Pith. sign in

REVIEW 3 major objections 4 minor 63 references

HGCN(O): A Self-Tuning GCN HyperModel Toolkit for Outcome Prediction in Event-Sequence Data

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A self-tuning toolkit of graph convolutional networks claims perfect accuracy on balanced process logs and weighted F1 near 0.87 on skewed event-sequence data.

desk verdict The toolkit is real, but the experiments don't support it: no baselines, and the perfect balanced-data scores look like label leakage from including the final activity in the graph. read the letter →

arxiv 2507.22524 v3 pith:JFXZ74UC submitted 2025-07-30 cs.LG

Fang Wang , Paolo Ceravolo , Ernesto Damiani This is my paper
classification cs.LG
keywords BusinessProcessMonitoringGraphConvolutionalNetworksOutcome-OrientedEvent-SequencePredictionHyperModelpredictiveself-tuninghyperparameteroptimizationeventlogsoutcome
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

The paper proposes HGCN(O), a self-tuning toolkit that encodes event sequences as graphs and trains graph convolutional networks to predict each sequence's outcome, targeting predictive business process monitoring (PBPM). Its claim is that combining hierarchical node- and graph-level attribute encoding, time-gap edge weights, and automated hyperparameter search lets eight GCN variants adapt to both heavily imbalanced and perfectly balanced datasets without manual tuning. On the six-class imbalanced Patients benchmark the paper reports weighted F1 scores near 0.86–0.87, with GCNConv models the most stable and two-level or embedding inputs the strongest on minority classes; on the balanced BPIC12-A and BPIC12-O benchmarks every model reaches perfect precision, recall, and F1. A reader should care because outcome prediction from event logs is a standing problem in process monitoring, and the toolkit offers a unified entry point plus a benchmark for future graph-based approaches.

What carries the argument

The load-bearing object is the trace-as-path-graph encoding: one node per event holding a composite feature vector, one directed edge from each event to its successor, and an edge weight equal to the normalized time gap between the two events' start times, $w(i \to i+1) = T^s_{i+1} - T^s_i$. On top of this encoding, the self-tuning loop — an automated Bayesian search over layer counts, hidden units, activations, pooling, skip connections, optimizers, schedulers, and regularizers — selects a configuration for each dataset. The four architectures differ only in where attributes enter: O-GCN folds graph-level attributes into every node, T-GCN keeps separate node and graph streams, TP-GCN adds a TF-IDF duration pseudo-embedding, and TE-GCN embeds the key activity attribute. The mechanism as a whole is a combinatorial family of input pipelines, layer types, and hyperparameter settings, screened automatically rather than fixed by a designer.

What would settle it

Re-run the balanced experiments with the final event's node and activity attribute excluded or masked to a padding value: if accuracy stays at 1.0, the perfect scores reflect real learning, and if it drops, the label was readable from the input. A complementary check runs a non-graph baseline such as a random forest on prefix encodings or an LSTM over the same traces and splits, testing the claimed superiority over traditional approaches.

Watch

Extended reading notes

Core claim

The central claim is that outcome prediction on event-sequence data can be solved as graph classification, and that making the pipeline self-tuning is what lets graph models adapt to dataset imbalance. For each trace the paper builds a path graph: nodes are events carrying concatenated attribute vectors $v_{N_i} = [A_i, B_i, U_i]$, edges connect consecutive events, and edge weights $w(i \to i+1) = T^s_{i+1} - T^s_i$ record the elapsed time between event starts. Four architectures — O-GCN, T-GCN, TP-GCN, TE-GCN — determine how graph-level attributes, TF-IDF duration pseudo-embeddings, and activity embeddings enter the network, and each is run with both GCNConv and GraphConv layers, giving eight hypermodels. An automated search over network depth, units, activations, pooling, skip connections, optimizers, schedulers, and regularizers configures each model per dataset. The paper reports that all eight hypermodels reach perfect precision, recall, and F1 on the balanced BPIC12 benchmarks, while on the imbalanced six-class Patients benchmark they reach weighted F1 near 0.86–0.87, with GCNConv variants the most stable and two-level or embedding inputs the strongest on minority classes.

