pith. sign in

arxiv: 2606.04647 · v1 · pith:QTKXWX5Znew · submitted 2026-06-03 · 💻 cs.LG

ALINC: Active Learning for Inductive Node Classification via Graph Sampling

Pascal Plettenberg , Denis Huseljic , Andr\'e Alcalde , Bernhard Sick , Josephine M. Thomas This is my paper

Pith reviewed 2026-06-28 07:05 UTC · model grok-4.3

classification 💻 cs.LG
keywords active learninginductive node classificationgraph samplingaggregation mechanismsnode classificationgraph neural networksmolecular predictioncircuit design
0
0 comments X

The pith

ALINC selects entire graphs for annotation in inductive node classification by turning node utility scores into graph scores via aggregation.

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

The paper addresses active learning when datasets contain thousands of separate graphs rather than one large graph. In these inductive settings, labeling any node in a graph yields labels for the entire graph at once, so the selection step must operate at the graph level. ALINC adapts standard node-level utility measures by applying aggregation functions that produce a single score per graph. Benchmarks across ten strategies, three aggregation methods, and four datasets identify CoreSet, TypiClust, and BADGE as the strongest performers when aggregated this way. The work further shows that the aggregation function itself drives large differences in final accuracy and total labels needed, and demonstrates the approach on molecular metabolism prediction and printed circuit board schematic design.

Core claim

ALINC bridges the existing methodological gap by elevating node-level utility measures to graph-level selection criteria through various aggregation mechanisms. In an extensive benchmark including ten strategies, three aggregation methods, and four datasets, CoreSet, TypiClust, and BADGE are identified as the top-performing graph sampling strategies. The analysis reveals that the choice of aggregation method is pivotal because it substantially affects model performance and annotation costs. The framework is shown to be effective in two concrete use cases: site-of-metabolism prediction in molecules and design automation of printed circuit board schematics.

What carries the argument

Aggregation mechanisms that combine node-level utility scores into one graph-level selection score.

If this is right

  • CoreSet, TypiClust, and BADGE become the leading strategies once node utilities are aggregated to the graph level.
  • Switching the aggregation function can produce large swings in both final model accuracy and total annotation cost.
  • The same graph-selection approach yields measurable gains on site-of-metabolism prediction tasks.
  • The same graph-selection approach yields measurable gains on printed circuit board schematic design tasks.

Where Pith is reading between the lines

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

  • The aggregation idea could be tested on other bundle-labeled domains such as protein complex prediction or scene graph annotation.
  • Different graph sizes within one dataset may require size-aware aggregation to avoid bias toward larger graphs.
  • Pairing the method with graph encoders that operate efficiently on small independent graphs could further lower labeling budgets.

Load-bearing premise

That annotating one node in a graph automatically supplies labels for every other node in the same graph.

What would settle it

A controlled experiment on a collection of independent graphs in which selecting and annotating individual nodes (without receiving the rest of the graph labels) produces higher accuracy at the same total annotation budget than any graph-level selection method.

Figures

Figures reproduced from arXiv: 2606.04647 by Andr\'e Alcalde, Bernhard Sick, Denis Huseljic, Josephine M. Thomas, Pascal Plettenberg.

