iN2V: Bringing Transductive Node Embeddings to Inductive Graphs

Ansgar Scherp; Nicolas Lell

REVIEW 2 major objections 4 minor 51 references

iN2V: Bringing Transductive Node Embeddings to Inductive Graphs

T0 review · 2 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Retrained node2vec plus a neighbor-averaging loop embeds nodes unseen at training time, improving inductive node classification by 1 point on average.

desk verdict A useful post-hoc propagation idea, but the evaluation protocol may let test-node structure seep into the training features, putting the inductive claim on shaky ground. read the letter →

arxiv 2506.05039 v1 pith:QSERXY4M submitted 2025-06-05 cs.LG

Nicolas Lell , Ansgar Scherp This is my paper

classification cs.LG

keywords inductivenodeembeddingsnode2vecfeaturepropagationclassificationshallowgraphpost-hocembeddingextensionmissingfeatureshomophilyandheterophily

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Shallow embedding methods such as node2vec are normally transductive: the whole graph, test nodes included, must be present during training, and a node that appears later has no embedding. This paper claims that a simple post-hoc loop can break that restriction. After training node2vec on the training subgraph only, iN2V assigns each unseen node the average embedding of its already-embedded neighbors, repeats this for a few iterations so information flows along longer paths, and also nudges the training embeddings themselves at each step, weighted by a factor $\lambda < 1$. Two training-time changes — randomly replacing a node's embedding by its neighborhood average, and losses that pull nodes toward their neighborhoods while keeping the neighbors distinct — prepare the embeddings for this averaging. The paper reports that, on nine benchmark graphs, iN2V embeddings improve node classification by 1 point on average over the Feature Propagation baseline, with up to 6 points of gain on individual datasets; because it needs only edges, the method also applies where node features are missing, and it can be concatenated with existing features to enrich them.

What carries the argument

The load-bearing object is the iterative propagation of Eq. 1, driven by a lookup vector that tracks which nodes already have embeddings. The update assigns an unembedded node the mean embedding of its embedded neighbors, sets an embedded node with embedded neighbors to $\lambda h_v + (1-\lambda) m_{N_s(v)}$, and leaves a node with no embedded neighbors unchanged; the parameter $\lambda$ and the number of extra “delay” iterations control how far the training embeddings may drift. Two training losses prepare the embedding space for this cheap extension: $L_{\text{close}}(v) = -\log \sigma(h_v \cdot m_{N(v)})$ pulls each node's embedding toward its mean neighborhood embedding, and $L_{\text{div}}(v)$ penalizes the average pairwise cosine similarity of the neighbors, blocking the trivial solution where all neighbors collapse to one point; both are added to the standard node2vec loss with weights $\alpha$ and $\beta$. The machinery works because after this training, the mean of a node's embedded neighbors is a good predictor of where an unseen node's embedding should sit.

What would settle it

Split a graph so that an entire connected component of test nodes has no edge to the training subgraph, then run iN2V and classify: every node in that component keeps its zero initialization, so accuracy there falls to majority-class guessing, contradicting the claim that iN2V provides embeddings to unseen nodes. A more graded test is to bucket test nodes by distance to the nearest training node and compare iN2V against Feature Propagation within each bucket; since the claimed gains come from adapting training embeddings, the improvement should appear only at short-to-moderate distances, and the two methods should converge where long chains of unseen nodes dominate.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that embeddings trained on the training subgraph can be extended to unseen nodes by iterative neighborhood averaging with a damped update for the training nodes themselves. The update in Eq. 1 gives an unembedded node the mean embedding of its embedded neighbors, moves an embedded node a fraction $1-\lambda$ toward that mean, and leaves nodes with no embedded neighbors unchanged; $\lambda < 1$ outperforms both frozen propagation ($\lambda = 1$) and Feature Propagation. The paper argues this adaptation matters: a test node whose neighbors were mostly negative samples during training would otherwise inherit an off-distribution vector, and iterating the update pulls its embedding between the meaningful cluster embeddings of the two training components it bridges. The claimed result is that iN2V effectively brings transductive embeddings to the inductive setting: averaged over datasets, splits, and classification models, it beats Feature Propagation by about 1 point on homophilic and 0.7 points on heterophilic graphs, and in some 10%-training splits it even surpasses using the original dataset features.

Load-bearing premise

The load-bearing premise is that at inference time every test node has at least one path to a training node, because the iterative loop can only assign an embedding to a node that, directly or through a chain of unseen nodes, touches the training subgraph; the paper states this limitation itself, and when a test component has no such path both iN2V and Feature Propagation fail.

