cut improves group F1 over the global cut in all six settings, for example from 0.865 to 0.902 on Amazon, from 0.696 to 0.716 on DBLP, and from 0.739 to 0.746 on LiveJournal with 100 training communities, and the nested cut raises it further to 0.923, 0.741, and 0.763. The training-selected cut also loses little to the oracle cut chosen with test labels (Appendices E.2 and E.5): at density 0.5, group F1 is 0.832 versus 0.852, and on held-out Amazon communities it detects 699 of 759 versus 701 with the oracle cut. On DBLP, it detects 1,410 versus 1,418 of 2,500. All 759 held-out Amazon communities appear as blocks at some level of the learned hierarchy.
Reviewer_zL2tOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper formulates group detection as learned graph coarsening, aiming to directly recover coordinated sets of nodes rather than first classifying individual nodes. It characterizes a suitable coarsening target subspace through two properties: high capture of the complete-group direction and low confusability with within-group variations. The target subspace is constructed using transferable Chebyshev graph filters learned from annotated training groups. At inference, the learned filters are applied to an unseen graph, and a target-aware variant of Ward clustering greedily merges adjacent blocks according to the amount of target information that each merge would discard. The paper further derives sufficient conditions under which a target group forms a supernode before any boundary-crossing merge occurs. Experimental results demonstrate that the proposed method achieves certain improvements.

Critical strengths
  1. The paper treats groups as the unit of prediction and formulates group detection as a learned graph coarsening problem. This perspective enables the model to jointly infer group membership and group structure, which is closer to the practical requirements of such tasks, and it also provides a new supervised framework for group discovery that is distinct from node classification and traditional community detection.
  2. Through the two complementary quantities, capture and confusability, the paper respectively characterizes whether the target subspace preserves the overall information of a group and whether it suppresses intra-group differences; it further connects these properties to coarsening error, Ward merge scores, and the conditions for exact group recovery.
  3. The paper uses learnable Chebyshev graph filters to construct the target subspace, avoiding reliance on node identities from training graphs or explicit eigendecompositions on new graphs, while remaining applicable to relatively large-scale graphs through sparse graph operations. The experiments cover controlled synthetic graphs, a financial network with a temporal transfer setting, and a community graph containing hundreds of thousands of nodes; results show that the learned target has clear advantages over static spectral targets, post-hoc node classification, and baselines such as Louvain.
Critical weaknesses
  1. Most experiments match each real group to the predicted block with the largest overlap and compute precision, recall, and F1 based on this best match. The large number of output blocks that match no real group essentially do not reduce these metrics. Therefore, even if the method produces tens of thousands of candidate blocks, it can still obtain a very high "group precision." Apart from the node classifier additionally used in Elliptic++, the paper also does not explain how to distinguish target supernodes from ordinary background blocks.

  2. Capture, confusability, and the exact recovery conditions are all based on the jointly Gram-normalized screened response and the exact objective in Eq. (16), but the experiments actually optimize the channel-pooled quadratic surrogate given in the appendix.

  3. The theoretical part assumes that target groups are connected and pairwise disjoint, but the communities in Amazon and DBLP clearly overlap, and a single graph partition cannot simultaneously represent multiple overlapping groups; moreover, the same predicted block may be repeatedly selected as the best match by multiple real groups. At the same time, the main comparison method on the community datasets is essentially only Louvain, lacking more direct and stronger baselines such as recent supervised community detection, subgraph discovery, learned hierarchical clustering, and edge-classification-then-clustering approaches.

Questions for authors

See weaknesses.

6 comments

We thank the reviewer for the valuable and constructive feedback. In response, we substantially expanded the empirical evaluation (Appendices E.6--E.11): comparisons with learned baselines on community and Elliptic++ benchmarks, an overlapping-group cut, runtime and memory measurements, ablations of the learned target, and measurements connecting our results to the theoretical guarantees. These new results further demonstrate the performance of our method and connect to the empirical findings to our theory. Our point-by-point responses follow. All references to figures, appendices, etc., refer to the updated version of the paper.

W1

Our method learns a spectral geometry from example groups, encouraging high capture of group-level contrasts and low confusability with within-group variations. Coarsening under this geometry produces supernodes whose formation incurs relatively little loss of information from the learned target subspace, so target blocks are distinguished from background blocks by explicit selection rules that use only the training groups, and our evaluation accounts for irrelevant output blocks.

How target blocks are distinguished from background blocks. When node-level class labels are available for training, as on Elliptic++, we additionally learn a node classifier on the wallet features and the learned screened responses and flag a supernode as illicit when a majority of its members are predicted illicit. This is the explicit selection mechanism that separates candidate illicit groups from background blocks (Appendix E.4). When such labels are unavailable, as on the community benchmarks, the training communities supervise the spectral geometry, and inference produces a partition of the graph whose blocks are potential groups compatible with that geometry. The learned cut then scores every block of the hierarchy with a quality model trained on the training communities, using the block's size, graph structure, and merge history, selects the blocks that maximize predicted quality net of a per-group penalty, discards singletons, and ranks the output by this score (Appendix E.6). This requires no node-level class labels.

Metrics that penalize irrelevant blocks. In addition to best-match recovery, Appendix E.6 reports average precision (AP) and precision at depth ∣T∣|T| (P@∣T∣|T|), where predictions are matched one-to-one to test communities, and every unmatched prediction, including a repeated detection of an already matched community, is a false positive. On Elliptic++, matched-group precision and F1 use one-to-one matching and account for unmatched predicted groups (Appendices E.4 and E.7).

Results. With 100 training communities, our learned cut reaches AP 0.810 on Amazon, 0.376 on DBLP, and 0.465 on LiveJournal, versus at most 0.500, 0.211, and 0.289 for ProCom, CLARE, and SEAL. On Amazon, it returns 1,147 groups, compared with 5,000 for ProCom. On Elliptic++, our method returns 197 groups on the held-out days, compared with 3,950 for GNN + connected components, with matched-group precision 0.266 versus 0.061 and the highest matched-group F1 (0.206 versus at most 0.098 for all baselines).

W2

The surrogate is a faithful and much cheaper way to optimize the exact objective, and our guarantees hold for the subspace it produces. The recovery analysis holds for any target subspace RΘ\mathcal R_\Theta and does not require that Θ\Theta optimize Eq.~(16). The surrogate response energies bound the exact capture and confusability, and in experiments the closed-form surrogate retains 87--96% of the best exact objective, matches the recovery of direct optimization, and is 100--600$\times$ faster.

