pith. sign in

arxiv: 1906.09602 · v1 · pith:OD2HGJQFnew · submitted 2019-06-23 · 💻 cs.LG · stat.ML

Ego-CNN: Distributed, Egocentric Representations of Graphs for Detecting Critical Structures

Ruo-Chun Tzeng , Shan-Hung Wu This is my paper

Pith reviewed 2026-05-25 17:39 UTC · model grok-4.3

classification 💻 cs.LG stat.ML
keywords graph embeddingego-convolutionscritical structuresgraph neural networkssocial networksscale-free priorsCNN visualization
0
0 comments X

The pith

Ego-CNNs detect task-specific critical structures in graphs by stacking ego-convolutions in an egocentric manner.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces Ego-CNN as a graph embedding model that applies ego-convolutions at each layer and stacks them ego-centrically to detect precise critical structures at the global scale. This allows the model to be jointly trained with a task model, providing explanations and discoveries for the task through visualization. Sympathetic readers would care because it addresses the inability of existing models to precisely detect task-specific structures while maintaining performance and efficiency. The approach also incorporates scale-free priors for better training on social network data. Results indicate comparable performance to state-of-the-art models with added interpretability benefits.

Core claim

Ego-CNN employs the ego-convolutions at each layer and stacks up layers using an ego-centric way to detects precise critical structures efficiently. An Ego-CNN can be jointly trained with a task model and help explain or discover knowledge for the task.

What carries the argument

Ego-convolutions applied at each layer and stacked in an ego-centric manner, which enables precise detection of task-specific critical structures at the global scale.

If this is right

  • Ego-CNNs achieve comparable task performance to state-of-the-art graph embedding models.
  • The model works with CNN visualization techniques to illustrate the detected structures.
  • Training efficiency improves when incorporating scale-free priors common in social network datasets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Egocentric stacking could support distributed computation on very large graphs.
  • The visualization capability might transfer to knowledge discovery in non-social graph domains such as molecular structures.
  • Joint training with downstream tasks could improve robustness when graphs change over time.

Load-bearing premise

That ego-convolutions applied at each layer and stacked in an ego-centric manner overcome the limitation of existing models and enable precise detection of task-specific critical structures at the global scale.

What would settle it

An experiment on a benchmark graph dataset with known task-specific structures where Ego-CNN fails to highlight those structures more precisely than competing embedding models.

Figures

Figures reproduced from arXiv: 1906.09602 by Ruo-Chun Tzeng, Shan-Hung Wu.

Figure 2
Figure 2. Figure 2: Neighborhood of a node n in (a) Message-Passing NNs (Duvenaud et al., 2015; Li et al., 2016; Pham et al., 2017; Gilmer et al., 2017): g (l) n ∈ R D the aggregated hidden representations of adjacent nodes in the previous layer; (b) Patchy-San (Niepert et al., 2016): A(n) ∈ R K×K the adjacency matrix of K nearest neighbors of node n, and (c) Ego-CNNs: E (n,l) ∈ R (K+1)×D the hidden representation of the l-ho… view at source ↗
Figure 3
Figure 3. Figure 3: The model architecture of an Ego-CNN. With our egocentric design, neighborhoods are egocentrically enlarged by 1-hop after each Ego-Convolution layer. The dashed horizontal lines across layers indicate neighborhoods of the same node; the ? mark indicates an arbitrary dimension. the filters W(d) ∈ R N×N hard to learn. Efficient detection of task-specific, precise critical structures at global scale remains … view at source ↗
Figure 4
Figure 4. Figure 4: The receptive field of a neuron in an Ego-CNN effec￾tively enlarges at a deeper layer. (a)-(c) Receptive fields of neurons at the 1st, 2nd, and 5th layer corresponding to the same node. (d) Receptive field of another neuron at the 5th layer that partially cov￾ers the graph. The difference in the coverage reflects the position of the corresponding node [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The degree distribution of Reddit dataset follows the power-law distribution. each node. In fact, Ego-CNN can take any kind of node embeddings as input, as shown in the left of [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of critical structures on (a)(b) two Alcohol compounds for the task distinguishing Alcohol from Alkane, and (c) a Symmetric Isomer and (d) an Asymmetric Isomer compounds for the task classifying the types of Isomer. Critical structures are colored in grey and the node/edge size is proportional to its impor￾tance. The OH-base on Alcohols is always captured precisely and considered critical. On… view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of critical structures on Reddit dataset. The critical structures are colored in grey and the node/edge size is proportional to its importance. The results show that the variety of different opinions are the key to discriminant discussion-based threads from QA-based threads [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Filters and neighborhood in (a) Spatial GCN (Bruna et al., 2013). A filter W is a sparse matrix, which is not aimed to detect local patterns, but to learn the connectivity of clusters. (b) DCNN (Atwood & Towsley, 2016) scans through the M × N diffusion matrix of each node. (c) Patchy-San (Niepert et al., 2016) scans the K × K adjacency matrix of local neighborhood of each node. (d) Neural Fingerprints (Duv… view at source ↗
Figure 9
Figure 9. Figure 9: More experiments. paper. As we can see in [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
read the original abstract