Load-bearing premise

The load-bearing premise is that the evaluation measures genuine prediction: each trace's label is its final activity, the graph input contains every event in the trace, and the paper does not state that the final event is removed or masked, so the perfect balanced-data scores could in principle come from reading the label off the input rather than from learned process dynamics.

Editorial extensions

If this is right

  • Applying the toolkit to a new event log requires no manual hyperparameter tuning: the automated search adapts architecture depth, units, activations, pooling, and regularizers to the dataset's class balance and attribute structure.
  • Encoding elapsed time between consecutive event starts as edge weights lets the models represent single-event durations and inter-event gaps, which the paper argues supports outcome prediction for processes with concurrent or overlapping activities.
  • The results yield concrete model-selection guidance: two-level and embedding variants improve minority-class recall, T-GCN variants favor precision, and GCNConv layers give more stable training on skewed data.
  • Because every model scores perfectly on the balanced datasets, the published hyperparameter matrices stand as a reference for how architecture choices affect training speed, convergence, and stability once accuracy differences are removed.

Reading between the lines

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

  • The balanced-data protocol defines the label as the trace's final activity while the graph contains every event, so the reported 1.0 accuracy is consistent with the model reading the label off the input; masking the final event's attribute would separate that possibility from genuine learning.
  • The same trace-to-path-graph encoding and tuning loop transfer to other event-sequence domains, such as clinical careflows or user-session logs, since nothing in the encoding is specific to business processes.
  • Adding an external baseline — a sequence model like an LSTM or a classical tree ensemble on the same splits — would turn the internal model-to-model comparison into the external benchmark the abstract's claim of superiority over traditional approaches assumes but the experiments do not directly supply.
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.

Referee Report

3 major / 4 minor

Summary. The paper proposes HGCN(O), a self-tuning toolkit for outcome prediction in event-sequence data in Predictive Business Process Monitoring. It defines four GCN architectures (O-GCN, T-GCN, TP-GCN, TE-GCN) instantiated over two convolutional layers (GCNConv and GraphConv), giving eight hypermodels, and tunes their hyperparameters with Optuna. The experiments cover one imbalanced dataset (Patients) and two balanced datasets (BPI12-A and BPI12-O). The paper reports that GCNConv models are best on unbalanced data, that all models achieve perfect accuracy on balanced data, and that HGCN(O) outperforms traditional approaches. The central evidence for these claims is, however, undermined by a likely label-leakage artifact in the balanced-data setup and by the complete absence of non-GCN baseline comparisons.

Significance. If the results were valid, the paper would provide a useful benchmark-style toolkit for graph-based outcome prediction, with a welcome emphasis on automated hyperparameter search and on comparing multiple graph-input designs. The paper deserves credit for reporting the tuned hyperparameter configurations in detail, for making the dependence on edge weights explicit, and for covering several architectural variants in one framework. However, the balanced-data perfect scores are explained more simply by the label being recoverable from the input, and the claimed superiority over traditional approaches is not tested at all. As it stands, the paper does not establish its headline claims, so its significance is contingent on a substantial reworking of the experimental protocol.