Why the surrogate controls capture and confusability. Write

FΘ=1dZΘ,KΘ=FΘ⊤MτFΘ,F_\Theta=\frac{1}{\sqrt d}Z_\Theta, \qquad K_\Theta=F_\Theta^\top M_\tau F_\Theta,

let aΘ>0a_\Theta>0 and bΘb_\Theta be the smallest positive and largest eigenvalues of KΘK_\Theta, and define the surrogate response energy of a graph signal xx as Esur(x)=∣FΘ⊤Mτx∣22E_{\mathrm{sur}}(x)=\\|F_\Theta^\top M_\tau x\\|_2^2. The exact retained energy is

=(F_\Theta^\top M_\tau x)^\top K_\Theta^\dagger (F_\Theta^\top M_\tau x)$$ so $$\frac{E_{\mathrm{sur}}(x)}{b_\Theta} \le \lVert Q^\tau_{\mathcal R_\Theta}x\rVert_{M_\tau}^2 \le \frac{E_{\mathrm{sur}}(x)}{a_\Theta}$$Applying this to the normalized complete-group direction $\widehat v_S$, and to the internal fluctuations through $\chi_{\mathrm{sur}}(S)=\max_{w\in\widehat F_S^\tau}E_{\mathrm{sur}}(w)$, gives

\frac{E_{\mathrm{sur}}(\widehat v_S)}{b_\Theta} \le C^\tau_{\mathcal R_\Theta}(S) \le \frac{E_{\mathrm{sur}}(\widehat v_S)}{a_\Theta}, \qquad \frac{\chi_{\mathrm{sur}}(S)}{b_\Theta} \le \chi^\tau_{\mathcal R_\Theta}(S) \le \frac{\chi_{\mathrm{sur}}(S)}{a_\Theta}.

Thus, the surrogate response energies control the exact capture and confusability through the conditioning of the joint screened Gram matrix. The surrogate objective acts on these energies through boundary and internal edge contrasts. Its internal penalty suppresses the variation of $\ell_\Theta^{\mathrm{sur}}=\widetilde D^{-1/2}M_\tau F_\Theta$ along internal edges, and for a connected group a zero internal penalty makes this response constant on the group, which implies zero exact confusability. The boundary term encourages separation from neighboring nodes. **Empirical comparison with the exact objective (Appendix E.10, Table 12).** We compared the closed-form surrogate with direct Adam optimization through the full Gram whitening, using the same filter bank, features, training groups, coarsening procedure, and cut-selection rule. - *Closeness.* For the difference objective in Eq. (16), with $\gamma$ set to half the closed-form training ratio, the surrogate attains **95.8\%** of the best objective found on synthetic graphs and **86.6\%** on ProCom Amazon. For the ratio objective used in our experiments, the values are **90.1\%** and **87.8\%**. These are relative to the best solution found within 1,500 optimization steps. - *Cost.* The closed-form fit takes **0.5 s on synthetic graphs and **4.4 s** on Amazon, versus about **280 s** and **463 s** for Adam, which is 100--600\$\times\$ more time. - *Recovery.* Direct optimization from random starts yields group F1 comparable to the closed form, 0.735--0.766 vs.\ 0.738 on synthetic graphs and 0.874--0.877 vs.\ 0.875 on Amazon. Initializing Adam at the closed-form solution raises the training objective by 11.6\% (synthetic) and 14.3\% (Amazon) but gives no recovery gain: F1 is 0.616 vs.\ 0.738 on synthetic graphs and 0.874 vs 0.875 on Amazon. ## W3 **Overlapping groups.** Disjointness is not fundamental to our method. Disjointness enters only through the cut rule that selects blocks from the hierarchy. In Appendix E.6, we therefore designed a nested cut rule for overlapping groups, which allows up to $k$ selected blocks along each root-to-leaf path, so the same hierarchy can return overlapping groups. On Amazon, DBLP, and LiveJournal, whose ground-truth communities overlap, the nested cut improves group F1 over the disjoint cut in all six settings (by 0.016--0.025) and has the highest group F1 of all compared methods in all six. With 100 training communities, it reaches 0.923, 0.741, and 0.763 (Amazon, DBLP, LiveJournal), versus 0.904, 0.460, and 0.451 for the strongest learned baseline (ProCom). On Amazon, it has higher precision than ProCom (0.908 vs 0.867) at equal recall (0.960) with 2,507 output groups instead of 5,000, and on DBLP its detection rate is 0.687 versus 0.375 for ProCom. **Repeated matches.** Appendix E.6 reports ranked-detection metrics (average precision and P@$|T|$) in which predictions are matched one-to-one to test communities, and repeated or unmatched predictions count as false positives. On Elliptic++, matched-group F1 likewise uses a one-to-one matching (Appendix E.7). With 100 training communities, our learned cut reaches average precision 0.376 on DBLP, 0.465 on LiveJournal, and 0.810 on Amazon, versus at most 0.211, 0.289, and 0.500 for ProCom, CLARE, and SEAL.

W3 (Continue)

Baselines. Louvain is no longer our main comparison. We added three recent methods that learn from example communities, ProCom, CLARE, and SEAL, on Amazon, DBLP, and LiveJournal under identical splits and supervision budgets (Appendix E.6, Tables 5--6). On DBLP and LiveJournal, our learned cut gives higher group F1, B$^3$ F1, average precision, and P@∣T∣|T| than all three at both budgets, and on Amazon it matches ProCom's group F1 (0.902 vs 0.904) with much higher average precision. On Elliptic++, we added ProCom, CLARE, SEAL, and RevTrack's RevClassify, a subgraph-classification approach. Our method obtains the highest B$^3$ F1 (0.416) and matched-group F1 (0.206), versus at most 0.204 and 0.098 for any baseline (Appendix E.7, Tables 7--8).

Reviewer_X6L9OpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a completely new approach to the problem of group detection by framing it as a problem of learned graph coarsening. This paper defines a certain type of group as one with high capture and low confusability, and formulates the group detection problem as an objective function that aims to preserve the identity of the group as a whole to the greatest extent possible while simultaneously obscuring the identities within the group to the greatest extent possible. The authors derive conditions that theoretically guarantee exact recovery for this new formulation using the Chebyshev filter, spectral coarsening, and Ward merge ordering. The authors analyze the novelty of this framework by comparing it with the latest methods (SEAL, CLARE, ProCom, RevTrack, SPMiner, and MotiFiesta). The proposed method is effectively demonstrated in real-world applications such as Amazon and Ellipti++.

Critical strengths
  • I believe it is quite novel to formulate the group detection problem as a learned graph coarsening problem (essentially, group recovery from subspace approximation). As the authors argue—and it is highly likely that this paper is the first to adopt such a formulation—I am not aware of any prior research along these lines.

  • I find it truly remarkable how the mechanism for distinguishing between the interior and exterior of a group is theoretically well-organized and clearly articulated. The theoretical analysis of boundary merge and internal merge is non-trivial and insightful. Furthermore, the authors have identified sufficient conditions for the exact recovery of supernodes. In particular, Theorem 11, which provides an end-to-end theoretical analysis, is a very powerful theorem (albeit under several very strong assumptions).

Critical weaknesses
  • I have not been able to fully grasp whether the concept of “group” described in this paper corresponds to the insights that users (engineers seeking to analyze data) wish to discover. In fact, the experiments provide different interpretations of “group” for each application case, and these seem somewhat arbitrary (to put it critically). In terms of actual behavior, it appears to simply be reconstructing the proximity derived from pre-assigned labels. For example, as mentioned in the paper’s motivation, I have some concerns about whether this technique can be applied to data mining scenarios such as the discovery of unknown groups.