Editorial extensions

If this is right

iN2V gives a structure-only route to inductive node classification: graphs without node features, or with partially missing feature vectors, can still feed a classifier using only edges and the trained embeddings.
Because the post-hoc loop is agnostic to how the base embeddings were trained, the same extension can be applied to other shallow embedding methods, and the paper's loss and sampling modifications can also be paired with Feature Propagation as the extension step.
The gains are largest in the low-data regime: averaged over datasets, iN2V leads Feature Propagation by 1.2 points when only 10–20% of nodes are available for training, and by 0.8 points when 60–80% are used.
When only the extended embeddings are used as input, an MLP performs on par with or better than GraphSAGE, because the neighbor information is already encoded in the embeddings and the post-hoc loop already performs a form of aggregation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the paper draws its splits randomly from a single graph, most test nodes sit close to training nodes; on truly time-ordered splits, where new nodes arrive in clusters far from older ones, I would expect the reported margin over Feature Propagation to shrink.
A natural setting the paper does not test is temporal graphs: each new snapshot's nodes are exactly the 'unseen' nodes of the previous step, and the same loop could be rerun as nodes arrive; the paper mentions temporal GNNs only as motivation.
The $L_{\text{div}}$ term encodes a broader design principle — keep neighbor embeddings non-redundant so that later averaging is informative — which could be tested as a drop-in regularizer for other shallow embedding methods or GNN pre-training.
On heterophilic graphs the averaging rule should actively hurt, since neighbors tend to have different classes; the near-majority-class results the paper reports for Actor and Roman-empire are consistent with that, and a heterophily-aware variant (averaging over structurally similar rather than adjacent nodes) would be the direct test.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Alex,

Here's my take on arXiv:2506.05039 (iN2V). The paper proposes a post-hoc iteration that extends node2vec embeddings to unseen nodes, plus two training-time modifications (a close-neighbor loss and a sampling trick) that make the embeddings more amenable to that extension. The idea is simple, clearly described, and the experimental sweep is generous: nine datasets, five split sizes, four downstream classifiers, ablation studies, and code. The limitations section is honest about the path-connectivity assumption. If the method works as advertised, it's a useful tool for feature-less or missing-feature nodes.

The soft spot the stress-test note flags is real: the evaluation protocol seems to train the downstream MLP/GraphSAGE on embeddings that have already been propagated over the full graph, including test nodes. With λ<1, Equation (1b) updates the training-node embeddings whenever test nodes are present, so the final training features used to fit the classifier can encode test-graph information. Feature Propagation keeps training embeddings fixed, so the comparison isn't clean. The paper says the embeddings are 'train[ed] and store[d] for all 10 splits' and then evaluated; it never says the classifier is fit before the post-hoc step. This matters because the 'frozen' λ=1 variant, which avoids the leak, actually trails Feature Propagation on the aggregate tables (61.74 vs 62.97 for MLP, 62.88 vs 63.20 for GraphSAGE). The averaged gains of the adaptive variant over FP (about 1 point) are exactly the regime where test-aware training features could explain the difference. So the central inductive claim, as written, is not fully verified.

Two smaller issues: there's no significance testing, and many per-dataset gains sit within one standard deviation of the baselines; and the 'best iN2V variant' selection in Table 2 is a form of validation-set model selection that should be reported more carefully, though it isn't a fatal flaw on its own.

My overall read: the paper is worth serious referee time because the mechanism is novel and the experiments are broad, but the authors need to clarify the classifier-training order and ideally add an evaluation where the classifier is fit before the post-hoc extension, or explicitly argue why the test-aware update doesn't inflate the results.

If you'd like to see how the leak plays out, the GitHub repo has the code; I'd encourage a quick check of when the MLP is trained. I would not cite this as an inductive method until that's resolved.

Best

Referee Report

2 major / 4 minor

Summary. The paper proposes iN2V, a post-hoc iterative propagation scheme (Eq. 1) that extends node2vec embeddings to nodes unseen during training, together with two training-time modifications (a sampling-based replacement and the L_close/L_div losses in §3.4) intended to make the learned embeddings more amenable to this extension. The authors evaluate on nine benchmark graphs with five training-set sizes and four downstream classifiers, reporting an average 1-point accuracy improvement over Feature Propagation and a plug-in combination with raw features.

Significance. If the reported gains are real under a strictly inductive protocol, iN2V would be a simple and useful plug-in for feature-less or partially observed graphs: the post-hoc iteration is cheap, the code is released, and the experimental coverage is broad (9 datasets, 5 splits, 4 models, and ablations). The paper is also honest about the path-to-training-node limitation. However, the central inductive claim is not yet established because the evaluation protocol may allow test-graph information to leak into the training of the downstream classifiers, and because the headline improvements are small relative to the reported variance and are not accompanied by significance testing.

major comments (2)