Figure 1
Figure 1. Figure 1: Pairwise comparison matrix averaged across 4 datasets and 10 trials. Ele￾ment (i, j) corresponds to the proportion of total runs (in %), where strategy i yields a higher AULC compared to strategy j. For each strategy, we considered the best￾performing aggregation mechanism for the respective trial. For PATTERN and CLUSTER, performance is measured via macro-averaged accuracy, and for PascalVOC-SP and COCO-S… view at source ↗
Figure 2
Figure 2. Figure 2: Relative accuracy learning curves for different query strategies in the ALINC framework on PATTERN and CLUSTER using GatedGCN. different datasets. For example, while the purely uncertainty-based strategies, Entropy and Margin, perform poorly on PATTERN, they significantly outper￾form random sampling on CLUSTER. Overall, we observe the most consistent performance improvements over random sampling for TypiCl… view at source ↗
Figure 3
Figure 3. Figure 3: Top and middle row: Pairwise comparison matrix averaged across 4 datasets and 10 trials for all strategies. Bottom row: Pairwise comparison matrix averaged across 10 strategies and 10 trials for all datasets, and total pairwise comparison matrix across all strategies and datasets (bottom row, rightmost column). In each subplot, element (i, j) corresponds to the proportion of total runs (in %), where aggreg… view at source ↗
Figure 4
Figure 4. Figure 4: Top row: Relative learning curves of selected strategies with high sensitivity on the aggregation type. Bottom row: Selected relative learning curves of strategies with low sensitivity on the aggregation type. 2 1 0 1 2 AULC Improvement over Random (%) 110 130 150 170 Avg. Sampled Graph Size PATTERN (GatedGCN) Random Entropy Margin Graph-Density Node-Density Degree CoreSet AGE ANRMAB TypiClust BADGE 2 0 2 … view at source ↗
Figure 5
Figure 5. Figure 5: Annotation costs in terms of averaged sampled graph size against the relative AULC improvement over random sampling for different strategies and aggregation mechanisms with GatedGCN on PATTERN and CLUSTER. Some strategy-aggregation combinations are not displayed due to poor performance. sampled graphs across all AL cycles [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: a) An example molecule from the Zaretzki dataset with metabolism sites highlighted in light blue. b) An excerpt from a PCB schematic with two pull-up resistors (R4, R5) connecting the I2C signal lines to a 5V supply. 16 80 160 240 320 # Labeled Graphs 0 2 4 AUPRC Improv. (%) +5.9 % AULC SoM (GatedGCN) 320 640 1040 1440 1840 # Labeled Graphs 2.5 0.0 2.5 5.0 AUPRC Improv. (%) +2.9 % AULC PCB (GPS) Random Max… view at source ↗
Figure 7
Figure 7. Figure 7: Relative learning curves for different ALINC strategies on the Zaretzki SoM dataset using GatedGCN (left), and on the PCB pull-ups/-downs dataset using GPS (right). The strategies with the highest AULC are highlighted for better visibility. The used performance metric is the area under the precision-recall curve (AUPRC). Key Take-Away: While some strategies are biased towards selecting larger graphs, this … view at source ↗
Figure 8
Figure 8. Figure 8: Relative accuracy learning curves on PATTERN using GatedGCN with the strategies Mean-Graph-Density and Max-Node-Density for different values of k for the k-means clustering. B Additional Benchmark Results Supplementing the relative learning curves provided in the main paper, we also provide the relative learning curves for all strategies (with the best aggregation method, respectively) on PascalVOC-SP and … view at source ↗
Figure 9
Figure 9. Figure 9: Relative F1 learning curves for different query strategies in the ALINC framework on PascalVOC-SP and COCO-SP using GPS. 5 25 50 75 100 # Labeled Graphs 60 70 80 Accuracy (%) PATTERN (GatedGCN) 50 250 500 750 1000 # Labeled Graphs 40 50 60 CLUSTER (GatedGCN) Random Entropy Margin Node-Density Graph-Density Degree CoreSet AGE ANRMAB TypiClust BADGE [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Absolute accuracy learning curves for different query strategies in the ALINC framework on PATTERN and CLUSTER using GatedGCN. 100 500 1000 1500 2000 # Labeled Graphs 10 20 30 F1 Improvement (%) PascalVOC-SP (GPS) 100 500 1000 1500 2000 # Labeled Graphs 2.5 5.0 7.5 10.0 COCO-SP (GPS) Random Entropy Margin Node-Density Graph-Density Degree CoreSet AGE ANRMAB TypiClust BADGE [PITH_FULL_IMAGE:figures/full_f… view at source ↗
Figure 11
Figure 11. Figure 11: Absolute F1 learning curves for different query strategies in the ALINC framework on PascalVOC-SP and COCO-SP using GPS [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Relative accuracy learning curves for all strategies with different aggregation methods on PATTERN with GatedGCN. -4 -2 0 2 4 Acc. Improv. (%) Entropy Margin Node-Density Graph-Density Degree 50 250 500 750 1000 # Labeled Graphs -4 -2 0 2 4 Acc. Improv. (%) CoreSet 50 250 500 750 1000 # Labeled Graphs AGE 50 250 500 750 1000 # Labeled Graphs ANRMAB 50 250 500 750 1000 # Labeled Graphs TypiClust 50 250 500… view at source ↗
Figure 13
Figure 13. Figure 13: Relative accuracy learning curves for all strategies with different aggregation methods on CLUSTER with GatedGCN [PITH_FULL_IMAGE:figures/full_fig_p021_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Relative F1 learning curves for all strategies with different aggregation methods on PascalVOC-SP with GPS. -1 0 1 F1 Improv. (%) Entropy Margin Node-Density Graph-Density Degree 100 500 1000 1500 2000 # Labeled Graphs -1 0 1 F1 Improv. (%) CoreSet 100 500 1000 1500 2000 # Labeled Graphs AGE 100 500 1000 1500 2000 # Labeled Graphs ANRMAB 100 500 1000 1500 2000 # Labeled Graphs TypiClust 100 500 1000 1500 … view at source ↗
Figure 15
Figure 15. Figure 15: Relative F1 learning curves for all strategies with different aggregation methods on COCO-SP with GPS [PITH_FULL_IMAGE:figures/full_fig_p022_15.png] view at source ↗
read the original abstract

