Diffusion-Convolutional Neural Networks

James AtwoodDon Towsley

article2015NeurIPS1,387 citations

Introduces Diffusion-Convolutional Neural Networks, an efficient framework that models information diffusion across graph structures to learn isomorphism-invariant representations and outperform traditional relational models on node classification tasks.

Listen

Real-world data increasingly takes the form of complex networks and relational graphs, such as citation indices, social networks, and molecular structures. While exploiting graph structure can significantly improve predictive accuracy, traditional machine learning methods often face a difficult trade-off between predictive power and computational tractability. Existing relational models frequently suffer from exponential computational complexity during training and inference, while standard deep learning architectures are primarily designed for regular, grid-like inputs such as images.

The article demonstrates and evaluates Diffusion-Convolutional Neural Networks (DCNNs), a new neural framework for graph-structured data. The main objective is to establish whether scanning information diffusion processes across graph neighborhoods can capture relational patterns effectively without incurring the prohibitive computational overhead of traditional probabilistic relational models.

To evaluate this framework, the authors conducted controlled experiments across standard benchmark datasets. They tested node-level categorization on citation networks (Cora and Pubmed) containing thousands of documents and links, and whole-graph classification across chemical and biological collections (including NCI1, NCI109, MUTAG, PTC, and ENZYMES). The approach compares DCNN performance against established baselines, such as regularized logistic regression, graph kernels, and partially-observed conditional random fields.

The findings show that DCNNs achieve superior performance on node-level classification. On citation networks, a two-hop diffusion architecture achieved statistically significant gains over all baseline methods, reaching approximately 86.8% accuracy on Cora and 89.8% on Pubmed. Second, most of the performance gains were realized quickly within a local neighborhood search depth of two to three hops before leveling off. Third, DCNNs achieved these accuracy gains while maintaining polynomial-time computational complexity, executing substantially faster than exponential probabilistic alternatives. However, for whole-graph classification tasks, DCNNs did not consistently outperform specialized graph kernel methods, demonstrating that simply averaging node-level diffusion states is insufficient for summarizing entire graphs.

These results demonstrate that DCNNs provide an efficient, high-performance solution for entity-level prediction tasks in relational networks, allowing organizations to leverage GPU acceleration without expensive inference routines. For decision-makers, DCNNs are recommended for node-level relational problems such as document tagging, entity labeling, or user categorization. When dealing with whole-graph categorization, such as molecular property prediction, established graph kernel methods remain preferable until more sophisticated aggregation mechanisms are developed.

Confidence in node classification is high given the statistically validated benchmarks. However, key operational limitations remain. Because the framework represents diffusion states as dense tensors, GPU memory requirements scale quadratically with node count, making the current implementation practical for networks with tens to hundreds of thousands of nodes, but unsuitable for massive graphs with millions or billions of nodes without further optimization.

University of Massachusetts Amherst classification convolutional neural networks gpu Diffusion-Convolutional Neural Networks graph diffusion node classification

arXiv: 1511.02136

Paper: Spectral Networks and Locally Connected Networks on Graphs, Joan Bruna et al. (2014). This foundational paper establishes spectral and spatial neural network formulations on graphs, providing essential background for understanding how diffusion processes adapt convolution to non-Euclidean domains.
Paper: The Graph Neural Network Model, Franco Scarselli et al. (2009). It introduces the foundational framework of processing graph-structured data through iterative information diffusion mechanisms across neighboring nodes.
Paper: Learning with Local and Global Consistency, Dengyong Zhou et al. (2003). It formalizes label propagation and consistency via iterative graph diffusion operators, which directly underpins diffusion-based node classification in graphs.
Paper: Discrete Signal Processing on Graphs, Aliaksei Sandryhaila et al. (2012). It defines shift operators and filtering on graph adjacency structures, providing fundamental discrete signal processing concepts used to model graph convolutions.
Paper: Deep Convolutional Networks on Graph-Structured Data, Mikael Henaff et al. (2015). It explores extending spectral graph convolutions to unstructured data, establishing the groundwork for GPU-efficient graph-based deep learning.