major comments (3)
  1. [6.1 and 4.3] The balanced-data evaluation appears to contain label leakage. Section 6.1 states that each BPI12-A/O trace is classified into one of three outcomes 'corresponding to the final activity in the trace.' Section 4.3 constructs the graph from all event nodes, and Section 6.1 decomposes the activity attribute into verb and description components that are part of the node vectors. No step in Sections 6.1 or 6.2 masks the final activity, drops its node, or truncates the trace to a prefix. Since the readout is a pooled representation of all nodes (Table 1 allows max pooling), a model can essentially copy the label from the final activity's features. The reported 1.0 precision, recall, and F1 in Tables 4 and 5 are therefore the expected behavior of a model reading the last event, not evidence that the toolkit predicts future outcomes.
  2. [7 and Abstract] The abstract claims 'superior performance of HGCN(O) over traditional approaches,' but Section 7 contains no comparison with any non-GCN method. Tables 2–5 compare only the eight proposed GCN/GraphConv variants. There is no Random Forest, XGBoost, SVM, LSTM, or other PBPM baseline, so the superiority claim is unsupported by the experiments reported in the paper. The introduction's discussion of classical and deep-learning alternatives makes this comparison particularly necessary.
  3. [5.5 and 6.2] The model-selection protocol is ambiguous and potentially circular. Section 5.5 states that models are ranked by 'test accuracy' for balanced datasets and by weighted F1 for imbalanced datasets, while Section 6.2 describes only an 80/20 train/validation split and does not mention a separate held-out test set. If the same split is used both for hyperparameter selection and for the reported results, the reported numbers are optimistically biased. The authors should clarify whether the selection metric is computed on a validation set and should evaluate the final models on an independent test set.
minor comments (4)
  1. [Algorithm 1] Line 14 of Algorithm 1 contains an apparent typo: 'if any f ˚b ≈ any f ˚b' compares a quantity with itself; the intended comparison with f' b or f b should be stated.
  2. [Equation (1)] In Equation (1), the symbol G(l) is used before it is defined; the surrounding text defines G(l)= (V(l),E,W) only after the equation, which makes the notation hard to follow.
  3. [Tables 3–5] The hyperparameter matrices in Tables 3–5 are very difficult to parse: the columns are not clearly separated, many abbreviations are defined only in long footnotes, and the F1 column appears to contain a tuple of loss standard deviation and accuracy without a clear header. A cleaner layout would improve reproducibility.
  4. [General] No code or data availability statement is provided. Given the paper's toolkit orientation, making the implementation available would substantially strengthen reproducibility.

Circularity Check

2 steps flagged · score 8.0 of 10

Balanced-data perfect accuracy is forced by construction: the outcome is defined as the final activity while the graph input contains every event node with its activity features, so the label is recoverable from the input; model selection also tunes on test metrics.

  1. self definitional [Section 4.1, 4.3 and 6.1]
    "we conventionally refer to such attribute as “activity”, and denote it for each node as Ai... For each graph node N i, we concatenate all encoded attributes into a unified composite vector, denoted as: vNi = [A i, Bi, Ui]. ... The matrix of node vectors for graph G j is defined as V N(G j ) ∈ R n×dN, where V N(G j ) = [vN1, . . . ,vNn]⊺. ... Each trace is classified into one of three outcomes — “approved (accepted)”, “declined” or “canceled” - corresponding to the final activity in the trace."

    The balanced-data label is defined as the final activity in the trace. The graph object constructed in Section 4.3 contains all n event nodes, and each node vector contains the activity attribute Ai (later decomposed into verb and description). No truncation, masking, or exclusion of the final event is described in Sections 6.1 or 6.2. A graph readout that pools over all nodes can therefore recover the identity of the final activity from the node-feature matrix, making the reported 1.0 precision/recall/F1 on BPI12-A and BPI12-O (Tables 4-5) a consequence of the data encoding rather than evidence of predictive learning. The central claim that 'all models perform consistently on balanced data' is thus circular with respect to the outcome definition.

  2. fitted input called prediction [Section 5.5 (hyperparameter selection; contrasted with Section 6.2 setup)]
    "For balanced datasets, models were ranked according to test accuracy, with ties broken by the standard deviation of test loss to ensure stability; if both metrics were equal, the model with the lowest test loss was selected."

    The hyperparameter search objective is test accuracy/test loss, while Section 6.2 describes only an 80/20 train/validation split for tuning. As written, the evaluation set is used in the selection criterion, so the reported predictions are not independent of the model-selection process: the test metrics have already influenced which model and hyperparameters were chosen. This makes the reported performance a fitted quantity rather than a genuinely out-of-sample prediction.