  • To be frank, I would like the authors to reconsider the title of this paper. For example, (if my understanding is correct), this technique does not appear to handle scenarios where mixed membership (that is, a standard concept in community detection) occurs. In essence, this paper suggests a development in a direction distinct from various classification-based methods; it does not seem to propose a complete superset of conventional node classification. If this paper intends to assert “beyond node classification” in its title, it follows that the methodology presented within it will inevitably be evaluated from a more rigorous perspective.

Questions for authors

I have found the paper’s new and insightful methodological shift to be very interesting. In particular, I have thoroughly enjoyed reading the novel formulation and its theoretical analysis. On the other hand, I found the experimental section somewhat disappointing. I would be very pleased if the authors could address this point. I expect that the primary use of this method is to learn, within a supervised coarsening framework, the very definition of the “group” that one wishes to capture—even without relying on node classification (in fact, I interpreted it as such a technique and thoroughly enjoyed reading the methodology section). However, in the actual demonstrations, I get the impression that the supervised information provided is essentially very close to node class labels. For example, in the Ellipic++ scenario, it appears that group labels derived from the subgraphs introduced as an auxiliary measure are being used. Although the authors have honestly clarified that these are based on label-derived proximity, it also seems that, in effect, they are being used to solve a task akin to node classification as a supplementary step. Viewed in this light, while the proposed method is indeed very elegant in theory, I get the impression that, in practice, its use is limited to learning a surrogate method for node classification. Am I misunderstanding something here? I would be very grateful if the authors could respond to this point.

2 comments

We thank the reviewer for the valuable feedback. In response to the reviews, we substantially expanded the empirical evaluation with new experiments in Appendices E.6--E.11. All figure, table, and section references in our answers refer to the updated version of the paper submitted for the rebuttal. The new experiments cover the following.

  • Comparison with learned and unsupervised baselines (E.6, E.7). We compare against ProCom, CLARE, SEAL, RevTrack's RevClassify, Louvain, GNN + connected components, and Ward + GCN, under the same training groups and the same one-to-one matching rule. These experiments use ranked metrics that penalize irrelevant outputs, and they include an overlapping cut rule for mixed membership. The datasets are Amazon, DBLP, and LiveJournal (community benchmarks, with 100 or about 50% of the communities for training) and Elliptic++ (train on days 26--30, test on the unseen days 31--40).
  • Complexity (E.8). Runtime and memory on the full SNAP Amazon and DBLP graphs, Elliptic++, and the synthetic graphs, on a laptop CPU.
  • Isolating the learned target (E.9, E.10). Learned filters versus the fixed resolvent, resolvent powers, random filters, and low-pass banks, with the capture and confusability that our theory identifies, and the closed-form surrogate versus direct optimization with Adam, on the synthetic graphs and Amazon.
  • Generalization (E.11). Transfer across group shapes and sizes on controlled synthetic graphs, against the static spectral target.

These results show the performance of our method and connect it to our theoretical guarantees.

W1

Thank you for this thoughtful question. Yes, our approach discovers unknown groups. It learns from example groups what a group looks like spectrally and finds new groups with similar behavior in graphs and time periods it has not seen. It does not need to reconstruct a proximity derived from pre-assigned node labels.

Learning is at the group level. In our objective, the training groups enter only through their collective behavior, not through individual nodes: the complete-group direction and its internal fluctuations define capture and confusability, and the learning and the detection are both driven by these quantities. The filters are learned from how each training group's group-level signal and internal-variation signals are spread over the graph frequencies, not from individual nodes, so groups with similar spectral content are detected alike, whatever their shape or size. Groups that share a similar spectral response are therefore detected as similar groups, whatever their shape or size. Consistent with this, filters trained on cycles recover cliques with F1 0.890, close to the 0.907 of clique-trained filters, and the learned target beats the static spectral target in all nine train/test size combinations (Appendix E.11).

Where node labels are used. On the community benchmarks (Amazon, DBLP, LiveJournal), no node labels are given at all. The training communities are the only supervision, and the unseen communities are discovered from the graph structure. On Elliptic++, node labels are used only after the groups are detected, to classify the detected groups as malicious or benign (Appendix E.4). They do not shape the detection of the groups.

Evidence of discovery on unseen groups.

  • Communities (Appendix E.6): with only 100 training communities (about 2--10% of the communities), our learned cut reaches AP 0.810, 0.376, and 0.465 on Amazon, DBLP, and LiveJournal, versus at most 0.500, 0.211, and 0.289 for ProCom, CLARE, and SEAL.
  • Elliptic++ (Appendix E.7): we train on days 26--30 and detect gangs on the unseen days 31--40. Our method returns 197 groups with matched-group precision 0.266, versus 3,950 groups and 0.061 for GNN + connected components.
  • Synthetic graphs: the node features are random vectors that carry no group information, so recovery comes from the graph structure through the learned filters (Appendices E.1 and E.11). We will add a paragraph to the paper that states this definition of a group explicitly.

W2

Thank you for raising this. By ``beyond node classification'' we mean that the unit of prediction is the group rather than the node. We do not claim to replace node classifiers, and we are happy to make this explicit in the paper and, if the reviewer prefers, to revise the title. Our method addresses problems where the object of interest is a coordinated set of nodes whose members may look unremarkable in isolation, so classifying nodes one at a time does not recover the group.

Group detection and node classification are complementary. The two address different prediction units and can be used together. On Elliptic++, we use a node classifier only to label the detected groups as malicious or benign, and group-level detection is what recovers the gangs (Appendix E.7). On the community benchmarks, no node labels are used at all.

Mixed membership is supported. Disjointness is not fundamental to our method. In the new experiments in Appendix E.6, we design a cut rule, that is, a rule for when to stop the coarsening in the hierarchy, that accounts for overlapping groups. The theory and the algorithm stay the same, and only the cut rule changes. With this rule, our method reaches group F1 of 0.923, 0.741, and 0.763 on Amazon, DBLP, and LiveJournal, versus 0.904, 0.460, and 0.451 for the strongest baseline on each dataset.

Title. We will clarify the meaning of ''beyond'' in the abstract and introduction. We are also open to changing the title, for example, to ``Group Detection through Learned Graph Coarsening''.

Q1

Thank you for the kind words and for this question. Your interpretation is correct. The method learns the definition of a group from example groups, without node classification, and the supervision in our experiments is at the group level.

The supervision is a set of groups. In training, the supervised information is the training groups themselves. They enter the objective only through their collective behavior (the complete-group direction and its internal fluctuations, that is, capture and confusability), and no individual node label is used for detection. On the community benchmarks (Amazon, DBLP, LiveJournal), there are no node class labels at all: the training communities are the only supervision, and the method discovers unseen communities (Appendix E.6). With only 100 training communities, it reaches AP 0.810, 0.376, and 0.465, versus at most 0.500, 0.211, and 0.289 for ProCom, CLARE, and SEAL.