Active learning (AL) for node classification typically focuses on selecting the most informative nodes for annotation within one or a few large graphs (e.g., in social network analysis). However, in other domains, such as molecular chemistry or electronic design automation, datasets consist of thousands of independent graphs. In many of these inductive settings, annotating an individual node requires a full-graph analysis, which effectively yields the remaining node labels on-the-fly. Therefore, these scenarios require AL strategies that select entire graphs instead of single nodes, a problem which has not been tackled in the literature so far. Thus, we introduce ALINC, an AL framework for inductive node classification via graph sampling. It bridges the existing methodological gap by elevating node-level utility measures to graph-level selection criteria through various aggregation mechanisms. In an extensive benchmark including ten strategies, three aggregation methods, and four datasets, we identify CoreSet, TypiClust, and BADGE as the top-performing graph sampling strategies. Our detailed analysis further reveals that the choice of the aggregation method is pivotal, as it substantially affects model performance and annotation costs. Finally, we demonstrate the effectiveness of ALINC in two use case studies: site-of-metabolism prediction in molecules and design automation of printed circuit board schematics.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The paper introduces ALINC, an active learning framework for inductive node classification on datasets of thousands of independent graphs (e.g., molecules, PCB schematics). It argues that annotating one node requires full-graph analysis that yields all node labels on-the-fly, necessitating selection of entire graphs rather than nodes; ALINC elevates node-level utilities (CoreSet, TypiClust, BADGE, etc.) to graph level via aggregation functions. An extensive benchmark with ten strategies, three aggregations, and four datasets identifies CoreSet, TypiClust, and BADGE as top performers, shows aggregation choice is pivotal, and demonstrates effectiveness in two use-case studies.

Significance. If the motivating premise holds, the work fills a genuine gap between standard node-level AL and multi-graph inductive settings, with the benchmark providing actionable guidance on strategy and aggregation selection. The identification of aggregation method as a first-order factor is a useful empirical contribution.

major comments (2)
  1. [Abstract / Introduction] Abstract (and presumably §1): The central premise that 'annotating an individual node requires a full-graph analysis, which effectively yields the remaining node labels on-the-fly' is asserted without citation, empirical support, or domain-specific argument. This premise is load-bearing for the claimed methodological gap; if annotation in the cited domains (site-of-metabolism, PCB design) does not automatically produce the full label set, existing node-level AL strategies remain directly applicable and the motivation for graph-level selection via aggregation collapses.
  2. [Abstract / Experiments] Benchmark description (abstract): The claim that CoreSet, TypiClust, and BADGE are the top-performing graph sampling strategies rests on an 'extensive benchmark' whose experimental controls, statistical reporting, multiple-run variance, and handling of post-hoc strategy selection are not verifiable from the provided text; without these details the ranking cannot be treated as robust.
minor comments (2)
  1. [Experiments] Clarify whether the four datasets are all multi-graph inductive collections or include single-graph transductive cases, and report per-dataset statistics (number of graphs, average nodes/edges).
  2. [Method] The three aggregation mechanisms should be formally defined (e.g., mean, max, sum) with explicit equations rather than left as 'various aggregation mechanisms'.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and indicate the revisions we will incorporate.

read point-by-point responses

  1. Referee: [Abstract / Introduction] Abstract (and presumably §1): The central premise that 'annotating an individual node requires a full-graph analysis, which effectively yields the remaining node labels on-the-fly' is asserted without citation, empirical support, or domain-specific argument. This premise is load-bearing for the claimed methodological gap; if annotation in the cited domains (site-of-metabolism, PCB design) does not automatically produce the full label set, existing node-level AL strategies remain directly applicable and the motivation for graph-level selection via aggregation collapses.