Paper: Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting, Yaguang Li et al. (2017). This work extends diffusion convolution operators to directed spatiotemporal graphs by integrating bidirectional random walk diffusions into recurrent architectures for traffic forecasting.
Paper: Graph WaveNet for Deep Spatial-Temporal Graph Modeling, Zonghan Wu et al. (2019). It advances diffusion-based spatial modeling by learning self-adaptive graph adjacency matrices combined with temporal convolutions for dynamic graph representation.
Paper: Semi-Supervised Classification with Graph Convolutional Networks, Thomas N. Kipf et al. (2017). It simplifies spectral and localized convolutions into a highly efficient first-order approximation that became the standard baseline following early spatial models like DCNNs.
Paper: Contrastive Multi-View Representation Learning on Graphs, Kaveh Hassani et al. (2020). It incorporates generalized graph diffusion processes to construct multi-view contrastive representations for self-supervised node and graph learning.
Paper: An End-to-End Deep Learning Architecture for Graph Classification, Muhan Zhang et al. (2018). It develops an end-to-end spatial graph convolution framework with SortPooling that provides isomorphism-invariant graph-level classification.
Paper: Geometric Deep Learning on Graphs and Manifolds Using Mixture Model CNNs, Federico Monti et al. (2017). It presents a unified spatial mixture model framework (MoNet) that generalizes previous localized and diffusion-like graph convolution architectures.
Paper: A Comprehensive Survey on Graph Neural Networks, Zonghan Wu et al. (2019). This comprehensive survey categorizes the broader landscape of spatial and spectral graph neural networks, contextualizing diffusion convolutions within the field.

Cover for Diffusion-Convolutional Neural Networks

Abstract

We present diffusion-convolutional neural networks (DCNNs), a new model for graph-structured data. Through the introduction of a diffusion-convolution operation, we show how diffusion-based representations can be learned from graph-structured data and used as an effective basis for node classification. DCNNs have several attractive qualities, including a latent representation for graphical data that is invariant under isomorphism, as well as polynomial-time prediction and learning that can be represented as tensor operations and efficiently implemented on the GPU. Through several experiments with real structured datasets, we demonstrate that DCNNs are able to outperform probabilistic relational models and kernel-on-graph methods at relational node classification tasks.

Topics

classification convolutional neural networks gpu relational data models classification models image classification

Keywords

Diffusion-Convolutional Neural Networks graph diffusion node classification deep graph kernels diffusion kernel semi-supervised learning belief propagation graphical models

1 Introduction
2 Model
3 Experiments
3.1 Node classification
3.2 Graph Classification
4 Limitations
5 Related Work
6 Conclusion and Future Work
7 Appendix: Representation Invariance for Isomorphic Graphs
References

Knowls

Knowl 1 — Diffusion-Convolution Operation for Node Classification
model/method
The diffusion-convolution operation defines a latent representation for graph nodes by mapping entity features through a graph diffusion process. Let $G = (V, E)$ be a graph with $N$ vertices, feature matrix $X \in \mathbb{R}^{N \times F}$ , and degree-normalized transition probability matrix $P \in \mathbb{R}^{N \times N}$ (where $P_{ij}$ denotes the one-step transition probability from node $i$ to node $j$ ). Let $P^* \in \mathbb{R}^{N \times H \times N}$ denote the 3D tensor containing the power series of $P$ across $H$ diffusion hops, where slice $j \in \{0, \dots, H-1\}$ corresponds to $P^j$ .

The diffusion-convolutional activation $Z_{ijk}$ for node $i$ , hop $j$ , and feature $k$ is given by: $Z_{ijk} = f\left(W^c_{jk} \sum_{l=1}^N P^*_{ijl} X_{lk}\right)$

In tensor notation: $Z = f\left(W^c \odot (P^* X)\right)$ where $W^c \in \mathbb{R}^{H \times F}$ is a learnable real-valued weight tensor, $\odot$ denotes element-wise multiplication broadcast across the node dimension, and $f$ is a differentiable activation function (such as $\tanh$ ). The model requires $\mathcal{O}(H \times F)$ parameters, making the parameter count independent of the number of nodes $N$ .

Node label predictions $\hat{Y}$ and class conditional probabilities $\mathbb{P}(Y \mid X)$ are obtained via a dense classification layer with weight tensor $W^d$ : $\hat{Y} = \arg\max f(W^d \odot Z), \qquad \mathbb{P}(Y \mid X) = \text{softmax}(f(W^d \odot Z))$
Knowl 2 — Graph-Level Classification with Diffusion-Convolutional Networks
model/method
Diffusion-Convolutional Neural Networks (DCNNs) extend to whole-graph classification by aggregating node diffusion activations across all vertices via mean pooling.