full rationale

The core graph-convolution machinery and the four HGCN(O) architectures are internally self-contained: Equations (1)-(2) are standard GCNConv/GraphConv definitions, and the architectural variations (O, T, TP, TE) are described without relying on a circular derivation. The self-citations [7] and [59] supply preprocessing conventions (activity decomposition, duration binning) and are not load-bearing for the mathematical claims. However, the balanced-data experimental claim is circular by construction: Section 6.1 defines each outcome as the final activity, and Section 4.3 places every event node, including the encoded attributes of the final activity, into the input graph, with no masking step specified. Perfect scores on BPI12-A/O therefore do not demonstrate that the toolkit predicts future outcomes; they are consistent with the model reading the label back from the input. The additional use of test-set metrics as the model-selection criterion in Section 5.5 further compromises the claim that the reported results are out-of-sample predictions. The abstract's statement that HGCN(O) is superior to traditional approaches is not backed by any non-GCN baseline in Section 7, but that is an evidence gap rather than a circularity. Overall score 8: the balanced-data result is effectively forced by definition and input construction.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central claims rest on several fitted choices: hyperparameters tuned by Optuna (which themselves are selected using test metrics), hand-set pseudo-embedding binning, and an ad hoc activity decomposition. The graph representation and the label definition are domain assumptions, with the label definition creating the likely leakage on balanced data.

free parameters (4)
  • Pseudo-embedding duration binning cutoff and bin count (T_cut, N_b) = T_cut = 5 minutes; N_b = 24 for Patients dataset
    Chosen by hand in Section 6.1 to construct the TF-IDF duration pseudo-embedding used by TP-GCN.
  • Optuna search ranges for GCN/dense layers, units, activations, optimizer, learning rate, and schedulers = 200 trials over ranges in Table 1; best values in Tables 3, 4, and 5
    Tuned to the data; the central performance claims depend on these fitted hyperparameters.
  • Model selection metric (test accuracy for balanced, weighted F1 for imbalanced) = Chosen per dataset balance in Section 5.5
    Selection on test metrics is a fitted criterion that bakes the evaluation set into the final model choice.
  • Activity split into verb and description components = Not specified beyond reference [59]
    Ad hoc encoding choice from the authors' prior work, applied to all datasets in Section 6.1.
assumptions (4)
  • domain assumption Event sequences are faithfully represented as path graphs with edges between consecutive events and edge weights equal to start-time differences.
    Section 4.3 constructs graphs this way; the validity of the whole approach depends on this representation preserving the information needed for outcome prediction.
  • ad hoc to paper The outcome label is correctly defined by the final activity of the trace.
    Section 6.1 states labels 'corresponding to the final activity in the trace'. If the final activity is also an input node, the task becomes trivial; this premise is load-bearing for the balanced-data perfect scores.
  • standard math GCNConv and GraphConv operations as implemented in PyTorch Geometric correctly compute the weighted normalization described in Section 3.
    The paper relies on Eqns (1) and (2) and the PyTorch Geometric implementations [51,52] without re-deriving them.
  • domain assumption The 80/20 train/validation split produces representative class distributions after per-class stratified splitting.
    Section 6.2 describes this split; the reliability of the reported metrics depends on this assumption, especially for the 5-instance class 2.
invented entities (1)
  • Pseudo-embedding duration bin matrix (Algorithm 1)
    purpose: Provides an extra node feature space for TP-GCN built from duration bins and TF-IDF term weighting.
    It is an engineered representation with no falsifiable handle outside the downstream task; its only justification is the observed accuracy, which is entangled with the evaluation issues.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HGCN(O): A Self-Tuning GCN HyperModel Toolkit for Outcome Prediction in Event-Sequence Data." pith.science (2026). https://pith.science/paper/JFXZ74UC

@misc{pith2026250722524,
  author       = {Pith},
  title        = {Pith review of: HGCN(O): A Self-Tuning GCN HyperModel Toolkit for Outcome Prediction in Event-Sequence Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JFXZ74UC}},
  note         = {Machine review of arXiv:2507.22524}
}
read the original abstract

We propose HGCN(O), a self-tuning toolkit using Graph Convolutional Network (GCN) models for event sequence prediction. Featuring four GCN architectures (O-GCN, T-GCN, TP-GCN, TE-GCN) across the GCNConv and GraphConv layers, our toolkit integrates multiple graph representations of event sequences with different choices of node- and graph-level attributes and in temporal dependencies via edge weights, optimising prediction accuracy and stability for balanced and unbalanced datasets. Extensive experiments show that GCNConv models excel on unbalanced data, while all models perform consistently on balanced data. Experiments also confirm the superior performance of HGCN(O) over traditional approaches. Applications include Predictive Business Process Monitoring (PBPM), which predicts future events or states of a business process based on event logs.