    Authors: We agree that the premise requires stronger substantiation to support the claimed gap. While the manuscript is grounded in domain practices for the cited applications, we will revise the abstract and introduction to include domain-specific arguments and citations describing typical annotation workflows in molecular chemistry and PCB design. revision: yes

  2. Referee: [Abstract / Experiments] Benchmark description (abstract): The claim that CoreSet, TypiClust, and BADGE are the top-performing graph sampling strategies rests on an 'extensive benchmark' whose experimental controls, statistical reporting, multiple-run variance, and handling of post-hoc strategy selection are not verifiable from the provided text; without these details the ranking cannot be treated as robust.

    Authors: The full manuscript contains the experimental details on controls, variance across runs, and statistical reporting. To improve verifiability of the abstract claim, we will revise the abstract to briefly reference the robustness measures and direct readers to the experimental section for full controls and post-hoc handling. revision: partial

Circularity Check

0 steps flagged

Empirical benchmark study with no derivations or self-referential reductions

full rationale

The paper presents an empirical active learning benchmark for graph sampling in inductive node classification settings. No equations, derivations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text. The central methodological claim rests on a stated domain assumption about annotation costs rather than any reduction of outputs to inputs by construction. The work is self-contained as a comparative study of existing strategies (CoreSet, TypiClust, BADGE) under aggregation mechanisms, with no circular steps identified.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No mathematical derivations, free parameters, or new postulated entities are described in the abstract; the contribution is an empirical framework and benchmark.

pith-pipeline@v0.9.1-grok · 5765 in / 1065 out tokens · 24386 ms · 2026-06-28T07:05:03.708636+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

47 extracted references · 3 canonical work pages · 2 internal anchors

  1. [1]

    In: ICLR (2021)

    Alon, U., Yahav, E.: On the bottleneck of graph neural networks and its practical implications. In: ICLR (2021)

    2021

  2. [2]

    In: ICLR (2020)

    Ash, J.T., Zhang, C., Krishnamurthy, A., Langford, J., Agarwal, A.: Deep batch active learning by diverse, uncertain gradient lower bounds. In: ICLR (2020)

    2020

  3. [3]

    Bemis, G.W., Murcko, M.A.: The properties of known drugs. 1. molecular frame- works. J. Med. Chem.39(15), 2887–2893 (1996)

    1996

  4. [4]

    Residual Gated Graph ConvNets

    Bresson, X., Laurent, T.: Residual gated graph convnets. arXiv preprint arXiv:1711.07553 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  5. [5]

    Brody, S., Alon, U., Yahav, E.: How attentive are graph attention networks? In: ICLR (2022)

    2022

  6. [6]

    In: VISIGRAPP (2019)

    Brust, C.A., Käding, C., Denzler, J.: Active learning for deep object detection. In: VISIGRAPP (2019)

    2019

  7. [7]

    Active Learning for Graph Embedding

    Cai, H., Zheng, V.W., Chang, K.C.C.: Active learning for graph embedding. arXiv preprint arXiv:1705.05085 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  8. [8]

    In: ICML (2022)

    Chen, D., O’Bray, L., Borgwardt, K.: Structure-aware transformer for graph repre- sentation learning. In: ICML (2022)

    2022

  9. [9]

    JCIM64(2), 348–358 (2024)

    Chen, Y., Seidel, T., Jacob, R.A., Hirte, S., Mazzolari, A., Pedretti, A., Vistoli, G., Langer, T., Miljkovic, F., Kirchmair, J.: Active learning approach for guiding site-of-metabolism measurement and annotation. JCIM64(2), 348–358 (2024)

    2024

  10. [10]

    JMLR24(43), 1–48 (2023)

    Dwivedi, V.P., Joshi, C.K., Luu, A.T., Laurent, T., Bengio, Y., Bresson, X.: Bench- marking graph neural networks. JMLR24(43), 1–48 (2023)

    2023

  11. [11]

    In: NeurIPS (2022)

    Dwivedi, V.P., Rampášek, L., Galkin, M., Parviz, A., Wolf, G., Luu, A.T., Beaini, D.: Long range graph benchmark. In: NeurIPS (2022)

    2022

  12. [12]

    In: ICML (2024)

    Fuchsgruber, D., Wollschläger, T., Charpentier, B., Oroz, A., Günnemann, S.: Uncertainty for active learning on graphs. In: ICML (2024)