Elliptic++: node labels only name the groups. On Elliptic++, a node classifier is used only as a supplementary step to label the detected groups as malicious or benign. The detection of the groups comes from the learned coarsening. The group-level pipeline is much better than a node-classification pipeline: node classification followed by connected components reaches B$^3$ F1 0.204 and matched-group F1 0.098, versus 0.416 and 0.206 for our method, and it returns 3,950 groups instead of 197 (Appendix E.7). If our method were only a surrogate for node classification, it would not outperform node classification followed by grouping.

The learned definition of a group does not depend on node labels or on the group's shape. In the controlled synthetic experiments, the node features are random vectors with no group information, so recovery comes from the graph structure and the learned collective spectral response. The learned target transfers to unseen group shapes and sizes and beats the static spectral target (Appendix E.11). For evaluation without node-level supervision, we are also running experiments on Elliptic2, which provides labeled suspicious subgraphs, and will update the paper when the results are ready.

Reviewer_wRoFOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies group-level discovery on graphs from examples of known target groups. It learns a graph-specific target subspace using transferable Chebyshev graph filters, encourages agreement on within-group edges and separation at group boundaries, and applies an adjacency-constrained, deflated Ward-style coarsening algorithm to produce a hierarchy of connected supernodes. The theoretical analysis introduces capture and confusability, relates these quantities to projection distortion and merge-score bounds, and derives sufficient conditions for a target group to emerge as a single supernode. Experiments cover planted synthetic groups, temporally held-out Elliptic++ wallet graphs, and large SNAP community networks.

Critical strengths
  1. A principled connection between target-subspace geometry and hierarchical group recovery. The distinction between preserving the complete-group direction through capture and suppressing within-group variations through confusability provides an interpretable formulation of group-aware graph coarsening. The subsequent connection to Ward-style merge scores and conditional recovery guarantees offers a coherent mathematical framework.
  2. Evaluation across diverse graph regimes. The experiments include controlled synthetic graphs, temporal transfer in financial networks, and large real-world community graphs.
Critical weaknesses
  1. Experiments do not adequately establish the claimed group-detection advantage over the closest prior methods. Semi-supervised targeted community detection has already addressed learning from known groups to discover unknown groups. However, none of these directly relevant approaches is included in the experimental comparisons under a matched protocol. The experimental baselines are restricted to unsupervised methods or naive node classification. In Section 2, the authors explicitly discuss closely related semi-supervised or few-shot community and subgraph discovery methods, such as SEAL, CLARE, ProCom and RevTrack. Comparing a supervised method trained on 50% of the ground-truth communities solely against unsupervised Louvain (where Louvain still slightly outperforms the proposed method on com-DBLP in macro F1 and B3B^3 F1) does not sufficiently establish its empirical advantage over existing learning-based group detectors.
  2. Limited algorithmic novelty and an imbalance between theoretical presentation and practical implementation. At its core, the proposed method combines two established techniques: supervised discriminative learning through a generalized eigenvalue problem over Chebyshev graph features, followed by adjacency-constrained Ward-type agglomerative clustering. Although the proposed target-subspace characterization and modified merge score provide a distinct theoretical formulation, the extent of the algorithmic improvement over these classical components remains insufficiently demonstrated. Moreover, parts of the theoretical development appear unnecessarily elaborate for standard mathematical results. For example, Lemma 16 devotes a formal proof to the equivalence between Chebyshev and monomial polynomial bases, while Theorem 19 largely rederives the classical Ky Fan principle for generalized eigenvalue optimization. In contrast, Appendix D.1 provides little explanation of how the partition-dependent MτM_\tau -orthogonal deflation and corresponding merge scores are efficiently updated in Algorithm 2. This is a critical computational issue for graphs containing hundreds of thousands of nodes. The paper would benefit from a clearer distinction between established mathematical tools and its genuinely new contributions, together with stronger evidence that the proposed modifications are both necessary and computationally practical.
Questions for authors
  1. What is the specific methodological improvement over existing targeted community detection methods? Can the authors compare against more existing methods using identical graph splits, supervision budgets, features, and group-level evaluation protocols, or provide concrete empirical evidence explaining why those methods cannot apply to these settings?
  2. Can the authors ensure that supervised and semi-supervised baselines receive comparable supervision? Please report group-level precision, recall, F1, and runtime under consistent evaluation settings to establish whether the proposed method offers a meaningful advantage over existing group discovery approaches.
  3. How much improvement comes specifically from the learned target-subspace formulation and the MτM_\tau-orthogonal deflation? Could the authors add more ablation studies for comparisons, such as conventional supervised discriminative embeddings followed by connectivity-constrained Ward clustering, and the learned embeddings with the raw Ward score instead of the deflated score? These comparisons would help establish whether the added mathematical and computational complexity produces a substantial benefit.
6 comments

We thank the reviewer for the valuable and constructive feedback. In response, we substantially expanded the empirical evaluation (Appendices E.6--E.11): comparisons with learned baselines on community and Elliptic++ benchmarks, an overlapping-group cut, runtime and memory measurements, ablations of the learned target, and measurements connecting our results to the theoretical guarantees. These new results further demonstrate the performance of our method and connect to the empirical findings to our theory. Our point-by-point responses follow. All references to figures, appendices, etc, refer to the updated version of the paper.

Setting and benchmarks. Learning to detect groups from labeled example groups and transferring to unseen graphs is a new problem setting, and to our knowledge there are very few dedicated methods and no established benchmarks for it. The closest existing methods come from community detection and search (ProCom, CLARE, SEAL) and from financial group detection (RevTrack), and we compare against all of them. Because no ready-made benchmark exists, we construct our evaluation from the strongest available sources: the community benchmarks used by ProCom (Amazon, DBLP, LiveJournal) with fixed train/test community splits, Elliptic++ with held-out time steps, and synthetic graphs where the ground-truth groups and their shapes and sizes are controlled.

W1

We added comparisons with the closest learning-based methods that the reviewer asks for, SEAL, CLARE, ProCom, and RevTrack's RevClassify, under a matched protocol (Appendices E.6 and E.7). All baselines are trained on the same communities as our method and scored with the same matching rule. These experiments establish that our method outperforms the closest learning-based group detectors. To address the concern about supervision, we evaluate two budgets: a few-shot setting with only 100 training communities, which is about 8% of the communities on Amazon, 2% on DBLP, and 10% on LiveJournal, and a larger setting with about 50% of the communities.

Community benchmarks (Appendix E.6, Tables 5--6). We compared against ProCom, CLARE, and SEAL on the pruned Amazon, DBLP, and LiveJournal benchmarks distributed with ProCom. All methods use the same graphs, the same seeded community splits (three seeds), and the same available supervision, and test communities are used only for evaluation. Our learned cut has higher average precision than ProCom, CLARE, and SEAL on Amazon, DBLP, and LiveJournal, at both training budgets. With 100 training communities, it reaches AP 0.810, 0.376, and 0.465, versus at most 0.500, 0.211, and 0.289 for the baselines. On DBLP and LiveJournal, group F1 rises from 0.460 to 0.716 and from 0.451 to 0.746. On the pruned DBLP benchmark with 50% of the communities for training, our method reaches 0.727 group F1 versus 0.694 for Louvain.