Figures

Figures reproduced from arXiv: 2507.22524 by the authors.

Figure 1
Figure 1. Learning Curves of GCN Models Applied on Imbalanced Dataset [PITH_FULL_IMAGE:figures/full_fig_p020_1.png] view at source ↗
Figure 2
Figure 2. Learning Curve Of GCN HyperModels on Balanced Datasets [PITH_FULL_IMAGE:figures/full_fig_p028_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 55 canonical work pages

  1. [1]

    Letham, C.Rudin, D.Madigan, Sequentialeventprediction, Machine learning 93 (2013) 357–380

    B. Letham, C.Rudin, D.Madigan, Sequentialeventprediction, Machine learning 93 (2013) 357–380

    work page 2013

  2. [2]

    Ceravolo, M

    P. Ceravolo, M. Comuzzi, J. De Weerdt, C. Di Francescomarino, F. M. Maggi, Predictive process monitoring: concepts, challenges, and future research directions, Process Science 1 (2024) 1–22

    work page 2024

  3. [3]

    F. M. Maggi, C. Di Francescomarino, M. Dumas, C. Ghidini, Predictive monitoring of business processes, in: Advanced Information Systems Engineering: 26th International Conference, CAiSE 2014, Thessaloniki, Greece, June 16-20, 2014. Proceedings 26, Springer, 2014, pp. 457–472

    work page 2014

  4. [4]

    A. Pika, W. M. van der Aalst, M. T. Wynn, C. J. Fidge, A. H. ter Hofstede, Evaluating and predicting overall process risk using event logs, Information Sciences 352 (2016) 98–120

    work page 2016

  5. [5]

    Teinemaa, M

    I. Teinemaa, M. Dumas, M. L. Rosa, F. M. Maggi, Outcome-oriented predictive process monitoring: Review and benchmark, ACM Transac- tions on Knowledge Discovery from Data (TKDD) 13 (2019) 1–57. 30

    work page 2019

  6. [6]

    Pasquadibisceglie, A

    V. Pasquadibisceglie, A. Appice, G. Castellano, D. Malerba, G. Mod- ugno, Orange: outcome-oriented predictive process monitoring based on image encoding and cnns, IEEE Access 8 (2020) 184073–184086

    work page 2020

  7. [7]

    Ceravolo, S

    P. Ceravolo, S. Barbon, E. Damiani, W. Van der Aalst, Tuning machine learning to address process mining requirements, IEEE Access (2024)

    work page 2024

  8. [8]

    Leontjeva, R

    A. Leontjeva, R. Conforti, C. Di Francescomarino, M. Dumas, F. M. Maggi, Complex symbolic sequence encodings for predictive monitor- ing of business processes, in: Business Process Management: 13th International Conference, BPM 2015, Innsbruck, Austria, August 31– September 3, 2015, Proceedings 13, Springer, 2015, pp. 297–313

    work page 2015

Show all 63 references
  1. [9]

    Senderovich, C

    A. Senderovich, C. Di Francescomarino, C. Ghidini, K. Jorbina, F. M. Maggi, Intra and inter-case features in predictive process monitoring: A tale of two dimensions, in: Business Process Management: 15th International Conference, BPM 2017, Barcelona, Spain, September 10– 15, 2...

    2017

  2. [10]

    Grigori, F

    D. Grigori, F. Casati, U. Dayal, M.-C. Shan, Improving business process quality through exception understanding, prediction, and prevention, in: VLDB, volume 1, 2001, pp. 159–168

    2001

  3. [11]

    Grigori, F

    D. Grigori, F. Casati, M. Castellanos, U. Dayal, M. Sayal, M.-C. Shan, Business process intelligence, Computers in industry 53 (2004) 321–343

    2004

  4. [12]