[§4.2 (procedure) and Eq. (1)] The manuscript never states whether the MLP/GraphSAGE classifiers are trained before or after the post-hoc extension in Eq. (1). Since with λ<1 Eq. (1b) updates training-node embeddings using neighbor embeddings that, after the first iteration, include test-node values, training the classifiers on the 'stored' embeddings (§4.2) means the training features depend on the test graph. This violates the inductive setting defined in §1 ('test data is entirely unseen during training') and makes the comparison with Feature Propagation unfair, since FP keeps training embeddings fixed. If instead the classifiers are trained before the extension, then the paper must analyze the train/test feature shift (training embeddings vs. post-extension test embeddings) and its effect on the reported accuracies. This concern is load-bearing: Table 3 shows that the frozen variant (λ=1), which is the only one that truly keeps training embeddings unchanged, actually underperforms FP (61.74 vs 62.97 for MLP; 62.88 vs 63.20 for GraphSAGE), so the claimed advantage over FP comes entirely from the test-aware adaptation of training embeddings.
[Table 3 and §5.1] The central quantitative claim is a 1-point average improvement over Feature Propagation (Table 3: post-hoc 64.38 vs FP 62.97 under MLP; 63.75 vs 63.20 under GraphSAGE). The per-cell standard deviations in Tables 2, 5, and 6 are frequently of the same magnitude as these differences, and some cells go against the trend (e.g., Amazon-ratings at 40% training with GraphSAGE: FP 45.72±0.75, iN2V 45.48±0.57). No significance tests, confidence intervals, or paired analyses across the 10 seeds are reported, so the headline improvement is not established beyond noise. Please provide paired tests (e.g., across the 10 splits per cell, or across dataset-split cells) and report effect sizes.

minor comments (4)

[Figure 1 and §3.2] The sentence 'the embedding of v17 =v 7/43 is close to zero' is difficult to parse; please clarify the notation intended for repeated averaging.
[§4.2] The description says 'the embeddings from the training set are extended to the validation set using iN2V' for hyperparameter selection; it would be helpful to state explicitly whether the logistic regression used for selection is trained on the pre-extension or post-extension training embeddings.
[§4.3] The ranges searched for λ and delay are not reported, although these are central to the post-hoc variant; please include them in the hyperparameter description.
[§5.1] The aggregated statement 'outperforms Feature Propagation by 1 point on homophilic and 0.7 points on heterophilic datasets' should clarify which datasets are in each group, since Actor shows no improvement and Roman-empire sometimes shows a degradation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Eq. 1 and the L_close/L_div losses are empirically evaluated on held-out labels; no fitted quantity is renamed as a prediction, and the only self-citations are contextual.

full rationale

The paper's central claim is empirical: the post-hoc averaging rule of Eq. 1 plus the modified N2V training (L_close, L_div, sampling) improves node classification over Feature Propagation on unseen nodes. No step of the derivation reduces to its own inputs by construction. Eq. 1 is a defined iterative rule, not a fitted quantity; lambda, delay, alpha, beta, and r are hyperparameters selected on the validation set, and all final numbers are reported on held-out test labels across nine datasets and four classifiers (MLP, GraphSAGE, GAT, GIN). The L_close and L_div losses are regularizers that bias training embeddings toward the post-hoc operation; this is an explicitly disclosed design choice ('The final loss for iN2V is L(v) = Ln2v(v) + alpha * Lclose(v) + beta * Ldiv(v)'), and the benefit of the modification is tested empirically (Tables 3, 4) rather than asserted identically. The contrast with Feature Propagation is explicit: FP keeps training embeddings fixed, while Eq. 1b with lambda<1 adapts them, and the ablation (Fig. 3, Table 3) shows the adaptive variant outperforms both the frozen variant and FP, which is an empirical finding against external baselines. The only self-citations (Lell & Scherp 2024; Polleres et al. 2023) appear in contextual related-work lists and are not load-bearing for any result; no uniqueness theorem or prior-work ansatz is imported. The stated limitation ('Our method generally can only provide embeddings to test nodes with a path to at least one train node') restricts applicability but does not make the derivation circular. A protocol ambiguity exists — Section 4.2 does not state whether MLP/GraphSAGE are fit before or after the test-aware propagation, which the skeptic notes — but that is a leakage/verification concern, not a reduction of a prediction to a fitted input, and therefore not circularity under the rubric.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

No new physical or mathematical entities are introduced. The method relies on standard hyperparameter tuning (lambda, delay, alpha, beta, r, N2V hyperparameters) and on the domain assumptions listed above. The most fragile assumptions are graph connectivity at inference time and the informativeness of random-walk embeddings on the given graph.

free parameters (8)

lambda (post-hoc blend) = tuned per dataset/split
Eq. 1b weights a node's own embedding against the mean neighbor embedding; lambda < 1 enables adaptation, the key difference from Feature Propagation.
delay = tuned per dataset/split
Number of extra propagation iterations after all reachable nodes are embedded; affects the convergence trade-off (Section 3.3, Figure 3).
alpha (L_close weight) = grid {0,0.1,1,10}
Controls the pull of a node's embedding toward its neighbor mean during training (Section 3.4).
beta (L_div weight) = grid {0,0.001,0.01,0.1}
Controls diversity among neighbor embeddings to prevent collapse (Section 3.4).
r (sampling probability) = grid {0.2,0.4,0.6,0.8}
Probability of replacing a node's embedding with its neighbor mean during training (Section 3.4).
N2V p and q = grid {0.2,1,5}
Random-walk bias hyperparameters tuned on validation (Section 4.3).
embedding dimension d = grid {64,256}
Size of the node embeddings; tuned on validation (Section 4.3).
N2V learning rate = grid {0.1,0.01,0.001}
Learning rate for embedding training; tuned on validation (Section 4.3).

