REVIEW 3 major objections 5 minor 34 references
Toward Data-centric Directed Graph Learning: An Entropy-driven Approach
T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash
Pith's one-line read EDEN claims that an entropy-built hierarchical knowledge tree, refined by mutual-information scores, distills data-level digraph knowledge into better node and link predictions.
desk verdict EDEN has a novel and empirically strong hierarchical KD pipeline for digraphs, but the core MI lower-bound theorems are invalid as written, so the paper's theoretical grounding collapses. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the Hierarchical Knowledge Tree (HKT), a height-h partition tree over the digraph's nodes built by minimizing directed structural entropy; it carries the argument by supplying the teacher-student structure in which parent nodes act as teachers and child nodes as students. The mechanism that refines the tree is the mutual-information neural estimator of Theorems 3.1-3.3, which scores nodes by the dependency between a node and its generalized neighbourhood, where the neighbourhood is defined by HKT partitions rather than by the adjacency matrix. Those scores set the affinity weights in Eq. (9), the distillation loss in Eq. (10) transfers refined parent knowledge to child nodes, and the tree-based random walk in Eq. (11) aggregates multi-level representations for leaf prediction.
What would settle it
Take a small digraph with known node features, compute the exact mutual information for every node, and compare it against the trained affinity scores produced by Eq. (9). If the ordering of the scores does not match the ordering of the true mutual-information values, or if the right-hand side of Eq. (6) exceeds the true mutual information, then the information-theoretic refinement step is not doing what the theorems claim.
Extended reading notes
Core claim
On the paper's own terms, the discovery is that digraph data knowledge can be organised as a Hierarchical Knowledge Tree and then distilled. EDEN first computes directed structural entropy from random walks that include reverse probability and sink self-loops, and minimises this entropy greedily to build a coarse tree. It then uses a GAN-style mutual-information neural estimator, justified by Theorems 3.1-3.3, to score how well each node represents its generalized neighbourhood; these scores refine the tree by keeping or reassigning nodes across partitions. Finally, parent nodes act as teachers and child nodes as students in an online knowledge-distillation loss, and a tree-based random walk gathers multi-level representations for leaf prediction. The paper reports that this pipeline attains better accuracy than the compared baselines on node and link tasks across homophilic and heterophilic digraphs, and that it improves existing directed and undirected GNNs when plugged in.
Load-bearing premise
The argument rests on the premise that the GAN-style objective in Eq. (6) is a valid lower bound on the mutual information between a node and its generalized neighbourhood; if that bound is not valid, the affinity scores that refine the HKT are not information-theoretically grounded.
Editorial extensions
If this is right
- A digraph learner built on EDEN should beat existing DiGNNs on node classification and on link existence, direction, and three-class link classification; the paper reports consistent wins across 10 digraph datasets.
- Existing (Di)GNNs can be upgraded by attaching EDEN as a plug-in module, with reported accuracy gains up to 4.96% over the same model without EDEN.
- EDEN transfers to undirected graphs and to heterophilic settings where connected nodes do not share labels; the experiments include both homophilic and heterophilic benchmarks.
- Under feature, edge, or label sparsity, EDEN degrades more gracefully than baselines because tree-based multi-level representations compensate for missing input signals.
- The weighting of the distillation loss matters by task: edge-existence prediction favours small weights, while node-level and finer link tasks favour larger weights, indicating the distilled knowledge is most useful where supervision is fine-grained.
Reading between the lines
- The parent-child distillation view suggests a practical recipe for semi-supervised graphs: leaf nodes with labels can refine their parents' soft targets, which then regularize unlabeled children; the paper's label-sparsity experiments are consistent with this, though they do not isolate it.
- The same two-stage construction of an entropy tree followed by mutual-information refinement could be applied to graph-level tasks, where the tree would summarise collections of graphs rather than nodes; that is a plausible extension the paper does not pursue.
- Because the coarse tree is built without labels, EDEN could be adapted to unsupervised representation learning by using the tree-based random walk as a graph-equivariant augmentation; the paper only evaluates supervised tasks.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript proposes EDEN, a data-centric framework for directed graph learning. It first constructs a coarse hierarchical knowledge tree (HKT) by minimizing a directed structural-entropy measure, then refines the tree using a claimed mutual-information neural estimator between node profiles and generalized neighborhoods, and finally performs parent-to-child knowledge distillation within the tree, followed by a random-walk-based leaf prediction. The paper claims state-of-the-art results on node- and link-level tasks over 14 (di)graph datasets and reports consistent improvements when EDEN is plugged into existing (Di)GNNs.