    Castellanos, N

    M. Castellanos, N. Salazar, F. Casati, U. Dayal, M.-C. Shan, Pre- dictive business operations management, in: Databases in Networked Information Systems: 4th International Workshop, DNIS 2005, Aizu- Wakamatsu, Japan, March 28-30, 2005. Proceedings 4, Springer, 2005, pp. 1–14

    2005

  5. [13]

    B. Kang, D. Kim, S.-H. Kang, Periodic performance prediction for real-time business process monitoring, Industrial Management & Data Systems 112 (2012) 4–23

    2012

  6. [14]

    Hinkka, T

    M. Hinkka, T. Lehto, K. Heljanko, A. Jung, Classifying process in- stances using recurrent neural networks, in: Business Process Manage- ment Workshops: BPM 2018 International Workshops, Sydney, NSW, Australia, September 9-14, 2018, Revised Papers 16, Springer, 2019, pp. 313–324. 31

    2018

  7. [15]

    Kratsch, J

    W. Kratsch, J. Manderscheid, M. Röglinger, J. Seyfried, Machine learn- ing in business process monitoring: a comparison of deep learning and classical approaches used for outcome prediction, Business & Informa- tion Systems Engineering 63 (2021) 261–276

    2021

  8. [16]

    J. Wang, D. Yu, C. Liu, X. Sun, Outcome-oriented predictive pro- cess monitoring with attention-based bidirectional lstm neural networks, in: 2019 IEEE international conference on web services (ICWS), IEEE, 2019, pp. 360–367

    2019

  9. [17]

    Bellandi, P

    V. Bellandi, P. Ceravolo, S. Maghool, S. Siccardi, Graph embeddings in criminal investigation: towards combining precision, generalization and transparency: special issue on computational aspects of network science, World Wide Web 25 (2022) 2379–2402

    2022

  10. [18]

    Khemani, S

    B. Khemani, S. Patil, K. Kotecha, S. Tanwar, A review of graph neu- ral networks: concepts, architectures, techniques, challenges, datasets, applications, and future directions, Journal of Big Data 11 (2024) 18

    2024

  11. [19]

    J. B. Lee, R. A. Rossi, S. Kim, N. K. Ahmed, E. Koh, Attention models in graphs: A survey, ACM Transactions on Knowledge Discovery from Data (TKDD) 13 (2019) 1–25

    2019

  12. [20]

    Chen, Y.-C

    F. Chen, Y.-C. Wang, B. Wang, C.-C. J. Kuo, Graph representation learning: a survey, APSIPA Transactions on Signal and Information Processing 9 (2020) e15

    2020

  13. [21]

    F. Xia, K. Sun, S. Yu, A. Aziz, L. Wan, S. Pan, H. Liu, Graph learning: A survey, IEEE Transactions on Artificial Intelligence 2 (2021) 109–127

    2021

  14. [22]

    Azzini, S

    A. Azzini, S. Barbon Jr, V. Bellandi, T. Catarci, P. Ceravolo, P. Cudré- Mauroux, S. Maghool, J. Pokorny, M. Scannapieco, F. Sedes, et al., Advances in data management in the big data era, in: Advancing Re- search in Information and Communication Technology: IFIP’s Exciting Fi...

    2021

  15. [23]

    Nikolentzos, G

    G. Nikolentzos, G. Siglidis, M. Vazirgiannis, Graph kernels: A survey, Journal of Artificial Intelligence Research 72 (2021) 943–1027. 32

    2021

  16. [24]

    Gärtner, P

    T. Gärtner, P. Flach, S. Wrobel, On graph kernels: Hardness results and efficient alternatives, in: Learning Theory and Kernel Machines: 16th Annual Conference on Learning Theory and 7th Kernel Workshop, COLT/Kernel 2003, Washington, DC, USA, August 24-27, 2003. Pro- ceedings,...