assumptions (4)

domain assumption The graph is treated as undirected.
Section 4.3: 'We make all datasets undirected.' This simplifies neighbor sets and is standard but changes graph structure.
domain assumption At inference time, edges from new nodes to known nodes (and among new nodes) are observed, and every test node has a path to some training node.
Section 3.3 requires N_s(v) non-empty for propagation; Section 6 acknowledges failure when no path exists.
domain assumption Random-walk embeddings (N2V) encode information useful for classification, i.e., the graph is at least partly homophilic or structurally informative.
Section 5.1 and Section 6: on Actor and Roman-empire, N2V-based models predict only the majority class, showing this assumption fails for strongly heterophilic graphs.
domain assumption The iterative averaging procedure with lambda < 1 preserves class-discriminative information rather than washing it out.
Section 5.2 ablation shows performance drops with too many iterations (convergence to global average); the method depends on choosing a good delay.

how reviews work

0 comments

Cite this review

Pith. "Pith review of iN2V: Bringing Transductive Node Embeddings to Inductive Graphs." pith.science (2026). https://pith.science/paper/QSERXY4M

@misc{pith2026250605039,
  author       = {Pith},
  title        = {Pith review of: iN2V: Bringing Transductive Node Embeddings to Inductive Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QSERXY4M}},
  note         = {Machine review of arXiv:2506.05039}
}

read the original abstract

Shallow node embeddings like node2vec (N2V) can be used for nodes without features or to supplement existing features with structure-based information. Embedding methods like N2V are limited in their application on new nodes, which restricts them to the transductive setting where the entire graph, including the test nodes, is available during training. We propose inductive node2vec (iN2V), which combines a post-hoc procedure to compute embeddings for nodes unseen during training and modifications to the original N2V training procedure to prepare the embeddings for this post-hoc procedure. We conduct experiments on several benchmark datasets and demonstrate that iN2V is an effective approach to bringing transductive embeddings to an inductive setting. Using iN2V embeddings improves node classification by 1 point on average, with up to 6 points of improvement depending on the dataset and the number of unseen nodes. Our iN2V is a plug-in approach to create new or enrich existing embeddings. It can also be combined with other embedding methods, making it a versatile approach for inductive node representation learning. Code to reproduce the results is available at https://github.com/Foisunt/iN2V .

Figures

Figures reproduced from arXiv: 2506.05039 by the authors.

**Figure 1.** Three example graphs illustrate the post-hoc extensions to the white test nodes after obtaining embeddings for the gray training nodes in the inductive setup. Figure a) shows a simple example where node 2 obtains the average embedding of nodes 1 and 3. In Figure b) node 5 got a distant embedding during embedding training, but during the post-hoc extension it should be updated to move between the embeddings of the no… view at source ↗

**Figure 2.** Influence of amount of training data. Results are averaged over all datasets. These dataset-specific differences are explained by the usefulness of N2V embeddings vs the usefulness of the original features. On Roman-empire, the graph is close to a sequence of words with few additional edges; the neighborhood information encoded by N2V embeddings does not bring much useful information. Influence of Amount of Train… view at source ↗

**Figure 4.** shows the effect of the loss weights α and β on logistic regression validation accuracy. As we already saw in our main results, the loss-based N2V modification gives a small performance improvement. The figure suggests that α has a bigger influence on the performance as long as β is not too high. 0.0 0.001 0.01 0.1 1.0 10.0 0.0 0.01 0.1 1.0 10.0 100.0 79.7 79.7 79.7 79.7 79.7 79.6 79.7 79.7 79.7 79.7 79.7 79.6 79.7 … view at source ↗

Figures from the paper (3 more)

**Figure 3.** Figure 3: Ablation of the effect of λ for different delays vs Feature Propagation and MatMul. We perform ablation studies on our post-hoc method and the loss modification. For this, we use the 40% training split, set the N2V hyperparameters p and q to 1, embeddings size to 256, …

**Figure 5.** Figure 5: Sensitivity to learning rate and embedding dimension. B. Complete Result Tables Tables 5 and 6 show the comparison of MLP and GraphSAGE on the iN2V embeddings vs baselines. N2V, applied in the inductive setting, and Feature Propagation are comparable because they have …

**Figure 6.** Figure 6: Sensitivity to the N2V p and q hyperparameters. D. Using other GNNs The effectiveness of iN2V is not limited by the chosen GNN. To demonstrate this, we additionally performed all experiments with GAT (Velickovic et al., 2018), see Tables 13, 15, 17, and 19 and with GIN…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 36 canonical work pages

[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...
[2]