For an input graph $G_t = (V_t, E_t)$ with $N_t$ nodes, feature design matrix $X_t \in \mathbb{R}^{N_t \times F}$ , and transition matrix power series tensor $P^*_t \in \mathbb{R}^{N_t \times H \times N_t}$ , the graph-level latent representation $Z_t \in \mathbb{R}^{H \times F}$ is defined as: $Z_t = f\left(W^c \odot \left(\frac{1}{N_t} \mathbf{1}_{N_t}^T P^*_t X_t\right)\right)$ where $\mathbf{1}_{N_t}$ is an $N_t \times 1$ all-ones vector, $W^c \in \mathbb{R}^{H \times F}$ is the learnable weight tensor, $\odot$ is element-wise multiplication, and $f$ is a differentiable activation function.

Predictions $\hat{Y}_t$ or class posterior distributions $\mathbb{P}(Y_t \mid X_t)$ for the whole graph are then computed using a dense weight tensor $W^d \in \mathbb{R}^{H \times F}$ : $\hat{Y}_t = \arg\max f(W^d \odot Z_t), \qquad \mathbb{P}(Y_t \mid X_t) = \text{softmax}(f(W^d \odot Z_t))$
Knowl 3 — Edge Classification and Feature Incorporation via Adjacency Augmentation
model/method
To handle edge classification and incorporate edge features in a Diffusion-Convolutional Neural Network, each edge is transformed into an intermediate node connected to its head and tail vertices.

Given a graph with $N_t$ vertices, vertex adjacency matrix $A_t \in \mathbb{R}^{N_t \times N_t}$ , and vertex-edge incidence matrix $B_t \in \mathbb{R}^{M_t \times N_t}$ for $M_t$ edges, an augmented adjacency matrix $A'_t \in \mathbb{R}^{(N_t + M_t) \times (N_t + M_t)}$ is defined as: $A'_t = \begin{pmatrix} A_t & B_t^T \\ B_t & 0 \end{pmatrix}$

The degree-normalized transition probability matrix $P'_t$ and its power series tensor $(P'_t)^*$ are derived from $A'_t$ . Applying the diffusion-convolution operation over this augmented graph produces a latent activation tensor $Z_t \in \mathbb{R}^{M_t \times H \times F}$ for the edge nodes, enabling edge label prediction and joint representation learning over node and edge attributes.
Knowl 4 — Isomorphism Invariance of Diffusion-Convolutional Representations
theoretical result
The latent diffusion-convolutional representations produced by a Diffusion-Convolutional Neural Network (DCNN) are invariant under graph isomorphism.

Let $G_1 = (V_1, E_1)$ and $G_2 = (V_2, E_2)$ be two isomorphic graphs with identical vertex feature mappings such that $|V_1| = |V_2| = N$ , $X_1 = X_2 = X \in \mathbb{R}^{N \times F}$ , and transition power series tensors $P^*_1 = P^*_2 = P^*$ . The graph-level diffusion-convolutional activations $Z_1, Z_2 \in \mathbb{R}^{H \times F}$ computed by: $Z_{jk} = f\left(W^c_{jk} \odot \frac{1}{N} \sum_{v \in V} \sum_{v' \in V} P^*_{v j v'} X_{v' k}\right)$ satisfy $Z_1 = Z_2$ identically for all diffusion hops $j \in \{0, \dots, H-1\}$ and features $k \in \{1, \dots, F\}$ . Because parameters are tied across diffusion search depth rather than node indices or spatial grid locations, DCNN activations are permutation-invariant across isomorphic graphs.
Knowl 5 — DCNN Training and Optimization Procedure
algorithm
Diffusion-Convolutional Neural Networks are trained end-to-end via stochastic minibatch gradient descent on backpropagated multiclass hinge loss using AdaGrad optimization and windowed early stopping.
Input: Dataset $G = \{G_t = (V_t, E_t, X_t, Y_t)\}_{t=1}^T$ , diffusion depth $H$ , learning rate $\alpha = 0.05$ , early stopping window $w$
Output: Optimized weight tensors $W^c$ and $W^d$
Initialize weight tensors $W^c, W^d \sim \mathcal{N}(0, 0.01)$
Compute degree-normalized transition matrices $P_t$ and power series tensors $P^*_t$ for each graph $G_t$
repeat
    Randomly partition entity indices into minibatches
    for each minibatch $B$ do
        Compute diffusion-convolution activations $Z = f(W^c \odot (P^*_B X))$
        Compute dense predictions $\hat{Y} = f(W^d \odot Z)$
        Compute multiclass hinge loss $\mathcal{L}(\hat{Y}, Y)$
        Compute gradients $\nabla_{W^c} \mathcal{L}$ and $\nabla_{W^d} \mathcal{L}$ via backpropagation
        Update parameters $W^c$ and $W^d$ using AdaGrad with learning rate $\alpha$
    end for
    Compute validation loss $\mathcal{L}_{\text{val}}$ on the validation set
until $\mathcal{L}_{\text{val}}$ exceeds the moving average of validation losses over the previous $w$ epochs
return $W^c, W^d$