Elliptic++ (Appendix E.7, Tables 7--8). We extended the temporal comparison (fit on days 26--30, test on days 31--40) to ProCom, CLARE, SEAL, and RevTrack's RevClassify, with the community baselines given the fitting-day gangs as examples. Our method has the highest B$^3$ F1 (0.416) and matched-group F1 (0.206) among all methods, versus 0.204 and 0.098 for GNN + connected components, 0.198 and 0.048 for RevClassify, and 0.093 and 0.076 for ProCom. It also returns 197 groups versus 3,950 for GNN + connected components.

Louvain. Louvain is an unsupervised reference, and its operating level is selected with the same training-F1 rule as ours. On the pruned DBLP benchmark, our learned cut reaches 0.716 group F1 with 100 training communities and 0.727 with 50%, versus 0.696 and 0.694 for Louvain, and it has higher group F1, average precision, and P@∣T∣|T| than Louvain in all six settings of Tables 5--6.

W2

Novelty. Our contribution is a unified framework in which the learning step and the coarsening step are tied to the same group-level quantities, and not a combination of off-the-shelf components. We introduce capture and confusability, which measure how much of a group's complete-group direction a target subspace retains and how much of its internal variation it also retains. We show that (i) they control the coarsening error and the purity of the block containing a group (Proposition 2, Theorem 5); (ii) they can be expressed through the learned filter response (Eq. 15), which gives our learning objective (Eq. 16) and its closed-form surrogate; and (iii) they determine the Ward merge scores through a deflated score linked to the coarsening error (Eq. 17, Propositions 8--9), which yields a sufficient condition for exact recovery of a group as a single supernode (Theorem 11). The learned filters and the deterministic coarsener are thus connected through the group parameters, and one analysis covers the whole pipeline from training groups to recovery. The objective, merge score, and recovery condition are derived from this analysis, and differ from classical supervised discriminative embeddings and from standard Ward linkage.

Established tools and new contributions. We agree that Lemma 16 and Theorem 19 use standard mathematics and will shorten them in the revision, adding a short paragraph that separates the established tools (Chebyshev filters, Ward linkage, generalized eigenproblems) from our new contributions. Lemma 16 only shows that Chebyshev and monomial filter banks span the same targets, so the basis choice affects numerical conditioning but not the model class. Theorem 19 reduces learning to a single small generalized eigenproblem with an exact global solution for the pooled surrogate, which avoids iterative training: 0.5 s versus 280 s on synthetic graphs and 4.4 s versus 463 s on Amazon, with comparable recovery (Appendix E.10).

Efficient implementation of Algorithm 2. As written, Algorithm 2 rescores every adjacent pair after each merge. We revised and expanded Appendix D.1 to show the incremental implementation we use, which never forms an N×NN\times N matrix and rescores only the pairs near each merge:

  1. Per-block state. Each block stores its volume, cut weight, neighbor weights, and two qq-vectors, pC=Uτ⊤vCp_C=U_\tau^\top v_C and mC=Uτ⊤MτvCm_C=U_\tau^\top M_\tau v_C, where UτU_\tau is an MτM_\tau-orthonormal basis of the target. A merge combines them linearly, in O(q+degree)O(q+\text{degree}) time.
  2. Closed-form coarse geometry. The block Gram matrix VP⊤MτVPV_{\mathcal P}^\top M_\tau V_{\mathcal P} equals the volume-normalized coarse Laplacian plus τI\tau I, so it is read directly from the block adjacency.
  3. Local deflation. The deflated score of a candidate merge needs one small solve with this matrix, supported on the two blocks and their coarse neighbors. Screening keeps it well conditioned, so we solve on a 2-hop ball capped at 32 blocks.
  4. Rank-one update. After a merge, the block projector changes by a rank-one term (Lemma 20), so the target-residual matrix is updated in O(q2)O(q^2), and the cumulative score and distortion are read off without global recomputation.
  5. Lazy priority queue. Only pairs involving the new block are rescored (at most 32 new queue entries per merge), and other entries keep their keys, which lower-bound their deflated scores. The search stops as soon as the best evaluated score is below the smallest remaining key.

The cost per merge is Rt(B3+Bq)+q2+δtqR_t(B^3+Bq)+q^2+\delta_t q, where RtR_t is the number of rescored candidates, B≤32B\le 32 the local ball size, and δt\delta_t the degree sum of the merged blocks. Together with the queue operations, this gives the near-linear bound and runtimes of Appendix E.8.

Computational cost. Our framework is inexpensive: all experiments run on a single laptop CPU (Apple M2 Pro, 16 GB), with no GPU. The target is built with sparse Chebyshev filtering, training is a closed-form eigenproblem (0.5 s on synthetic graphs, 4.4,s on Amazon), and each coarsening merge is a rank-one update of the block projector (Lemma 20) with local rescoring. Appendix E.8 gives the full complexity analysis, including the runtime bound for deflated Ward coarsening (Eq. 129), which is near-linear when the target width and local neighborhood stay bounded, and we will expand Appendix D.1 to describe the update procedure. In practice, the full pipeline takes 10.5 min with 3.1 GB on SNAP Amazon (335K nodes), 14.5 min with 3.7 GB on SNAP DBLP (317K nodes), and 32.6 min with 5.7 GB for all 15 Elliptic++ days.

W2 (Continue)

Evidence that the modifications matter. Replacing the learned target with fixed or random filters under the same coarsener lowers synthetic group F1 from 0.738 to at most 0.362 and the detection rate from 75.6% to at most 3.3%, and lowers Amazon group F1 from 0.875 to at most 0.857 (Appendix E.9). The learned targets also have higher capture and about fivefold lower confusability on synthetic graphs, the properties that our analysis identifies as required for recovery (Table 11. Against learning-based group detectors, the full method reaches 0.716 versus 0.460 group F1 on DBLP and 0.206 versus 0.098 matched-group F1 on Elliptic++ (Appendices E.6--E.7).

Q1

Methodological improvements. Existing targeted community detection methods (ProCom, CLARE, SEAL) generate or match candidate communities heuristically, from seeds or from a pretrained encoder, and return fixed-size candidate lists. Our approach differs in four ways.

  1. Group-level spectral learning, beyond node classification. The group, not the node, is the prediction unit. We learn from example groups a spectral target that preserves the signal distinguishing a whole group from its surroundings (high capture) while suppressing the signals distinguishing its members from each other (low confusability). A group is thus detected from its collective structure, even when its members look unremarkable individually, which is the situation in gang detection.
  2. A learning objective derived from group metrics. Capture and confusability are expressed through the learned filter response (Eq. 15), which gives our training objective (Eq. 16) with a closed-form solution (Theorem 19), so training takes seconds.
  3. Theoretical guarantees. The same quantities bound the coarsening error and the purity of the block containing a group (Proposition 2, Theorem 5) and give a sufficient condition for exact recovery of a group as a single supernode through our merge score (Theorem 11). Existing targeted methods have no such guarantee linking the learned representation to recovery.
  4. Inductive transfer. The learned Chebyshev filters apply unchanged to unseen graphs, where they construct graph-specific targets, and the hierarchy needs no prescribed number of outputs. We fit on Elliptic++ days 26--30 and detect gangs on unseen days 31--40, whereas ProCom, CLARE, and SEAL must be run transductively on the evaluation graphs (Appendix E.7).