V., and Galstyan, A

Abu - El - Haija, S., Perozzi, B., Kapoor, A., Alipourfard, N., Lerman, K., Harutyunyan, H., Steeg, G. V., and Galstyan, A. Mixhop: Higher-order graph convolutional architectures via sparsified neighborhood mixing. In Chaudhuri, K. and Salakhutdinov, R. (eds.), Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019...

work page 2019
[3]

Adhikari, B., Zhang, Y., Ramakrishnan, N., and Prakash, B. A. Sub2vec: Feature learning for subgraphs. In Phung, D. Q., Tseng, V. S., Webb, G. I., Ho, B., Ganji, M., and Rashidi, L. (eds.), Advances in Knowledge Discovery and Data Mining - 22nd Pacific-Asia Conference, PAKDD 2018, Melbourne, VIC, Australia, June 3-6, 2018, Proceedings, Part II , volume 10...

work page doi:10.1007/978-3-319-93037-4 2018
[4]

Make heterophilic graphs better fit GNN: A graph rewiring approach

Bi, W., Du, L., Fu, Q., Wang, Y., Han, S., and Zhang, D. Make heterophilic graphs better fit GNN: A graph rewiring approach. IEEE Trans. Knowl. Data Eng. , 36 0 (12): 0 8744--8757, 2024. doi:10.1109/TKDE.2024.3441766. URL https://doi.org/10.1109/TKDE.2024.3441766

arXiv 2024
[5]

Translating embeddings for modeling multi-relational data

Bordes, A., Usunier, N., Garc \' a - Dur \' a n, A., Weston, J., and Yakhnenko, O. Translating embeddings for modeling multi-relational data. In Burges, C. J. C., Bottou, L., Ghahramani, Z., and Weinberger, K. Q. (eds.), Advances in Neural Information Processing Systems 26: 27th Annual Conference on Neural Information Processing Systems 2013. Proceedings ...

work page 2013
[6]

Simple and deep graph convolutional networks

Chen, M., Wei, Z., Huang, Z., Ding, B., and Li, Y. Simple and deep graph convolutional networks. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event , volume 119 of Proceedings of Machine Learning Research, pp.\ 1725--1735. PMLR , 2020. URL http://proceedings.mlr.press/v119/chen20v.html

work page 2020
[7]

Fede: Embedding knowledge graphs in federated setting

Chen, M., Zhang, W., Yuan, Z., Jia, Y., and Chen, H. Fede: Embedding knowledge graphs in federated setting. In IJCKG'21: The 10th International Joint Conference on Knowledge Graphs, Virtual Event, Thailand, December 6 - 8, 2021, pp.\ 80--88. ACM , 2021. doi:10.1145/3502223.3502233. URL https://doi.org/10.1145/3502223.3502233

work page arXiv 2021
[8]

Refactor gnns: Revisiting factorisation-based models from a message-passing perspective

Chen, Y., Mishra, P., Franceschi, L., Minervini, P., Stenetorp, P., and Riedel, S. Refactor gnns: Revisiting factorisation-based models from a message-passing perspective. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A. (eds.), Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processi...

work page 2022

Show all 51 references

[9]

Adaptive universal generalized pagerank graph neural network

Chien, E., Peng, J., Li, P., and Milenkovic, O. Adaptive universal generalized pagerank graph neural network. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 . OpenReview.net, 2021. URL https://openreview.net/forum?...

2021
[10]

Learning structural node embeddings via diffusion wavelets

Donnat, C., Zitnik, M., Hallac, D., and Leskovec, J. Learning structural node embeddings via diffusion wavelets. In Guo, Y. and Farooq, F. (eds.), Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD 2018, London, UK, August 19-...

2018
[11]

and Leskovec, J

Grover, A. and Leskovec, J. node2vec: Scalable feature learning for networks. In Krishnapuram, B., Shah, M., Smola, A. J., Aggarwal, C. C., Shen, D., and Rastogi, R. (eds.), Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San...

2016
[12]

Hahn, S. H. and Paulheim, H. Rdf2vec embeddings for updateable knowledge graphs--reuse, don’t retrain! ESWC Posters and Demos, 2024

2024
[13]

L., Ying, Z., and Leskovec, J

Hamilton, W. L., Ying, Z., and Leskovec, J. Inductive representation learning on large graphs. In Guyon, I., von Luxburg, U., Bengio, S., Wallach, H. M., Fergus, R., Vishwanathan, S. V. N., and Garnett, R. (eds.), Advances in Neural Information Processing Systems 30: Annual Co...

2017
[14]

Graph-mlp: Node classification without message passing in graph

Hu, Y., You, H., Wang, Z., Wang, Z., Zhou, E., and Gao, Y. Graph-mlp: Node classification without message passing in graph. CoRR, abs/2106.04051, 2021. URL https://arxiv.org/abs/2106.04051

2021 arXiv
[15]

and Benson, A

Jia, J. and Benson, A. R. A unifying generative model for graph learning algorithms: Label propagation, graph convolutions, and combinations. SIAM J. Math. Data Sci. , 4 0 (1): 0 100--125, 2022. doi:10.1137/21M1395351. URL https://doi.org/10.1137/21m1395351

2022 doi
[16]

Kipf, T. N. and Welling, M. Semi-supervised classification with graph convolutional networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings . OpenReview.net, 2017. URL https://openreview....