Significance. The proposed direction—hierarchical data-level knowledge distillation for digraphs—is timely, and the experimental campaign is broad: 14 datasets, multiple tasks, ablations, and sparsity and sensitivity analyses. If the theoretical underpinning were sound, EDEN would be a notable contribution to data-centric graph learning. However, the central theoretical result, the mutual-information neural estimator in Theorems 3.1–3.3, is not a valid variational MI bound as stated; since the affinity scores and the fine-grained HKT refinement in Eqs. (7)–(9) rest on this result, the main conceptual claim is not supported. The method may still work as a heuristic self-distillation regularizer, but that is a substantially weaker claim than the one made in the paper.
major comments (3)
- [Section 3.1, Eqs. (5)–(6); Appendix A.7–A.8] The derivation of the GAN-based MI neural estimator is mathematically invalid. The f-divergence variational form in Eq. (17) of Appendix A.7 is D_f(P||Q) ≥ sup_F [E_P T(F) + E_Q U(F)] with a single function F appearing in both the positive and negative expectation terms. In Eqs. (5) and (6) (and Eq. (21) of Appendix A.8), the authors replace this single supremum by two independent suprema: max_w over the positive term plus max_w over the negative term. This is a different optimization problem, since sup_F a(F) + sup_F b(F) is an upper bound on sup_F(a(F)+b(F)) and the two maximizing functions need not coincide. Consequently, C(Ω) in Eq. (6) is not the standard variational lower bound of MI. In fact, with the sigmoid parameterization both terms are non-positive and can be driven toward 0 separately on any finite sample when the function class is rich enough, making C(Ω) degenerate and largely data-independent. Theorems 3.2 and 3.3 therefore do not establish that C(Ω) estimates MI, and the affinity scores in Eq. (9) lose their claimed information-theoretic grounding.
- [Section 3.1, Eqs. (7)–(10); Section 2.2] The paper's data-centric framing is weakened by the model dependence of the knowledge extraction. The topology-based HKT construction in Sec. 3.1 and Appendix A.5 is independent of the trained model, but the fine-grained refinement explicitly uses the trainable model M: F_intra and F_inter in Eqs. (7)–(8) are computed from M(X_v), and the knowledge generation in Eq. (9) and the KD loss in Eq. (10) are functions of the same trainable encoder. EDEN is therefore closer to a self-distillation and regularization scheme than to the extraction of intrinsic data knowledge. The authors should either provide a data-only variant of the profile refinement or substantially temper the claim that the method is data-centric.
- [Theorem 3.3 and Eq. (6)] Even if the split-supremum expression is treated as a heuristic score rather than an MI lower bound, the criterion C(Ω) as written is degenerate: both terms are non-positive (log σ ≤ 0 and log(1−σ) ≤ 0), and with independent maximization over w each term can approach its trivial optimum separately. Thus C(Ω) does not quantify the dependency between nodes and their generalized neighborhoods, and it cannot rank nodes by information content in the way claimed. A valid criterion would need to use the same network in both expectations, as in the standard f-divergence variational form, or be re-derived explicitly for the proposed architecture.
minor comments (5)
- [Abstract and Section 1] The phrase 'hot-and-plug' appears repeatedly and should likely be 'plug-and-play'.
- [Eq. (10)] The definition of U_{X_p}^p and the expression X_p^{X_p}/U_{X_p}^p are unclear; the notation mixes partition indices, node indices, and class sums, and should be rewritten with explicit dimensionalities and a precise definition of each symbol.
- [Appendix A.4] The figures are referenced as 'Fig. ??' in the text; the figure numbers are missing.
- [Appendix A.7] The symbol σ is used both for the activation mapping into the conjugate domain and for the sigmoid function; this collision makes the derivation of Eq. (19) difficult to follow.
- [Eq. (11)] The transition probabilities 1/p_rw, 1/s_rw, and 1/c_rw are not normalized as written; the authors should clarify whether these are weights, probabilities, or unnormalized scores.
Circularity Check
Mild self-reference: the KD teacher (parent representation) is computed from the student model M itself; the invalid MI lower bound is a separate correctness issue, not a circularity.