Knowl 6 — Relational Node Classification Performance on Cora and Pubmed

data/table

Node classification was evaluated on two citation network datasets: Cora (2,708 scientific papers across 7 classes, 5,429 citation links, 1,433 binary word features) and Pubmed (19,717 diabetes papers across 3 classes, 44,338 citation links, 500 TF-IDF features). In each trial, nodes were randomly split into equal-sized training, validation, and test partitions.

A 2-hop DCNN was benchmarked against feature-only $\ell_1$ - and $\ell_2$ -regularized logistic regression, graph-structure-only kernel methods (Exponential Diffusion Kernel KED and Laplacian Exponential Diffusion Kernel KLED), and a partially-observed Conditional Random Field with Loopy Belief Propagation (CRF-LBP).

	Cora			Pubmed
Model	Accuracy	F (micro)	F (macro)	Accuracy	F (micro)	F (macro)
l1logistic	0.7087	0.7087	0.6829	0.8718	0.8718	0.8698
l2logistic	0.7292	0.7292	0.7013	0.8631	0.8631	0.8614
KED	0.8044	0.8044	0.7928	0.8125	0.8125	0.7978
KLED	0.8229	0.8229	0.8117	0.8228	0.8228	0.8086
CRF-LBP	0.8449	–	0.8248	–	–	–
2-hop DCNN	0.8677	0.8677	0.8584	0.8976	0.8976	0.8943

The 2-hop DCNN achieved statistically significant improvements over all baselines on both datasets across accuracy, micro-averaged F1, and macro-averaged F1.

Knowl 7 — Graph Classification Benchmark Performance across Chemical and Biological Datasets

data/table

Whole-graph classification was evaluated on five standard benchmarks: NCI1 (4,100 chemical compounds, 37 node labels), NCI109 (4,127 chemical compounds, 38 node labels), MUTAG (188 nitro compounds, 7 node features), PTC (344 carcinogenicity compounds, 19 node features), and ENZYMES (600 protein structures, 3 node features). Graphs were randomly partitioned into equal-sized training, validation, and test splits.

DCNN models with diffusion breadths $H=2$ and $H=5$ were compared against linear classifiers on graph-averaged features (l1logistic, l2logistic) and the Weisfeiler-Lehman subtree deep graph kernel (deepwl).

	NCI1			NCI109
Model	Accuracy	F (micro)	F (macro)	Accuracy	F (micro)	F (macro)
l1logistic	0.5728	0.5728	0.5711	0.5555	0.5555	0.5411
l2logistic	0.5688	0.5688	0.5641	0.5586	0.5568	0.5402
deepwl	0.6215	0.6215	0.5821	0.5801	0.5801	0.5178
2-hop DCNN	0.6250	0.5807	0.5807	0.6275	0.5884	0.5884
5-hop DCNN	0.6261	0.5898	0.5898	0.6286	0.5950	0.5899
	MUTAG			PTC
Model	Accuracy	F (micro)	F (macro)	Accuracy	F (micro)	F (macro)
l1logistic	0.7190	0.7190	0.6405	0.5470	0.5470	0.4272
l2logistic	0.7016	0.7016	0.5795	0.5565	0.5565	0.4460
deepwl	0.6563	0.6563	0.5942	0.5113	0.5113	0.4444
2-hop DCNN	0.6635	0.7975	0.79747	0.5660	0.0500	0.0531
5-hop DCNN	0.6698	0.8013	0.8013	0.5530	0.0	0.0526
	ENZYMES
Model	Accuracy	F (micro)	F (macro)
l1logistic	0.1640	0.1640	0.0904
l2logistic	0.2030	0.2030	0.1110
deepwl	0.2155	0.2155	0.1431
2-hop DCNN	0.1590	0.1590	0.0809
5-hop DCNN	0.1810	0.1810	0.0991