2017
[17]

Edge-splitting MLP : Node classification on homophilic and heterophilic graphs without message passing

Kohn, M., Hoffmann, M., and Scherp, A. Edge-splitting MLP : Node classification on homophilic and heterophilic graphs without message passing. In The Third Learning on Graphs Conference, 2024. URL https://openreview.net/forum?id=BQEb4r21cm

2024
[18]

Canonical tensor decomposition for knowledge base completion

Lacroix, T., Usunier, N., and Obozinski, G. Canonical tensor decomposition for knowledge base completion. In Dy, J. G. and Krause, A. (eds.), Proceedings of the 35th International Conference on Machine Learning, ICML 2018, Stockholmsm \" a ssan, Stockholm, Sweden, July 10-15, ...

2018
[19]

Le, Q. V. and Mikolov, T. Distributed representations of sentences and documents. In Proceedings of the 31th International Conference on Machine Learning, ICML 2014, Beijing, China, 21-26 June 2014 , volume 32 of JMLR Workshop and Conference Proceedings , pp.\ 1188--1196. JMLR...

2014
[20]

and Scherp, A

Lell, N. and Scherp, A. Hyperaggregation: Aggregating over graph edges with hypernetworks. In International Joint Conference on Neural Networks, IJCNN 2024, Yokohama, Japan, June 30 - July 5, 2024 , pp.\ 1--9. IEEE , 2024. doi:10.1109/IJCNN60899.2024.10650980. URL https://doi....

2024
[21]

L., Gupta, V., Bhalerao, O., and Lim, S

Lim, D., Hohne, F., Li, X., Huang, S. L., Gupta, V., Bhalerao, O., and Lim, S. Large scale learning on non-homophilous graphs: New benchmarks and strong simple methods. In Ranzato, M., Beygelzimer, A., Dauphin, Y. N., Liang, P., and Vaughan, J. W. (eds.), Advances in Neural In...

2021
[22]

Graph neural networks for temporal graphs: State of the art, open challenges, and opportunities

Longa, A., Lachi, V., Santin, G., Bianchini, M., Lepri, B., Lio, P., Scarselli, F., and Passerini, A. Graph neural networks for temporal graphs: State of the art, open challenges, and opportunities. Trans. Mach. Learn. Res., 2023, 2023. URL https://openreview.net/forum?id=pHCdMat0gI

2023
[23]

Revisiting heterophily for graph neural networks

Luan, S., Hua, C., Lu, Q., Zhu, J., Zhao, M., Zhang, S., Chang, X., and Precup, D. Revisiting heterophily for graph neural networks. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A. (eds.), Advances in Neural Information Processing Systems 35: Annual ...

2022
[24]

and Cangea, C

Mernyei, P. and Cangea, C. Wiki-cs: A wikipedia-based benchmark for graph neural networks. arXiv preprint arXiv:2007.02901, 2020

2007 arXiv
[25]

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. In Bengio, Y. and LeCun, Y. (eds.), 1st International Conference on Learning Representations, ICLR 2013, Scottsdale, Arizona, USA, May 2-4, 2013, Workshop Track Proce...

2013 arXiv
[26]

and Prokhorenkova, L

Mironov, M. and Prokhorenkova, L. Revisiting graph homophily measures. In The Third Learning on Graphs Conference, 2024. URL https://openreview.net/forum?id=fiFBjLD0LV

2024
[27]

Query-driven active surveying for collective classification

Namata, G., London, B., Getoor, L., Huang, B., and EDU, U. Query-driven active surveying for collective classification. In 10th International Workshop on Mining and Learning with Graphs, 2012

2012
[28]

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. CoRR, abs/1606.08928, 2016. URL http://arxiv.org/abs/1606.08928

2016 arXiv
[29]

graph2vec: Learning distributed representations of graphs

Narayanan, A., Chandramohan, M., Venkatesan, R., Chen, L., Liu, Y., and Jaiswal, S. graph2vec: Learning distributed representations of graphs. CoRR, abs/1707.05005, 2017. URL http://arxiv.org/abs/1707.05005

2017 arXiv
[30]

C., Lei, Y., and Yang, B

Pei, H., Wei, B., Chang, K. C., Lei, Y., and Yang, B. Geom-gcn: Geometric graph convolutional networks. In ICLR 2020 . OpenReview.net, 2020. URL https://openreview.net/forum?id=S1e2agrFvS

2020
[31]

Deepwalk: online learning of social representations

Perozzi, B., Al - Rfou, R., and Skiena, S. Deepwalk: online learning of social representations. In Macskassy, S. A., Perlich, C., Leskovec, J., Wang, W., and Ghani, R. (eds.), The 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD '14, New Yor...