    2024

  13. [13]

    In: IJCAI (2018)

    Gao, L., Yang, H., Zhou, C., Wu, J., Pan, S., Hu, Y.: Active discriminative network representation learning. In: IJCAI (2018)

    2018

  14. [14]

    In: ICML (2022)

    Hacohen, G., Dekel, A., Weinshall, D.: Active learning on a budget: Opposite strategies suit high and low budgets. In: ICML (2022)

    2022

  15. [15]

    In: NeurIPS (2017) 16 P

    Hamilton, W., Ying, Z., Leskovec, J.: Inductive representation learning on large graphs. In: NeurIPS (2017) 16 P. Plettenberg et al

    2017

  16. [16]

    In: NeurIPS (2020)

    Hu, S., Xiong, Z., Qu, M., Yuan, X., Côté, M.A., Liu, Z., Tang, J.: Graph policy network for transferable active learning on graphs. In: NeurIPS (2020)

    2020

  17. [17]

    In: NeurIPS (2020)

    Hu, W., Fey, M., Zitnik, M., Dong, Y., Ren, H., Liu, B., Catasta, M., Leskovec, J.: Open graph benchmark: Datasets for machine learning on graphs. In: NeurIPS (2020)

    2020

  18. [18]

    In: ICLR (2019)

    Hu, W., Liu, B., Gomes, J., Zitnik, M., Liang, P., Pande, V., Leskovec, J.: Strategies for pre-training graph neural networks. In: ICLR (2019)

    2019

  19. [19]

    In: ECML PKDD (2024)

    Huseljic, D., Hahn, P., Herde, M., Rauch, L., Sick, B.: Fast fishing: Approximating bait for efficient and scalable deep active image classification. In: ECML PKDD (2024)

    2024

  20. [20]

    In: IAL @ ECML PKDD

    Huseljic, D., Herde, M., Hahn, P., Sick, B.: Role of hyperparameters in deep active learning. In: IAL @ ECML PKDD. pp. 19–24 (2023)

    2023

  21. [21]

    In: CVPR (2026)

    Huseljic, D., Herde, M., Rauch, L., Hahn, P., Sick, B.: Cleaning the pool: Progressive filtering of unlabeled pools in deep active learning. In: CVPR (2026)

    2026

  22. [22]

    In: ICLR (2017)

    Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. In: ICLR (2017)

    2017

  23. [23]

    Expert Opin

    Litsa, E.E., Das, P., Kavraki, L.E.: Machine learning models in the prediction of drug metabolism: challenges and future perspectives. Expert Opin. Drug Metab. Toxicol.17(11), 1245–1247 (2021)

    2021

  24. [24]

    TMLR (2023)

    Ma, J., Ma, Z., Chai, J., Mei, Q.: Partition-based active learning for graph neural networks. TMLR (2023)

    2023

  25. [25]

    In: ICLR (2020)

    Oono, K., Suzuki, T.: Graph neural networks exponentially lose expressive power for node classification. In: ICLR (2020)

    2020

  26. [26]

    In: ECML PKDD (2025)

    Plettenberg, P., Alcalde, A., Sick, B., Thomas, J.M.: Graph neural networks for automatic addition of optimizing components in printed circuit board schematics. In: ECML PKDD (2025)

    2025

  27. [27]

    TMLR (2025)

    Plettenberg, P., Köhler, D., Sick, B., Thomas, J.M.: Flow-attentional graph neural networks. TMLR (2025)

    2025

  28. [28]

    In: NeurIPS (2022)

    Rampášek, L., Galkin, M., Dwivedi, V.P., Luu, A.T., Wolf, G., Beaini, D.: Recipe for a general, powerful, scalable graph transformer. In: NeurIPS (2022)

    2022

  29. [29]

    In: ECML PKDD (2023)

    Rauch, L., Aßenmacher, M., Huseljic, D., Wirth, M., Bischl, B., Sick, B.: Activeglae: A benchmark for deep active learning with transformers. In: ECML PKDD (2023)

    2023

  30. [30]

    Reiser, P., Neubert, M., Eberhard, A., Torresi, L., Zhou, C., Shao, C., Metni, H., van Hoesel, C., Schopmans, H., Sommer, T., et al.: Graph neural networks for materials science and chemistry. Commun. Mater.3(1), 93 (2022)

    2022

  31. [31]

    KBS 21(7), 727–739 (2008)