Matched comparisons. We compared against ProCom, CLARE, and SEAL using identical graphs, community splits, supervision budgets, and group-level evaluation protocol on Amazon, DBLP, and LiveJournal (Appendix E.6, Tables 5--6). With only 100 training communities, about 2--10% of the communities, our learned cut improves group F1 from 0.460 to 0.716 on DBLP and from 0.451 to 0.746 on LiveJournal, and average precision from 0.211 to 0.376 and from 0.289 to 0.465. On Amazon, it matches ProCom's group F1 (0.902 vs 0.904) with much higher average precision (0.810 vs 0.500). Higher average precision holds at the 50% budget on all three datasets. On Elliptic++ (Appendix E.7), we added ProCom, CLARE, SEAL, and RevTrack's RevClassify on the same temporal split. Our method reaches B$^3$ F1 0.416 and matched-group F1 0.206, versus at most 0.204 and 0.098 for any baseline. All of these methods could be applied. For RevTrack, whose discovery component RevFilter needs candidate sender and receiver sets that are unavailable for unseen gangs, we evaluate RevClassify on externally generated candidates.

Q2

Comparable supervision. In Appendix E.6, all methods use the same graphs, the same seeded community splits (three seeds), and the same training communities, at two budgets: 100 training communities and about 50% of the communities. Test communities are used only for evaluation, and CLARE reserves 10% of its training communities for validation. Louvain, which uses no labels, has its operating level selected by the same training-F1 rule as ours. On Elliptic++ (Appendix E.7), ProCom, CLARE, and SEAL are given the fitting-day gangs (days 26--30) as examples and are evaluated on the unseen days 31--40.

Reviewer_TatzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses group detection: finding small, coordinated sets of nodes (e.g., money-laundering gangs) whose members look normal individually. It frames this as learned graph coarsening: instead of a fixed spectral target, a bank of learnable Chebyshev graph filters learns, from training groups, a target subspace under which unseen groups emerge as single supernodes. The authors characterize the target through capture (how much of the group-versus-rest signal it retains) and confusability (how much within-group variation it retains), relate both to coarsening error, and adapt Ward clustering with a merge score that comes with a sufficient condition for exact group recovery. Experiments on synthetic planted groups, Elliptic++ Bitcoin graphs (temporal transfer), and the com-Amazon graph (335k nodes) show gains over a static spectral target, GNN + connected components, and Louvain, respectively.

This is an interesting idea with solid theoretical grounding, but the empirical evaluation is not yet sufficient to support the paper’s claims, particularly its anti-money laundering (AML) motivation: each setting uses a single baseline, the financial "gangs" are defined by the authors as connected components of known-illicit wallets (label-derived proxies, as the authors acknowledge) rather than drawn from real laundering schemes, and several ablations are missing. The paper would be considerably stronger with stronger baselines, additional ablations, and evaluation in more realistic settings, as detailed below.

Critical strengths
  1. A novel and well-motivated formulation. Treating the group, rather than the node, as the unit of prediction addresses a real limitation of node-classification pipelines, which output flagged nodes but leave group recovery unsolved. Learning the coarsening target from example groups is an original mechanism that turns graph coarsening into a detector. It is also inductive and designed to scale: it needs only sparse filter operations and no eigendecomposition of the graph Laplacian, and it is trained in closed form.

  2. "Capture" and "Confusability" give a clear account of when a group survives coarsening as one supernode, and the authors connect them to coarsening distortion, block purity, a lower bound on contraction error, and a sufficient condition for exact recovery under greedy Ward merging. I checked the main proofs (Appendices A–D), including numerical tests on random graphs, and found them correct. This theoretical grounding distinguishes the work from purely heuristic group-detection methods and is the strongest part of the submission.

Critical weaknesses
  1. Insufficient empirical validation. Each setting is compared against a single baseline: a static spectral target (synthetic), GNN + connected components (Elliptic++), and Louvain (Amazon, DBLP, email-Eu-core). On DBLP, Louvain is slightly better (group F1 0.694 vs 0.685). Related methods cited in the paper (ProCom, SEAL, CLARE, RevTrack) are not compared. No ablation compares the learned target with the fixed operator (L + \tau I)^{-1}, which Fig. 1 depicts the learned filter as approximating. The only financial experiment uses "gangs" that the authors define as connected components of illicit wallets, rather than real laundering schemes, and the results are modest (matched-group F1 0.206 on held-out time steps). A more realistic benchmark exists: Elliptic2, which the paper cites, provides subgraphs labeled as suspicious of money laundering, but it is not used for evaluation. This most strongly drives my recommendation: it is unclear whether the method improves in practice on existing approaches, which the paper does not compare against.

  2. A gap between the setting and the AML motivation. The method assumes that target groups are disjoint and connected, whereas real laundering schemes may share accounts or avoid direct links. Even in the paper’s own Amazon experiment, about 40% of the communities overlap (share nodes with another community), which the method cannot represent. The paper also does not report whether the exact-recovery condition of Thm. 11 (U_S < L_S) holds for any evaluated group, so it is unclear whether the guarantee applies in practice. In addition, the experiments train with a closed-form surrogate rather than the objective J_edge (Eq. 16) that the theory motivates. As a result, the theoretical results are not yet shown to explain or guarantee the empirical performance, which weakens the paper’s central claim of principled group detection for applications such as AML.

Questions for authors
  1. Comparison with related methods. Can you compare against methods that learn from example groups or communities (e.g., ProCom, SEAL, CLARE) and, where applicable, RevTrack? If any of them cannot be applied to this setting, please explain why.

  2. Realistic AML evaluation. Can you evaluate detection on Elliptic2, e.g., by hiding the labeled subgraphs, running the coarsener on the background graph, and measuring recovery of the suspicious subgraphs? If the full graph (about 49M nodes, 196M edges) is too large, the method could still be applied to a portion of it, such as a subgraph around the labeled subgraphs. If scale is the obstacle, please state this as a limitation.

  3. Scalability. The paper describes the method as scalable to large graphs, but reports no runtime or memory usage, gives a cost analysis only for the target construction (not for the deflated Ward coarsening), and evaluates graphs of at most about 335K nodes. Could you report runtime and memory for each experiment and the complexity of the coarsening step?

  4. Is learning necessary? Fig. 1 depicts the learned filter as approximately g_{\theta}(\lambda) \approx 1/(\lambda + \tau), which corresponds to the fixed operator (L + \tau I)^{-1} and requires no training. If the learned filters are indeed close to this form, a fixed target R = span{(L + \tau I)^{-1} X} might perform similarly, which would weaken the claim that learning the target is what drives the gains. Could you (i) compare the learned target with this fixed target under the same coarsening procedure, and (ii) plot the actual learned filter responses g_{\theta}(\lambda) for each dataset, to clarify whether Fig. 1 reflects the learned filters or is only illustrative?