2014
[32]

Characterizing graph datasets for node classification: Homophily-heterophily dichotomy and beyond

Platonov, O., Kuznedelev, D., Babenko, A., and Prokhorenkova, L. Characterizing graph datasets for node classification: Homophily-heterophily dichotomy and beyond. In Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., and Levine, S. (eds.), Advances in Neural Informati...

2023
[33]

A critical look at the evaluation of gnns under heterophily: Are we really making progress? In ICLR 2023

Platonov, O., Kuznedelev, D., Diskin, M., Babenko, A., and Prokhorenkova, L. A critical look at the evaluation of gnns under heterophily: Are we really making progress? In ICLR 2023 . OpenReview.net, 2023 b . URL https://openreview.net/pdf?id=tJbbQfw-5wv

2023
[34]

How Does Knowledge Evolve in Open Knowledge Graphs? Transactions on Graph Data and Knowledge, 1 0 (1): 0 11:1--11:59, 2023

Polleres, A., Pernisch, R., Bonifati, A., Dell'Aglio, D., Dobriy, D., Dumbrava, S., Etcheverry, L., Ferranti, N., Hose, K., Jim\' e nez-Ruiz, E., Lissandrini, M., Scherp, A., Tommasini, R., and Wachs, J. How Does Knowledge Evolve in Open Knowledge Graphs? Transactions on Graph...

2023 doi
[35]

Ribeiro, L. F. R., Saverese, P. H. P., and Figueiredo, D. R. struc2vec: Learning node representations from structural identity. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Halifax, NS, Canada, August 13 - 17, 2017 , pp...

2017
[36]

o tzsch, M., L \' e cu \' e , F., Fl \

Ristoski, P. and Paulheim, H. Rdf2vec: RDF graph embeddings for data mining. In Groth, P., Simperl, E., Gray, A. J. G., Sabou, M., Kr \" o tzsch, M., L \' e cu \' e , F., Fl \" o ck, F., and Gil, Y. (eds.), The Semantic Web - ISWC 2016 - 15th International Semantic Web Confere...

2016 doi
[37]

I., Chamberlain, B

Rossi, E., Kenlay, H., Gorinova, M. I., Chamberlain, B. P., Dong, X., and Bronstein, M. M. On the unreasonable effectiveness of feature propagation in learning on graphs with missing node features. In Rieck, B. and Pascanu, R. (eds.), Learning on Graphs Conference, LoG 2022, 9...

2022
[38]

F., and Catalyurek, U

Sancak, K., Balin, M. F., and Catalyurek, U. Do we really need complicated graph learning models? -- a simple but effective baseline. In The Third Learning on Graphs Conference, 2024. URL https://openreview.net/forum?id=0664MgKEVz

2024
[39]

Collective classification in network data

Sen, P., Namata, G., Bilgic, M., Getoor, L., Gallagher, B., and Eliassi - Rad, T. Collective classification in network data. AI Mag. , 2008. doi:10.1609/aimag.v29i3.2157

2008 doi
[40]

Pitfalls of graph neural network evaluation

Shchur, O., Mumme, M., Bojchevski, A., and G \" u nnemann, S. Pitfalls of graph neural network evaluation. CoRR, 2018. URL http://arxiv.org/abs/1811.05868

2018 arXiv
[41]

Rotate: Knowledge graph embedding by relational rotation in complex space

Sun, Z., Deng, Z., Nie, J., and Tang, J. Rotate: Knowledge graph embedding by relational rotation in complex space. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019 . OpenReview.net, 2019. URL https://openreview.net/fo...

2019
[42]

LINE: large-scale information network embedding

Tang, J., Qu, M., Wang, M., Zhang, M., Yan, J., and Mei, Q. LINE: large-scale information network embedding. In Gangemi, A., Leonardi, S., and Panconesi, A. (eds.), Proceedings of the 24th International Conference on World Wide Web, WWW 2015, Florence, Italy, May 18-22, 2015 ,...

2015
[43]

Tian, Y., Zhang, C., Guo, Z., Zhang, X., and Chawla, N. V. Learning mlps on graphs: A unified view of effectiveness, robustness, and efficiency. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net, 202...

2023
[44]

Complex embeddings for simple link prediction

Trouillon, T., Welbl, J., Riedel, S., Gaussier, \' E ., and Bouchard, G. Complex embeddings for simple link prediction. In Balcan, M. and Weinberger, K. Q. (eds.), Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-...

2016
[45]

Graph attention networks

Velickovic, P., Cucurull, G., Casanova, A., Romero, A., Li \` o , P., and Bengio, Y. Graph attention networks. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings . OpenReview.net, ...

2018
[46]

Knowledge graph embedding by translating on hyperplanes

Wang, Z., Zhang, J., Feng, J., and Chen, Z. Knowledge graph embedding by translating on hyperplanes. In Brodley, C. E. and Stone, P. (eds.), Proceedings of the Twenty-Eighth AAAI Conference on Artificial Intelligence, July 27 -31, 2014, Qu \' e bec City, Qu \' e bec, Canada , ...