    Rodriguez, M.A.: Grammar-based random walkers in semantic networks. KBS 21(7), 727–739 (2008)

    2008

  32. [32]

    In: BMVC (2018)

    Roy, S., Unmesh, A., Namboodiri, V.P.: Deep active learning for object detection. In: BMVC (2018)

    2018

  33. [33]

    In: ICLR (2018)

    Sener, O., Savarese, S.: Active learning for convolutional neural networks: A core-set approach. In: ICLR (2018)

    2018

  34. [34]

    Technical Report, University of Wisconsin-Madison, Department of Computer Sciences (2009)

    Settles, B.: Active learning literature survey. Technical Report, University of Wisconsin-Madison, Department of Computer Sciences (2009)

    2009

  35. [35]

    In: ICML (2023)

    Shirzad, H., Velingker, A., Venkatachalam, B., Sutherland, D.J., Sinop, A.K.: Exphormer: Sparse transformers for graphs. In: ICML (2023)

    2023

  36. [36]

    JCIM59(8), 3400–3412 (2019)

    Sicho, M., Stork, C., Mazzolari, A., de Bruyn Kops, C., Pedretti, A., Testa, B., Vistoli, G., Svozil, D., Kirchmair, J.: Fame 3: predicting the sites of metabolism in synthetic compounds and natural products for phase 1 and phase 2 metabolic enzymes. JCIM59(8), 3400–3412 (2019)

    2019

  37. [37]

    In: NeurIPS (2023) Active Learning for Inductive Node Classification via Graph Sampling 17

    Song,Z.,Zhang,Y.,King,I.:Nochange,nogain:Empoweringgraphneuralnetworks with expected model change maximization for active learning. In: NeurIPS (2023) Active Learning for Inductive Node Classification via Graph Sampling 17

    2023

  38. [38]

    TMLR (2024)

    Tönshoff, J., Ritzert, M., Rosenbluth, E., Grohe, M.: Where did the gap go? reassessing the long-range graph benchmark. TMLR (2024)

    2024

  39. [39]

    In: ICLR (2018)

    Veličković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph attention networks. In: ICLR (2018)

    2018

  40. [40]

    NeurIPS33, 5776–5788 (2020)

    Wang, W., Wei, F., Dong, L., Bao, H., Yang, N., Zhou, M.: Minilm: Deep self- attention distillation for task-agnostic compression of pre-trained transformers. NeurIPS33, 5776–5788 (2020)

    2020

  41. [41]

    arXiv preprint arXiv:1910.07567 (2019)

    Wu, Y., Xu, Y., Singh, A., Yang, Y., Dubrawski, A.: Active learning for graph neural networks via node feature propagation. arXiv preprint arXiv:1910.07567 (2019)

    work page arXiv 1910

  42. [42]

    IEEE TBD8(4), 920–932 (2022)

    Xie, Y., Lv, S., Qian, Y., Wen, C., Liang, J.: Active and semi-supervised graph neural networks for graph classification. IEEE TBD8(4), 920–932 (2022)

    2022

  43. [43]

    Xu, K., Hu, W., Leskovec, J., Jegelka, S.: How powerful are graph neural networks? In: ICLR (2019)

    2019

  44. [44]

    Ying, C., Cai, T., Luo, S., Zheng, S., Ke, G., He, D., Shen, Y., Liu, T.Y.: Do transformers really perform badly for graph representation? In: NeurIPS (2021)

    2021

  45. [45]

    JCIM53(12), 3373–3383 (2013)

    Zaretzki, J., Matlock, M., Swamidass, S.J.: Xenosite: accurately predicting cyp- mediated sites of metabolism with neural networks. JCIM53(12), 3373–3383 (2013)

    2013

  46. [46]

    In: ICLR (2022)

    Zhang, W., Wang, Y., You, Z., Cao, M., Huang, P., Shan, J., Yang, Z., Cui, B.: Information gain propagation: a new way to graph active learning with soft labels. In: ICLR (2022)

    2022

  47. [47]

    In: VLDB (2021) 18 P

    Zhang, W., Yang, Z., Wang, Y., Shen, Y., Li, Y., Wang, L., Cui, B.: Grain: Improving data efficiency of graph neural networks via diversified influence maximization. In: VLDB (2021) 18 P. Plettenberg et al. A Number of Clusters in Density-based Strategies Density-based acquisition strategies performk-meansclustering to calculate the information density. W...

    2021