We study the problem of detecting critical structures using a graph embedding model. Existing graph embedding models lack the ability to precisely detect critical structures that are specific to a task at the global scale. In this paper, we propose a novel graph embedding model, called the Ego-CNNs, that employs the ego-convolutions convolutions at each layer and stacks up layers using an ego-centric way to detects precise critical structures efficiently. An Ego-CNN can be jointly trained with a task model and help explain/discover knowledge for the task. We conduct extensive experiments and the results show that Ego-CNNs (1) can lead to comparable task performance as the state-of-the-art graph embedding models, (2) works nicely with CNN visualization techniques to illustrate the detected structures, and (3) is efficient and can incorporate with scale-free priors, which commonly occurs in social network datasets, to further improve the training efficiency.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes Ego-CNN, a graph embedding model that applies ego-convolutions at each layer and stacks them in an ego-centric manner to detect task-specific critical structures at global scale. It claims the model can be jointly trained with a downstream task, yields comparable performance to SOTA graph embeddings, integrates with CNN visualization for interpretability, and gains efficiency from scale-free priors common in social networks.

Significance. If the architecture truly enables global-scale, task-specific structure detection with the claimed efficiency and interpretability, the work would address a recognized gap in graph representation learning. The combination of local ego-operations with visualization and scale-free priors is a potentially useful direction for social-network applications, though the absence of any experimental details in the provided text leaves the practical impact unassessable.

major comments (2)
  1. [Abstract / model architecture] Abstract (and implied model description): the central claim that ego-convolutions stacked ego-centrically overcome prior limitations and enable precise critical-structure detection at global scale lacks an explicit long-range mechanism. Ego-networks are definitionally local (node plus k-hop neighborhood); repeated local operations without described global aggregation, cross-ego message passing, or whole-graph readout leave the receptive field local, directly undermining claims (1) and (2).
  2. [Abstract] Abstract: the three experimental claims ((1) comparable task performance, (2) compatibility with CNN visualization, (3) efficiency gains with scale-free priors) are asserted without any reference to datasets, baselines, metrics, or experimental design, rendering the claims unverifiable from the supplied text.
minor comments (2)
  1. [Abstract] Abstract contains a duplicated phrase: 'employs the ego-convolutions convolutions'.
  2. [Abstract] Abstract: 'to detects precise' should be 'to detect precise'.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract and model description. We address each major comment below and will make targeted revisions to strengthen clarity without altering the core contributions.

read point-by-point responses

  1. Referee: [Abstract / model architecture] Abstract (and implied model description): the central claim that ego-convolutions stacked ego-centrically overcome prior limitations and enable precise critical-structure detection at global scale lacks an explicit long-range mechanism. Ego-networks are definitionally local (node plus k-hop neighborhood); repeated local operations without described global aggregation, cross-ego message passing, or whole-graph readout leave the receptive field local, directly undermining claims (1) and (2).