    2003

  17. [25]

    S. Cao, W. Lu, Q. Xu, Grarep: Learning graph representations with global structural information, in: Proceedings of the 24th ACM interna- tional on conference on information and knowledge management, 2015, pp. 891–900

    2015

  18. [26]

    Zhang, Y

    J. Zhang, Y. Dong, Y. Wang, J. Tang, M. Ding, Prone: Fast and scalable network representation learning., in: IJCAI, volume 19, 2019, pp. 4278–4284

    2019

  19. [27]

    Perozzi, R

    B. Perozzi, R. Al-Rfou, S. Skiena, Deepwalk: Online learning of social representations, in: Proceedings of the 20th ACM SIGKDD interna- tional conference on Knowledge discovery and data mining, 2014, pp. 701–710

    2014

  20. [28]

    A.Grover, J.Leskovec, node2vec: Scalablefeaturelearningfornetworks, in: Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining, 2016, pp. 855–864

    2016

  21. [29]

    J. Tang, M. Qu, M. Wang, M. Zhang, J. Yan, Q. Mei, Line: Large-scale information network embedding, in: Proceedings of the 24th interna- tional conference on world wide web, 2015, pp. 1067–1077

    2015

  22. [30]

    Bellandi, P

    V. Bellandi, P. Ceravolo, S. Maghool, M. Pindaro, S. Siccardi, Corre- lation and pattern detection in event networks, in: 2021 IEEE Interna- tional Conference on Big Data (Big Data), IEEE, 2021, pp. 4103–4112

    2021

  23. [31]

    Scarselli, M

    F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, G. Monfardini, The graph neural network model, IEEE transactions on neural networks 20 (2008) 61–80

    2008

  24. [32]

    Zhang, A

    C. Zhang, A. Swami, N. V. Chawla, Shne: Representation learning for semantic-associated heterogeneous networks, in: Proceedings of the twelfth ACM international conference on web search and data mining, 2019, pp. 690–698. 33

    2019

  25. [33]

    J. Wang, V. W. Zheng, Z. Liu, K. C.-C. Chang, Topological recurrent neural network for diffusion prediction, in: 2017 IEEE international conference on data mining (ICDM), IEEE, 2017, pp. 475–484

    2017

  26. [34]

    D. Wang, P. Cui, W. Zhu, Structural deep network embedding, in: Pro- ceedings of the 22nd ACM SIGKDD international conference on Knowl- edge discovery and data mining, 2016, pp. 1225–1234

    2016

  27. [35]

    K. Tu, P. Cui, X. Wang, F. Wang, W. Zhu, Structural deep embedding for hyper-networks, in: Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018

    2018

  28. [36]

    D. Q. Nguyen, T. D. Nguyen, D. Phung, Universal graph transformer self-attention networks, in: Companion Proceedings of the Web Confer- ence 2022, 2022, pp. 193–196

    2022

  29. [37]

    S. Yao, T. Wang, X. Wan, Heterogeneous graph transformer for graph- to-sequence learning, in: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 2020, pp. 7145–7154

    2020

  30. [38]

    V. Shiv, C. Quirk, Novel positional encodings to enable tree-based trans- formers, Advances in neural information processing systems 32 (2019)

    2019

  31. [39]

    H. Peng, G. Li, Y. Zhao, Z. Jin, Rethinking positional encoding in tree transformer for code representation, in: Proceedings of the 2022 Confer- ence on Empirical Methods in Natural Language Processing, 2022, pp. 3204–3214

    2022

  32. [40]

    M. S. Hussain, M. J. Zaki, D. Subramanian, Global self-attention as a replacement for graph convolution, in: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2022, pp. 655–665

    2022

  33. [41]

    H. Zeng, H. Zhou, A. Srivastava, R. Kannan, V. Prasanna, Graph- saint: Graph sampling based inductive learning method, arXiv preprint arXiv:1907.04931 (2019)

    2019 arXiv

  34. [42]

    Jiang, P

    H. Jiang, P. Cao, M. Xu, J. Yang, O. Zaiane, Hi-gcn: A hierarchical graph convolution network for graph embedding learning of brain net- workandbraindisordersprediction, ComputersinBiologyandMedicine 127 (2020) 104096. 34

    2020

  35. [43]

    Pasquadibisceglie, A

    V. Pasquadibisceglie, A. Appice, G. Castellano, D. Malerba, Using con- volutional neural networks for predictive process analytics, in: 2019 international conference on process mining (ICPM), IEEE, 2019, pp. 129–136

    2019

  36. [44]