2014 doi
[47]

Wu, F., Jr., A. H. S., Zhang, T., Fifty, C., Yu, T., and Weinberger, K. Q. Simplifying graph convolutional networks. In Chaudhuri, K. and Salakhutdinov, R. (eds.), Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, Cali...

2019
[48]

Representation learning on graphs with jumping knowledge networks

Xu, K., Li, C., Tian, Y., Sonobe, T., Kawarabayashi, K., and Jegelka, S. Representation learning on graphs with jumping knowledge networks. In Dy, J. G. and Krause, A. (eds.), Proceedings of the 35th International Conference on Machine Learning, ICML 2018, Stockholmsm \" a ssa...

2018
[49]

How powerful are graph neural networks? In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019

Xu, K., Hu, W., Leskovec, J., and Jegelka, S. How powerful are graph neural networks? In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019 . OpenReview.net, 2019. URL https://openreview.net/forum?id=ryGs6iA5Km

2019
[50]

Graph-less neural networks: Teaching old mlps new tricks via distillation

Zhang, S., Liu, Y., Sun, Y., and Shah, N. Graph-less neural networks: Teaching old mlps new tricks via distillation. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022 . OpenReview.net, 2022. URL https://openreview.ne...

2022
[51]

Beyond homophily in graph neural networks: Current limitations and effective designs

Zhu, J., Yan, Y., Zhao, L., Heimann, M., Akoglu, L., and Koutra, D. Beyond homophily in graph neural networks: Current limitations and effective designs. In Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., and Lin, H. (eds.), Advances in Neural Information Processing Syst...

2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[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...

[2] [2]

V., and Galstyan, A

Abu - El - Haija, S., Perozzi, B., Kapoor, A., Alipourfard, N., Lerman, K., Harutyunyan, H., Steeg, G. V., and Galstyan, A. Mixhop: Higher-order graph convolutional architectures via sparsified neighborhood mixing. In Chaudhuri, K. and Salakhutdinov, R. (eds.), Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019...

work page 2019

[3] [3]

Adhikari, B., Zhang, Y., Ramakrishnan, N., and Prakash, B. A. Sub2vec: Feature learning for subgraphs. In Phung, D. Q., Tseng, V. S., Webb, G. I., Ho, B., Ganji, M., and Rashidi, L. (eds.), Advances in Knowledge Discovery and Data Mining - 22nd Pacific-Asia Conference, PAKDD 2018, Melbourne, VIC, Australia, June 3-6, 2018, Proceedings, Part II , volume 10...

work page doi:10.1007/978-3-319-93037-4 2018

[4] [4]

Make heterophilic graphs better fit GNN: A graph rewiring approach

Bi, W., Du, L., Fu, Q., Wang, Y., Han, S., and Zhang, D. Make heterophilic graphs better fit GNN: A graph rewiring approach. IEEE Trans. Knowl. Data Eng. , 36 0 (12): 0 8744--8757, 2024. doi:10.1109/TKDE.2024.3441766. URL https://doi.org/10.1109/TKDE.2024.3441766

arXiv 2024

[5] [5]

Translating embeddings for modeling multi-relational data

Bordes, A., Usunier, N., Garc \' a - Dur \' a n, A., Weston, J., and Yakhnenko, O. Translating embeddings for modeling multi-relational data. In Burges, C. J. C., Bottou, L., Ghahramani, Z., and Weinberger, K. Q. (eds.), Advances in Neural Information Processing Systems 26: 27th Annual Conference on Neural Information Processing Systems 2013. Proceedings ...

work page 2013

[6] [6]

Simple and deep graph convolutional networks

Chen, M., Wei, Z., Huang, Z., Ding, B., and Li, Y. Simple and deep graph convolutional networks. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event , volume 119 of Proceedings of Machine Learning Research, pp.\ 1725--1735. PMLR , 2020. URL http://proceedings.mlr.press/v119/chen20v.html

work page 2020

[7] [7]

Fede: Embedding knowledge graphs in federated setting

Chen, M., Zhang, W., Yuan, Z., Jia, Y., and Chen, H. Fede: Embedding knowledge graphs in federated setting. In IJCKG'21: The 10th International Joint Conference on Knowledge Graphs, Virtual Event, Thailand, December 6 - 8, 2021, pp.\ 80--88. ACM , 2021. doi:10.1145/3502223.3502233. URL https://doi.org/10.1145/3502223.3502233

work page arXiv 2021

[8] [8]

Refactor gnns: Revisiting factorisation-based models from a message-passing perspective

Chen, Y., Mishra, P., Franceschi, L., Minervini, P., Stenetorp, P., and Riedel, S. Refactor gnns: Revisiting factorisation-based models from a message-passing perspective. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A. (eds.), Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processi...

work page 2022