    Authors: The manuscript's model section describes how the distributed egocentric stacking integrates local ego-convolutions across layers to build representations that capture task-specific structures at global scale, leveraging the graph's overall connectivity rather than requiring a single global aggregation step. This is achieved through the ego-centric layering that propagates information across overlapping ego-networks. We agree the abstract is too concise on this point and will revise it to explicitly note the mechanism by which stacking enables global receptive fields. revision: yes

  2. Referee: [Abstract] Abstract: the three experimental claims ((1) comparable task performance, (2) compatibility with CNN visualization, (3) efficiency gains with scale-free priors) are asserted without any reference to datasets, baselines, metrics, or experimental design, rendering the claims unverifiable from the supplied text.

    Authors: The full manuscript contains the experimental details in the dedicated evaluation section, including datasets, baselines, and metrics. We acknowledge that the abstract alone does not reference them and will revise the abstract to include a brief statement indicating that the claims are supported by experiments on standard graph datasets with comparisons to state-of-the-art models. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical validation of proposed architecture

full rationale

The paper introduces Ego-CNN via a novel stacking of ego-convolutions and validates performance, visualization utility, and efficiency through direct experimental comparisons to baselines on real datasets. No equations, uniqueness theorems, or first-principles derivations are offered that reduce by construction to fitted parameters, self-citations, or renamed inputs; the central claims rest on reported task metrics rather than internal definitional loops.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies insufficient detail to identify free parameters, axioms, or invented entities.

pith-pipeline@v0.9.0 · 5688 in / 1033 out tokens · 27002 ms · 2026-05-25T17:39:51.547013+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 28 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

    work page

  2. [2]

    and Towsley, D

    Atwood, J. and Towsley, D. Diffusion-convolutional neural networks. In Proceedings of NIPS, 2016

    work page 2016

  3. [3]

    Spectral networks and locally connected networks on graphs

    Bruna, J., Zaremba, W., Szlam, A., and LeCun, Y. Spectral networks and locally connected networks on graphs. In Proceedings of ICLR, 2013

    work page 2013

  4. [4]

    W., and Chang, K

    Cai, H., Zheng, V. W., and Chang, K. A comprehensive survey of graph embedding: problems, techniques and applications. IEEE Transactions on Knowledge and Data Engineering, 2018

    work page 2018

  5. [5]

    Cook, S. A. The complexity of theorem-proving procedures. In Proceedings of the third annual ACM symposium on Theory of Computing. ACM, 1971

    work page 1971

  6. [6]

    Discriminative embeddings of latent variable models for structured data

    Dai, H., Dai, B., and Song, L. Discriminative embeddings of latent variable models for structured data. In Proceedings of ICML, 2016

    work page 2016

  7. [7]

    Convolutional neural networks on graphs with fast localized spectral filtering

    Defferrard, M., Bresson, X., and Vandergheynst, P. Convolutional neural networks on graphs with fast localized spectral filtering. In Proceedings of NIPS, 2016

    work page 2016

  8. [8]

    K., Maclaurin, D., Iparraguirre, J., Bombarell, R., Hirzel, T., Aspuru-Guzik, A., and Adams, R

    Duvenaud, D. K., Maclaurin, D., Iparraguirre, J., Bombarell, R., Hirzel, T., Aspuru-Guzik, A., and Adams, R. P. Convolutional networks on graphs for learning molecular fingerprints. In Proceedings of NIPS, 2015

    work page 2015

  9. [9]

    S., Riley, P

    Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and Dahl, G. E. Neural message passing for quantum chemistry. In Proceedings of ICML, 2017

    work page 2017

  10. [10]

    A model of saliency-based visual attention for rapid scene analysis

    Itti, L., Koch, C., and Niebur, E. A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20 0 (11): 0 1254--1259, 1998

    work page 1998

  11. [11]

    M., Morris, C., Mutzel, P., and Neumann, M

    Kersting, K., Kriege, N. M., Morris, C., Mutzel, P., and Neumann, M. Benchmark data sets for graph kernels, 2016. URL http://graphkernels.cs.tu-dortmund.de

    work page 2016

  12. [12]