-
self definitional
[Sec. 3.1–3.2, Eqs. (6)–(10); Appendix A.8, Eq. (21)]
"we obtain C(Ω)=bI(Ω)GAN = max_w 1/|Ω| Σ_{v∈Ω} logσ(F_w(X_v,X_NT_v)) + max_w 1/|Ω|^2 Σ_{(v,v̄)∈Ω} log(1−σ(F_w(X_v,X_NT_v̄))) ... Sv1 = σ(Qintra(W1(M(Xv1)),W2(M(XN_T_v1)))) ... Finally, we obtain the parent representation of Xp by Xp = SΩp XΩp. ... Lkd = ||X_p^p/U_Xp^p − Qchild(X_{v1,v2}^p)||_F."
The teacher (parent) representation is computed by weighting node features with affinity scores S that are functions of the trainable model M (Eq. 9 depends on M(X_v) and M(X_NT_v)); the student side Qchild(X_{v1,v2}^p) in the KD loss is also a function of M's embeddings. Therefore Lkd aligns child embeddings to a parent aggregate built from the same encoder's outputs. The 'knowledge refined by MI neural estimation' is thus not an independent data-derived target but a function of the model being trained; optimizing Lkd is a self-consistency constraint on M.
full rationale
The principal empirical claim (SOTA on 14 (di)graph datasets and up to 4.96% plug-in improvement) is tested against external baselines, so the final accuracy numbers are not circular predictions. The coarse-grained HKT construction (Eqs. 1–3) uses only topological measurements—degrees, partition volumes, and structural entropy—and is independent of the trained model. The circularity burden lies in the profile-refinement and distillation loop: affinity scores S and the parent representation X_p are generated from the trainable encoder M (Eqs. 7–9), and the KD loss (Eq. 10) then trains the same model toward these model-generated aggregates. This makes the 'entropy-driven data knowledge' partly self-definitional, but it does not reduce the final label prediction to a fitted parameter, since supervised cross-entropy remains. Separately, Theorems 3.1–3.3 contain a serious correctness flaw: the proof in Appendix A.7–A.8 splits a single variational supremum into two independent suprema, so Eq. (6) is not a valid MI lower bound. That flaw undermines the information-theoretic framing but is a mathematical error, not a circularity; it should be weighed as a correctness risk rather than raising the circularity score further. No load-bearing self-citation chain was found.
Assumptions & free parameters
free parameters (6)
- HKT height h =
grid searched in [3,10]
- Knowledge domain expansion coefficient kappa =
grid searched in [1,2]
- Random walk transition probabilities (p_rw, s_rw, c_rw) =
grid searched in [0,1]
- KD loss weight alpha =
grid searched in [0,1]
- PPR teleport tau and propagation steps l =
tau=0.5, l=5
- Layer-wise propagation iterations L =
unspecified
assumptions (5)
- ad hoc to paper The information entropy H is captured by the directed topology and profile-based hierarchical encoding system, reflecting the uncertainty of complex systems (Assumption 2.1).
- ad hoc to paper The true structure T is obtained by minimizing H (Assumption 2.2).
- ad hoc to paper The GAN-like objective with two separate maximizations over w is a valid lower bound on MI (Theorems 3.2-3.3).
- domain assumption The amended random walk (reverse probability and self-loops) yields a structural entropy that is minimized by a useful partition tree.
- domain assumption The greedy algorithm (Algorithms 1-2) approximates the minimizer of H_h sufficiently well for KD to help.