  5. Does the theory apply in practice? For how many evaluated groups does the exact-recovery condition of Thm. 11 (U_S < L_S) hold? Reporting the empirical capture and confusability values would also help.

  6. Surrogate vs. J_edge. How close is the closed-form surrogate solution to the optimum of J_edge (Eq. 16), e.g., compared with direct gradient-based optimization?

  7. Generalization to unseen group types. How does the method perform on group structures or sizes not seen during training (e.g., training on cycles and testing on stars, or on larger groups)?

  8. Some notation is inconsistent, e.g., \bar{\mu}^{\tau} at line 203 vs. \bar{\mu}_{\tau} in Theorem 11 (line 409) and Eq. 23. Please double-check the notation throughout.

7 comments

We thank the reviewer for the valuable and constructive feedback. In response, we substantially expanded the empirical evaluation (Appendices E.6--E.11): comparisons with learned baselines on community and Elliptic++ benchmarks, an overlapping-group cut, runtime and memory measurements, ablations of the learned target, and measurements connecting our results to the theoretical guarantees. These new results further demonstrate the performance of our method and connect to the empirical findings to our theory. Our point-by-point responses follow. All references to figures, appendices, etc., refer to the updated version of the paper.

W1

We substantially expanded the empirical evaluation (Appendices E.6--E.11). The main results, detailed in our answers to Q1--Q6, are summarized below.

Setting and benchmarks. Learning to detect groups from labeled example groups and transferring to unseen graphs is a new problem setting, and to our knowledge there are very few dedicated methods and no established benchmarks for it. The closest existing methods come from community detection and search (ProCom, CLARE, SEAL) and from financial group detection (RevTrack), and we compare against all of them. Because no ready-made benchmark exists, we construct our evaluation from the strongest available sources: the community benchmarks used by ProCom (Amazon, DBLP, LiveJournal) with fixed train/test community splits, Elliptic++ with held-out time steps, and synthetic graphs where the ground-truth groups and their shapes and sizes are controlled.

Comparison with learned baselines (Q1). On Amazon, DBLP, and LiveJournal, we compared against ProCom, SEAL, and CLARE using identical community splits and training budgets (Appendix E.6). Our learned cut achieves higher average precision than all three on every dataset at both budgets, and with 100 training communities improved group F1 from 0.460 to 0.716 on DBLP and from 0.451 to 0.746 on LiveJournal. On the pruned DBLP benchmark with 50% training communities, it also reaches 0.727 group F1 versus 0.694 for Louvain. On Elliptic++, we added ProCom, CLARE, SEAL, and RevTrack's RevClassify (Appendix E.7). Our method obtains the highest B$^3$ F1 (0.416) and matched-group F1 (0.206), versus 0.204 and 0.098 for the strongest competing baseline (GNN + connected components).

Static spectral baseline. We use the rank-matched static spectral target on the synthetic graphs, where it is a controlled comparison under the same coarsener. It does not apply to Elliptic++ or the community benchmarks. It requires an eigendecomposition of every graph, which is computationally infeasible at these scales (up to 335K nodes, and 15 daily graphs for Elliptic++). It is also transductive: the target is computed from the spectrum of each graph itself and cannot be fitted on training groups and transferred to unseen days or held-out communities. On these datasets we therefore compare against the methods above, which can be trained on example groups.

Is learning necessary? (Q4). Under the same coarsening procedure and cut-selection rule, our learned filters outperform random filters, a fixed resolvent (L+τI)−1(L+\tau I)^{-1}, and a width-matched bank of resolvent powers (Appendix E.9). On synthetic graphs, group F1 was 0.738, versus 0.251 for random filters, 0.320 for the resolvent, and 0.362 for resolvent powers. On ProCom Amazon, it was 0.875, versus 0.778 for random filters and 0.857 for resolvent powers. Figure 1 is now purely illustrative. Figure 9 plots the measured learned responses, which differ from the resolvent.

Target quality and training (Q5, Q6). The learned targets have the highest test capture on both benchmarks, and about fivefold lower mean confusability than resolvent powers on synthetic graphs (Table 11). The closed-form training solution reaches 90.1% (synthetic) and 87.8% (Amazon) of the best exact objective found by direct Adam optimization, at roughly 0.5 s versus 280 s and 4.4 s versus 463 s, and further refinement gave no consistent gain in recovery (Appendix E.10).

Scalability (Q3). On full SNAP Amazon and DBLP (about 335K and 317K nodes), the full pipeline runs in 10.5 and 14.5 minutes with 3.1 and 3.7 GB peak memory on a laptop CPU (Appendix E.8).

Elliptic++ and Elliptic2 (Q2). Since no benchmark with labeled laundering groups for this learning setting is established, we evaluate on Elliptic++ by defining gangs from wallet labels, with training on early time steps and testing on held-out ones (Appendices E.4 and E.7). Elliptic2 is the natural next benchmark, and we are running experiments on it to measure recovery of its labeled suspicious subgraphs. We will update the paper once the results are ready.

W2

Overlapping groups. Disjointness is not fundamental to our method. Disjointness enters only through the cut rule that selects blocks from the hierarchy. In Appendix E.6, we therefore designed a nested cut rule for overlapping groups, which allows up to kk selected blocks along each root-to-leaf path, so the same hierarchy can return overlapping groups. On Amazon, DBLP, and LiveJournal, whose ground-truth communities overlap, the nested cut improves group F1 over the disjoint cut in all six settings (by 0.016--0.025) and has the highest group F1 of all compared methods in all six. With 100 training communities, it reaches 0.923, 0.741, and 0.763 (Amazon, DBLP, LiveJournal), versus 0.904, 0.460, and 0.451 for the strongest learned baseline (ProCom). On Amazon, it has higher precision than ProCom (0.908 vs.\ 0.867) at equal recall (0.960) with 2,507 output groups instead of 5,000, and on DBLP its detection rate is 0.687 versus 0.375 for ProCom.

Theory and practice (Q5). As detailed in our answer to Q5, we measured the quantities that Theorem 11 is built on, capture and confusability, for every group (Appendix E.9, Table 11). The learned targets raise test capture from 0.025 to 0.053 on synthetic graphs and from 0.028 to 0.071 on Amazon, compared with width-matched resolvent powers. On synthetic graphs, they also reduce mean confusability about fivefold. These are the properties the theory identifies as necessary for coarsening to return a group as a supernode, and the learned targets improve both. The merge-cost ordering the theory requires is also visible empirically (Appendix E.3): on a held-out synthetic graph, every planted group has a lower mean internal than boundary edge cost, with edge-level AUC 0.971 and raw Ward-score AUC 0.919. Theorem 11 is a sufficient condition for exact recovery, and we will report the fraction of evaluated groups satisfying US<LSU_S<L_S in the revised paper.