Unlike in node classification, no single model dominated across all graph classification datasets. DCNNs performed competitively on NCI1, NCI109, and MUTAG, but underperformed deepwl on ENZYMES, demonstrating that mean-pooling node diffusion activations is suboptimal for summarizing whole-graph structures.

Knowl 8 — Computational Complexity of DCNN vs. Partially-Observed CRFs
theoretical result
Training a Diffusion-Convolutional Neural Network involves numerical gradient descent requiring a single forward and backward pass per instance, resulting in polynomial time complexity dominated by dense tensor multiplications: $\mathcal{O}(N_t^2 F)$ for a graph with $N_t$ vertices and $F$ features.

In contrast, learning parameters for partially-observed Conditional Random Fields (CRFs) involves minimizing negative marginal log-likelihood using a contrast-of-partition-functions objective that requires running loopy belief propagation twice per step (once over the entire graph and once with observed labels conditioned). This induces an exponential time complexity per optimization step of: $\mathcal{O}(E_t N_t^{C_t})$ where $E_t$ is the number of edges and $C_t$ is the size of the maximal clique in graph $G_t$ . DCNNs thereby achieve superior classification accuracy without incurring the combinatorial inference costs of probabilistic relational graphical models.
Knowl 9 — Scalability and Locality Limitations in DCNNs
limitation
Diffusion-Convolutional Neural Networks have two primary structural limitations:
1. Memory Scalability: Storing the full transition probability power series tensor $P^* \in \mathbb{R}^{N_t \times H \times N_t}$ as a dense tensor requires $\mathcal{O}(N_t^2 H)$ memory. On GPU hardware, this results in out-of-memory constraints for graphs larger than tens to hundreds of thousands of nodes, preventing application to million- or billion-node graphs.
2. Locality Bias: Constructing latent representations from diffusion processes centered at individual nodes restricts the model to local topological neighborhoods. Consequently, DCNNs may fail to capture non-local graph behaviors or long-range spatial dependencies between distant entities.
Knowl 10 — Impact of Diffusion Hop Count on Classification Accuracy
empirical result
Evaluating DCNN performance across varying diffusion depths $H$ demonstrates that for node classification tasks (e.g., on Cora and Pubmed), accuracy rises sharply as the search breadth increases from $H=0$ (isolated node features with no graph structure) to $H=3$ hops. Beyond 3 hops, performance gains plateau as the random walk transition matrix powers converge toward stationarity. For whole-graph classification tasks, increasing the diffusion breadth beyond $H=2$ to $H=5$ or $H=10$ provides no significant gain in predictive accuracy.

Coverage note — All core contributions, including the node, graph, and edge DCNN model formulations, isomorphism invariance property, optimization algorithm, complexity analysis, limitations, and empirical benchmark evaluations, have been included; no substantial contributed material was omitted.