invented entities (2)
-
Hierarchical Knowledge Tree (HKT)
-
Data knowledge K (Assumption 2.3)
Cite this review
Pith. "Pith review of Toward Data-centric Directed Graph Learning: An Entropy-driven Approach." pith.science (2026). https://pith.science/paper/PDV6B2K4
@misc{pith2026250500983,
author = {Pith},
title = {Pith review of: Toward Data-centric Directed Graph Learning: An Entropy-driven Approach},
year = {2026},
howpublished = {\url{https://pith.science/paper/PDV6B2K4}},
note = {Machine review of arXiv:2505.00983}
}
read the original abstract
The directed graph (digraph), as a generalization of undirected graphs, exhibits superior representation capability in modeling complex topology systems and has garnered considerable attention in recent years. Despite the notable efforts made by existing DiGraph Neural Networks (DiGNNs) to leverage directed edges, they still fail to comprehensively delve into the abundant data knowledge concealed in the digraphs. This data-level limitation results in model-level sub-optimal predictive performance and underscores the necessity of further exploring the potential correlations between the directed edges (topology) and node profiles (feature and labels) from a data-centric perspective, thereby empowering model-centric neural networks with stronger encoding capabilities. In this paper, we propose \textbf{E}ntropy-driven \textbf{D}igraph knowl\textbf{E}dge distillatio\textbf{N} (EDEN), which can serve as a data-centric digraph learning paradigm or a model-agnostic hot-and-plug data-centric Knowledge Distillation (KD) module. The core idea is to achieve data-centric ML, guided by our proposed hierarchical encoding theory for structured data. Specifically, EDEN first utilizes directed structural measurements from a topology perspective to construct a coarse-grained Hierarchical Knowledge Tree (HKT). Subsequently, EDEN quantifies the mutual information of node profiles to refine knowledge flow in the HKT, enabling data-centric KD supervision within model training. As a general framework, EDEN can also naturally extend to undirected scenarios and demonstrate satisfactory performance. In our experiments, EDEN has been widely evaluated on 14 (di)graph datasets (homophily and heterophily) and across 4 downstream tasks. The results demonstrate that EDEN attains SOTA performance and exhibits strong improvement for prevalent (Di)GNNs.
Figures
Figures from the paper (5 more)
Reference graph
Works this paper leans on
-
[1]
This three-phase process ensures the efficient construction of a height-limited partition tree while minimizing directed structural measurement. A.6. The Proof of Theorem 3.1 As discussed in Sec. 3.1, node profiles in a digraph act as essential identifiers. These profiles are not only instrumental in distinguishing nodes but also play a critical role in t...
work page 2025
-
[4]
Notably, to provide a more thorough assessment, we introduce two additional evaluation metrics, Area Under Curve (AUC) and Average Precision (AP), alongside the commonly known Accuracy (ACC). We default to using AUC and AP in the evaluation of the link prediction tasks, and ACC to evaluate the predictive performance of node-level tasks. Regarding the expe...
work page 2025
-
[5]
The visualization experiments of the interruption issue of single-direction random walk on digraphs. The circle represents a special topology where a node can walk back to itself, and its existence will alleviate walk interruption. The y-axis denotes the proportion of non-walk interruptions. (7) For eachα,Tα is the union ofTβ for allβ such thatβ+ =α. Thus...
work page 2025
-
[6]
Maekawa, S., Sasaki, Y ., and Onizuka, M. Why using either aggregated features or adjacency lists in directed or undi- rected graph? empirical study and simple classification method. arXiv preprint arXiv:2306.08274,
-
[7]
Model performance (%) in three directed link-level downstream tasks. Datasets (→) Slashdot Epinions Tasks (→) Exist Direct Link-C Exist Direct Link-C Models (↓) AUC AP AUC AP ACC AUC AP AUC AP ACC GCNII 88.6±0.1 88.4±0.0 90.3±0.1 90.4±0.1 84.0±0.1 91.3±0.1 91.3±0.0 85.9±0.2 86.3±0.1 82.7±0.1 GATv2 88.2±0.2 88.5±0.1 90.6±0.1 90.4±0.1 83.7±0.3 91.8±0.2 91.6...
work page 2025
-
[8]
On evaluating the donsker-varadhan i-function
10 Submission and Formatting Instructions for ICML 2025 Pinsky, R. On evaluating the donsker-varadhan i-function. The Annals of Probability, pp. 342–362,
work page 2025
-
[9]
Link-C ACC and others AUC (%) in three directed link-level downstream tasks. Datasets Tasks GCNII GATv2 AGT DGCN DIMPA D-HYPR DiGCN MagNet EDEN CoraML Existence84.01±0.22 84.58±0.33 84.50±0.2487.65±0.20 88.06±0.20 87.99±0.24 87.65±0.28 88.05±0.2190.84±0.19 Direction 83.25±0.36 84.94±0.60 85.57±0.5190.43±0.49 90.88±0.50 90.94±0.54 89.75±0.71 90.83±0.4992.3...