    A box-covering algorithm for fractal scaling in scale-free networks

    Kim, J., Goh, K.-I., Kahng, B., and Kim, D. A box-covering algorithm for fractal scaling in scale-free networks. Chaos: An Interdisciplinary Journal of Nonlinear Science, 2007

    work page 2007

  13. [13]

    Kipf, T. N. and Welling, M. Semi-supervised classification with graph convolutional networks. In Proceedings of ICLR, 2017

    work page 2017

  14. [14]

    and Pan, H

    Kondor, R. and Pan, H. The multiscale laplacian graph kernel. In Proceedings of NIPS, 2016

    work page 2016

  15. [15]

    Kronecker graphs: An approach to modeling networks

    Leskovec, J., Chakrabarti, D., Kleinberg, J., Faloutsos, C., and Ghahramani, Z. Kronecker graphs: An approach to modeling networks. Journal of Machine Learning Research, 11 0 (Feb): 0 985--1042, 2010

    work page 2010

  16. [16]

    C., and Willinger, W

    Li, L., Alderson, D., Doyle, J. C., and Willinger, W. Towards a theory of scale-free graphs: Definition, properties, and implications. Internet Mathematics, pp.\ 431--523, 2005

    work page 2005

  17. [17]

    Gated graph sequence neural networks

    Li, Y., Tarlow, D., Brockschmidt, M., and Zemel, R. Gated graph sequence neural networks. In Proceedings of ICLR, 2016

    work page 2016

  18. [18]

    Efficient estimation of word representations in vector space

    Mikolov, T., Chen, K., Corrado, G., and Dean, J. Efficient estimation of word representations in vector space. 2013

    work page 2013

  19. [19]

    subgraph2vec: Learning distributed representations of rooted sub-graphs from large graphs

    Narayanan, A., Chandramohan, M., Chen, L., Liu, Y., and Saminathan, S. subgraph2vec: Learning distributed representations of rooted sub-graphs from large graphs. In Workshop on Mining and Learning with Graphs, 2016

    work page 2016

  20. [20]

    Learning convolutional neural networks for graphs

    Niepert, M., Ahmed, M., and Kutzkov, K. Learning convolutional neural networks for graphs. In Proceedings of ICML, 2016

    work page 2016

  21. [21]

    Q., and Venkatesh, S

    Pham, T., Tran, T., Phung, D. Q., and Venkatesh, S. Column networks for collective classification. In Proceedings of AAAI, 2017

    work page 2017

  22. [22]

    Shervashidze, N., Schweitzer, P., Leeuwen, E. J. v., Mehlhorn, K., and Borgwardt, K. M. Weisfeiler-lehman graph kernels. JMLR, 12 0 (Sep): 0 2539--2561, 2011

    work page 2011

  23. [23]

    Graph attention networks

    Velickovic, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., and Bengio, Y. Graph attention networks. In Proceedings of ICLR, 2018

    work page 2018

  24. [24]

    Watts, D. J. and Strogatz, S. H. Collective dynamics of small-worldnetworks. Nature, 393 0 (6684): 0 440, 1998

    work page 1998

  25. [25]

    and Lehman, A

    Weisfeiler, B. and Lehman, A. A reduction of a graph to a canonical form and an algebra arising during this reduction. Nauchno-Technicheskaya Informatsia, 2 0 (9): 0 12--16, 1968

    work page 1968

  26. [26]

    and Vishwanathan, S

    Yanardag, P. and Vishwanathan, S. Deep graph kernels. In Proceedings of SIGKDD. ACM, 2015

    work page 2015

  27. [27]

    Hierarchical graph representation learning with differentiable pooling

    Ying, Z., You, J., Morris, C., Ren, X., Hamilton, W., and Leskovec, J. Hierarchical graph representation learning with differentiable pooling. In Proceedings of NIPS, pp.\ 4805--4815, 2018

    work page 2018

  28. [28]

    D., Taylor, G

    Zeiler, M. D., Taylor, G. W., and Fergus, R. Adaptive deconvolutional networks for mid and high level feature learning. In Proceedings of ICCV. IEEE, 2011

    work page 2011