References

1.John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. The Journal of Machine Learning Research, 2011.
2.James Bergstra, Olivier Breuleux, Frédéric Bastien, Pascal Lamblin, Razvan Pascanu, Guillaume Desjardins, Joseph Turian, David Warde-Farley, and Yoshua Bengio. Theano: a CPU and GPU math expression compiler. In Proceedings of the Python for Scientific Computing Conference (SciPy), 2010.
3.P Sen and L Getoor. Link-based classification. Technical Report, 2007.
4.François Fouss, Kevin Francoisse, Luh Yen, Alain Pirotte, and Marco Saerens. An experimental investigation of kernels on graphs for collaborative recommendation and semisupervised classification. Neural Networks, 31:53–72, July 2012.
5.Prithviraj Sen, Galileo Mark Namata, Mustafa Bilgic, Lise Getoor, Brian Gallagher, and Tina Eliassi-Rad. Collective Classification in Network Data. AI Magazine, 2008.
6.Pinar Yanardag and S V N Vishwanathan. Deep Graph Kernels. In the 21th ACM SIGKDD International Conference, pages 1365–1374, New York, New York, USA, 2015. ACM Press.
7.Nikil Wale, Ian A Watson, and George Karypis. Comparison of descriptor spaces for chemical compound retrieval and classification. Knowledge and Information Systems, 14(3):347–375, August 2007.
8.Asim Kumar Debnath, Rosa L Lopez de Compadre, Gargi Debnath, Alan J Shusterman, and Corwin Hansch. Structure-activity relationship of mutagenic aromatic and heteroaromatic nitro compounds. correlation with molecular orbital energies and hydrophobicity. Journal of medicinal chemistry, 34(2):786–797, 1991.
9.Hannu Toivonen, Ashwin Srinivasan, Ross D King, Stefan Kramer, and Christoph Helma. Statistical evaluation of the predictive toxicology challenge 2000–2001. Bioinformatics, 19(10):1183–1193, 2003.
10.Karsten M Borgwardt, Cheng Soon Ong, Stefan Schönauer, SVN Vishwanathan, Alex J Smola, and Hans-Peter Kriegel. Protein function prediction via graph kernels. Bioinformatics, 21(suppl 1):i47–i56, 2005.
11.Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally connected networks on graphs. arXiv.org, 2014.
12.M Henaff, J Bruna, and Y LeCun. Deep Convolutional Networks on Graph-Structured Data. arXiv.org, 2015.
13.F Scarselli, M Gori, Ah Chung Tsoi, M Hagenbuchner, and G Monfardini. The Graph Neural Network Model. IEEE Transactions on Neural Networks, 2009.
14.A Micheli. Neural Network for Graphs: A Contextual Constructive Approach. IEEE Transactions on Neural Networks, 2009.
15.David K Duvenaud, Dougal Maclaurin, Jorge Aguilera-Iparraguirre, Rafael Gómez-Bombarelli, Timothy Hirzel, Alán Aspuru-Guzik, and Ryan P Adams. Convolutional Networks on Graphs for Learning Molecular Fingerprints. NIPS, 2015.
16.Daphne Koller and Nir Friedman. Probabilistic Graphical Models: Principles and Techniques. The MIT Press, 2009.
17.Jakob Verbeek and William Triggs. Scene segmentation with crfs learned from partially labeled images. NIPS, 2007.
18.Trevor Cohn. Efficient Inference in Large Conditional Random Fields. ECML, 2006.

Citation

MLA

Atwood, J., and D. Towsley. “Diffusion-Convolutional Neural Networks”. arXiv, 2015, http://arxiv.org/abs/1511.02136v6.

APA

Atwood, J., & Towsley, D. (2015). Diffusion-Convolutional Neural Networks. arXiv. http://arxiv.org/abs/1511.02136v6

Chicago

Atwood, J., and D. Towsley. 2015. “Diffusion-Convolutional Neural Networks”. arXiv. http://arxiv.org/abs/1511.02136v6.

Harvard

Atwood, J. and Towsley, D. (2015) “Diffusion-Convolutional Neural Networks”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1511.02136v6.

Vancouver

1. Atwood J, Towsley D (2015) Diffusion-Convolutional Neural Networks. arXiv

BibTeX

@article{atwood2015diffusion,
  title = {Diffusion-Convolutional Neural Networks},
  author = {Atwood, James and Towsley, Don},
  year = {2015},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1511.02136v6},
  eprint = {1511.02136}
}

Metadata:arXiv

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: Authors

Diffusion-Convolutional Neural Networks

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Diffusion-Convolution Operation for Node Classification

Knowl 2 — Graph-Level Classification with Diffusion-Convolutional Networks

Knowl 3 — Edge Classification and Feature Incorporation via Adjacency Augmentation

Knowl 4 — Isomorphism Invariance of Diffusion-Convolutional Representations

Knowl 5 — DCNN Training and Optimization Procedure

Knowl 6 — Relational Node Classification Performance on Cora and Pubmed

Knowl 7 — Graph Classification Benchmark Performance across Chemical and Biological Datasets

Knowl 8 — Computational Complexity of DCNN vs. Partially-Observed CRFs

Knowl 9 — Scalability and Locality Limitations in DCNNs

Knowl 10 — Impact of Diffusion Hop Count on Classification Accuracy

References

Citation

BibTeX

Access the Paper