work page 1979
-
[10]
Digraph inception convolutional networks
Tong, Z., Liang, Y ., Sun, C., Li, X., Rosenblum, D., and Lim, A. Digraph inception convolutional networks. Advances in Neural Information Processing Systems, NeurIPS , 2020a. Tong, Z., Liang, Y ., Sun, C., Rosenblum, D. S., and Lim, A. Directed graph convolutional network. arXiv preprint arXiv:2004.13970, 2020b. Veliˇckovi´c, P., Cucurull, G., Casanova, ...
arXiv 2004
Show all 34 references
-
[11]
Data-centric graph learning: A survey
Yang, C., Bo, D., Liu, J., Peng, Y ., Chen, B., Dai, H., Sun, A., Yu, Y ., Xiao, Y ., Zhang, Q., et al. Data-centric graph learning: A survey. arXiv preprint arXiv:2310.04987,
-
[13]
Zheng, X., Liu, Y ., Pan, S., Zhang, M., Jin, D., and Yu, P. S. Graph neural networks for graphs with heterophily: A survey. arXiv preprint arXiv:2202.07082,
-
[14]
W.-C., and Pan, S
Zheng, X., Liu, Y ., Bao, Z., Fang, M., Hu, X., Liew, A. W.-C., and Pan, S. Towards data-centric graph ma- chine learning: Review and outlook. arXiv preprint arXiv:2309.10979,
-
[15]
Outline The appendix is organized as follows: A.1 HKT Construction and Hierarchical Graph Clustering
12 Submission and Formatting Instructions for ICML 2025 A. Outline The appendix is organized as follows: A.1 HKT Construction and Hierarchical Graph Clustering. A.2 Data-level Online Knowledge Distillation in EDEN. A.3 The Definition of partition tree. A.4 Breaking the Limitat...
2025
-
[19]
The chosen activation function isσ(·) =− log(1 + exp(·))
log(x + 1), with 20 Submission and Formatting Instructions for ICML 2025 the conjugated divergence function defined asf⋆(t) =− log(1− exp(t)) (Nowozin et al., 2016). The chosen activation function isσ(·) =− log(1 + exp(·)). The GAN-like divergence can be expressed as: DGAN P f...
2025
-
[20]
The statistical information of the experimental di(graph) benchmark datasets. Topology-Profile Datasets #Node #Features #Edges #N Classes #N Train/Val/Test #L Train/Val/Test #Task Description Homophily Photo 7,487 745 119,043 8 612/612/5,889 Undirected Transductive Co-purchase...
2005
-
[22]
Nodes correspond to tolokers (workers) who have engaged in at least one of the 13 selected projects
is derived from the Toloka crowdsourcing platform (Likhobaba et al., 2023). Nodes correspond to tolokers (workers) who have engaged in at least one of the 13 selected projects. An edge connects two tolokers if they have collaborated on the same task. The objective is to predic...
2023
-
[23]
Nodes are connected by an edge if the words either follow each other in the text or are linked in the sentence’s dependency tree
is based on the Roman Empire article from the English Wikipedia (Lhoest et al., 2021), each node in the graph corresponds to a non-unique word in the text, mirroring the article’s length. Nodes are connected by an edge if the words either follow each other in the text or are l...
2021
-
[24]
Nodes represent various products, and edges connect items frequently bought together
is derived from the Amazon product co-purchasing network metadata available in the SNAP4 datasets (Leskovec & Krevl, 2014). Nodes represent various products, and edges connect items frequently bought together. The task involves predicting the average rating given by reviewers,...
2014
-
[25]
Each paper comes with a 128- dimensional feature vector obtained by averaging the embeddings of words in its title and abstract
is a citation graphs indexed by MAG (Wang et al., 2020). Each paper comes with a 128- dimensional feature vector obtained by averaging the embeddings of words in its title and abstract. The embeddings of individual words are computed by running the skip-gram model. Slashdot (O...
2020
-
[26]
includes all users and discussions from the inception of Wikipedia. The network comprises n = 2, 388, 953 nodes, where each node represents a Wikipedia user, and a directed edge from nodevi to nodevj indicates that useri edited userj ’s talk page at least once. For our analysi...
2022
-
[27]
The descriptions of them can be found later in this section
(NAG), and AGT (Ma et al., 2023). The descriptions of them can be found later in this section. For link-level dataset split, we are aligned with previous work (Zhang et al., 2021c; He et al., 2022a; 2023). To alleviate the influence of randomness, we repeat each experiment 10 ...