Surrogate versus JedgeJ_{\mathrm{edge}} (Q6). The recovery analysis is stated for any target subspace RΘ\mathcal{R}_\Theta, so it applies to the subspace produced by the closed-form surrogate. As detailed in our answer to Q6, we compared the surrogate with direct Adam optimization of the exact objective (Appendix E.10). The closed-form solution reaches 90.1% (synthetic) and 87.8% (Amazon) of the best exact objective found, at 0.5,s versus 280,s and 4.4,s versus 463,s. Further exact-objective refinement did not consistently improve recovery.

Q1

We added comparisons with ProCom, SEAL, and CLARE on Amazon, DBLP, and LiveJournal using the same community splits and training budgets (Appendix E.6, Tables 5--6). Our learned cut achieves higher average precision than all three baselines on every dataset at both training budgets. With 100 training communities, it improves group F1 from 0.460 to 0.716 on DBLP and from 0.451 to 0.746 on LiveJournal over the strongest learned baseline on each dataset. On Amazon, it matches ProCom's group F1 (0.902 vs.\ 0.904) while achieving much higher average precision than the strongest learned baseline (0.810 vs.\ 0.500), returning 1,147 groups compared with 5,000 for ProCom.

We also extended the Elliptic++ comparison to ProCom, CLARE, SEAL, and RevTrack's RevClassify (Appendix E.7, Tables 7--8), using the same temporal split (fit on days 26--30, test on days 31--40). RevTrack's discovery component, RevFilter, requires candidate sender and receiver sets that are not available for unseen target gangs, so we evaluate RevClassify on candidate groups generated by GNN + connected components. Our method achieves the highest B$^3$ F1 and matched-group F1 among all methods: 0.416 and 0.206, compared with 0.198 and 0.048 for RevClassify, 0.093 and 0.076 for ProCom, 0.065 and 0.064 for CLARE, and 0.046 and 0.062 for SEAL. It also has the highest B$^3$ F1 on seven of the ten test days.

Q2

We are running experiments on Elliptic2 to measure how well our method recovers held-out suspicious subgraphs. We will update the paper once the results are ready.

Q3

Our pipeline is computationally light. It has three simple components, each with a explicit cost analysis in Appendix E.8: (i) sparse Chebyshev filtering, which needs no eigendecomposition and no dense N×NN\times N operator, at cost O(HKd∣E∣)O(HKd|E|) per graph; (ii) training through a closed-form solution of small generalized eigenproblems, which takes about 0.5,s on the synthetic graphs and 4.4,s on Amazon (Appendix E.10); and (iii) a greedy Ward coarsener with local updates and a priority queue. The entire pipeline runs on a laptop CPU, without a GPU.

Runtime and memory. Table 9 reports the runtime and peak memory of every experiment, together with the time spent in each stage, on an Apple M2 Pro with 16,GB of memory. The global-cut pipeline take 10.5 min with 3.11,GB peak memory on full SNAP Amazon (334,863 nodes) and 14.5 min with 3.71,GB on full SNAP DBLP (317,080 nodes). The three ProCom benchmarks take 2.9 min (Amazon), 6.9 min (DBLP), and 9.9 min (LiveJournal), using 0.78--2.08,GB. The eight synthetic graphs take 56.1,s with 0.57,GB. The 15 Elliptic++ days took 32.6 min with 5.70,GB, even with a wide target of width q=1840q=1840.

Complexity of the coarsening step. The runtime of connectivity-constrained deflated Ward coarsening is

O!(∣E∣q+Nq2+q3+∑t=1nm ⁣[Rt(B3+Bq)+q2+δtq]+[∣E∣+nmF+∑t=1nmRt]log⁡Q).O\\!\left( |E|q+Nq^2+q^3 +\sum_{t=1}^{n_m}\!\left[R_t(B^3+Bq)+q^2+\delta_tq\right] +\left[|E|+n_mF+\sum_{t=1}^{n_m}R_t\right]\log Q \right).

Here qq is the target width, nmn_m is the number of constrained merges, RtR_t counts rescored candidates, and δt\delta_t sums the degrees of the two merged blocks and their replacement. The implementation caps the local neighborhood at B=32B=32, and new queue entries per merge at F=32F=32, with Q=∣E∣+nmF+1Q=|E|+n_mF+1, so each merge involves only a small local problem. The bound is near-linear up to queue costs when qq, average rescoring, and average coarse degree stay bounded. Storage is O(Nq+∣E∣+q2+B2)O(Nq+|E|+q^2+B^2), with no dense matrix over graph nodes.

Empirical scaling. On 10K--100K-node YouTube subgraphs, target construction and coarsening times scale approximately linearly, with fitted log--log slopes of 0.94 and 1.07, respectively. The 100K-node subgraph runs in 13.6 min with 1.64,GB.

Q4

Yes, learning is what drives the gains. Under the same coarsening procedure, our learned filters outperform every fixed target we tested, including the resolvent (L+τI)−1X(L+\tau I)^{-1}X the reviewer proposes, and on the synthetic graphs they do so by a large margin.

(i) Learned versus fixed targets. We compared the learned filters with the fixed resolvent (L+τI)−1X(L+\tau I)^{-1}X, a width-matched bank of resolvent powers, untrained random Chebyshev filters, a multi-scale low-pass bank, and the first HH Chebyshev atoms. All used the same input features, coarsener, and training-based cut-selection rule (Appendix E.9, Table 10). On the synthetic graphs, learning more than doubles group F1: it reaches 0.738, versus 0.251 for random filters, 0.320 for the single resolvent, and 0.362 for the strongest fixed control, the resolvent powers. The gap is even larger in detection rate: the learned target detects 75.6% of test groups, versus 3.3% for resolvent powers and at most 1.1% for the single resolvent and every other fixed control. Learning is also the best on ProCom Amazon, with 0.875 group F1 and a 96.4% detection rate, versus 0.857 and 93.2% for resolvent powers, 0.831 and 90.7% for the single resolvent, and 0.778 and 77.3% for random filters. Learning also raises the exact training objective from 1.33 to 3.56 on synthetic graphs and from 1.79 to 13.56 on Amazon, compared with resolvent powers. Capture and confusability, reported in Table 11 and discussed in our answer to Q5, explain these gains: the learned targets retain more of the group direction while suppressing internal variation. A fixed target without training therefore does not reproduce the performance of the learned one.

(ii) Learned filter responses. Figure 1 was illustrative, and we removed the ≈1/(λ+τ)\approx 1/(\lambda+\tau) label in the revised version. Figure 9 plots the measured learned responses for the synthetic and Amazon runs alongside the fixed resolvent. The learned heads respond in different frequency bands. On synthetic graphs, they separate the zero-eigenvalue component from narrow bands of low nonzero frequencies. On Amazon, most of them concentrate near the bottom of the spectrum, and some also respond at high frequencies. To compare the targets independently of the choice of basis, we computed the mean squared cosines of the principal angles between the learned and resolvent subspaces under the MτM_\tau inner product. This overlap is only 0.087 on synthetic graphs and 0.203 on Amazon, so the learned target is a different subspace from the fixed resolvent, and this difference is what produces the gains in (i).}

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.