2023
-
[28]
This strategy provides better node representation capabilities and enhanced robustness when dealing with graph noise
introduces a variant with dynamic graph attention mechanisms to improve GAT. This strategy provides better node representation capabilities and enhanced robustness when dealing with graph noise. OptBasisGNN (Guo & Wei, 2023): OptBasisGNN revolutionizes GNNs by redefining polyn...
2023
-
[29]
consists of a learnable centrality encoding strategy and a kenneled local structure encoding mechanism to extract structural patterns from the centrality and subgraph views to improve node representations for the node-level downstream tasks. 25 Submission and Formatting Instru...
2025
-
[30]
𝑝!" 𝑝!" 𝑝!
Convergence curves on node- and link-level tasks. A.12. Hyperparameter settings The hyperparameters in the baseline models are set according to the original paper if available. Otherwise, we perform a hyperparameter search via the Optuna (Akiba et al., 2019). For our proposed ...
2019
-
[33]
Link-C ACC and others AUC (%) in three directed link-level downstream tasks. Datasets Tasks GCN GAT OptBG NAG NSTE Dir-GNN MGC HoloNet EDEN CoraML Existence83.26±0.18 83.96±0.25 83.55±0.1684.32±0.20 87.94±0.18 88.15±0.21 87.86±0.20 87.80±0.2490.84±0.19 Direction 82.73±0.32 84....
-
[1948]
Pitfalls of graph neural network evaluation
Shchur, O., Mumme, M., Bojchevski, A., and G¨unnemann, S. Pitfalls of graph neural network evaluation. arXiv preprint arXiv:1811.05868,
-
[2005]
and Cangea, C
Mernyei, P. and Cangea, C. Wiki-cs: A wikipedia-based benchmark for graph neural networks. arXiv preprint arXiv:2007.02901,
2007 arXiv
-
[2010]
V ., Jernite, Y ., Thakur, A., von Platen, P., Patil, S., Chaumond, J., Drame, M., Plu, J., Tunstall, L., et al
Lhoest, Q., del Moral, A. V ., Jernite, Y ., Thakur, A., von Platen, P., Patil, S., Chaumond, J., Drame, M., Plu, J., Tunstall, L., et al. Datasets: A community li- brary for natural language processing. arXiv preprint arXiv:2109.02846,
-
[2014]
Learning word vectors for 157 languages
Grave, E., Bojanowski, P., Gupta, P., Joulin, A., and Mikolov, T. Learning word vectors for 157 languages. arXiv preprint arXiv:1802.06893,
-
[2018]
Here, we use F (·,·) to compute the dependency
is defined as: DKL P fv,fN Tv ∥Pfv⊗PfN Tv ≥ sup F∈F EXv,XN Tv ∼P fv,fN Tv F Xv, XN Tv − sup F∈F EXv∼Pfv,XN T¯v ∼PfN Tv e F Xv,XN T¯v −1 , (15) whereF is an arbitrary class of functions that maps a pair of selected node features and its generalized neighborhood features to a re...
2025
-
[2019]
and Cremers, D
Koke, C. and Cremers, D. Holonets: Spectral convo- lutions do extend to directed graphs. arXiv preprint arXiv:2310.02232,
-
[2020]
The dataset consists of nodes corresponding to computer science articles, with edges based on hyperlinks and 10 classes representing different branches of the field
is a Wikipedia-based dataset for bench-marking GNNs. The dataset consists of nodes corresponding to computer science articles, with edges based on hyperlinks and 10 classes representing different branches of the field. The node features are derived from the text of the corresp...
2014
-
[2021]
9 Submission and Formatting Instructions for ICML 2025 Kipf, T. N. and Welling, M. Semi-supervised classifica- tion with graph convolutional networks. In International Conference on Learning Representations, ICLR,
2025
-
[2022]
Magnet: A neural network for directed graphs
11 Submission and Formatting Instructions for ICML 2025 Zhang, X., He, Y ., Brugnone, N., Perlmutter, M., and Hirn, M. Magnet: A neural network for directed graphs. Advances in Neural Information Processing Systems, NeurIPS, 2021c. Zhang, Z., Bu, J., Ester, M., Zhang, J., Yao,...
2025 arXiv
-
[2023]
URL https://github
doi: 10.5281/zenodo.7620795. URL https://github. com/Toloka/TolokerGraph. Lin, L. and Gao, J. A magnetic framelet-based convolutional neural network for directed graphs. InIEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP,
Reviewed August 16, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.