    Chiorrini, C

    A. Chiorrini, C. Diamantini, A. Mircoli, D. Potena, Exploiting instance graphs and graph neural networks for next activity prediction, in: In- ternational conference on process mining, Springer, 2021, pp. 115–126

    2021

  37. [45]

    Y. Deng, J. Wang, C. Wang, C. Zheng, M. Li, B. Li, Enhancing predic- tive process monitoring with sequential graphs and trace attention, in: 2024 IEEE International Conference on Web Services (ICWS), IEEE, 2024, pp. 406–415

    2024

  38. [46]

    Bellandi, S

    V. Bellandi, S. Montanelli, D. Shlyk, S. Siccardi, Using graph neural networks for heterogeneous event classification, volume 3741, 2024, p. 247 – 259

    2024

  39. [47]

    Philipp, R

    P. Philipp, R. X. M. Georgi, J. Beyerer, S. Robert, Analysis of con- trol flow graphs using graph convolutional neural networks, in: 2019 6th International Conference on Soft Computing & Machine Intelligence (ISCMI), IEEE, 2019, pp. 73–77

    2019

  40. [48]

    Rama-Maneiro, J

    E. Rama-Maneiro, J. C. Vidal, M. Lama, Embedding graph convolu- tional networks in recurrent neural networks for predictive monitoring, IEEE Transactions on Knowledge and Data Engineering 36 (2023) 137– 151

    2023

  41. [49]

    T. N. Kipf, M. Welling, Semi-supervised classification with graph con- volutional networks, arXiv preprint arXiv:1609.02907 (2016)

    2016 arXiv

  42. [50]

    Morris, M

    C. Morris, M. Ritzert, M. Fey, W. L. Hamilton, J. E. Lenssen, G. Rat- tan, M. Grohe, Weisfeiler and leman go neural: Higher-order graph neural networks, in: Proceedings of the AAAI conference on artificial intelligence, volume 33, 2019, pp. 4602–4609

    2019

  43. [51]

    M. Fey, J. E. Lenssen, Fast graph representation learning with pytorch geometric, arXiv preprint arXiv:1903.02428 (2019). 35

    2019 arXiv

  44. [52]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, et al., Pytorch: An imper- ative style, high-performance deep learning library, Advances in neural information processing systems 32 (2019)

    2019

  45. [53]

    Hamilton, Z

    W. Hamilton, Z. Ying, J. Leskovec, Inductive representation learning on large graphs, Advances in neural information processing systems 30 (2017)

    2017

  46. [54]

    Chiang, X

    W.-L. Chiang, X. Liu, S. Si, Y. Li, S. Bengio, C.-J. Hsieh, Cluster-gcn: An efficient algorithm for training deep and large graph convolutional networks, in: Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 2019, pp. 257–266

    2019

  47. [55]

    Y. Wang, Y. Sun, Z. Liu, S. E. Sarma, M. M. Bronstein, J. M. Solomon, Dynamic graph cnn for learning on point clouds, ACM Transactions on Graphics (tog) 38 (2019) 1–12

    2019

  48. [56]

    Veličković, G

    P. Veličković, G. Cucurull, A. Casanova, A. Romero, P. Lio, Y. Bengio, Graph attention networks, arXiv preprint arXiv:1710.10903 (2017)

    2017 arXiv

  49. [57]

    B.-H. Kim, J. C. Ye, Understanding graph isomorphism network for rs-fmri functional connectivity analysis, Frontiers in neuroscience 14 (2020) 630

    2020

  50. [58]

    Isufi, F

    E. Isufi, F. Gama, A. Ribeiro, Edgenets: Edge varying graph neural net- works, IEEE Transactions on Pattern Analysis and Machine Intelligence 44 (2021) 7457–7473

    2021

  51. [59]

    F. Wang, P. Ceravolo, E. Damiani, Comprehensive attribute encoding and dynamic lstm hypermodels for outcome oriented predictive business process monitoring, arXiv preprint arXiv:2506.03696 (2025)

    2025 arXiv

  52. [60]

    Van Dongen, Bpi challenge 2012, 2012

    B. Van Dongen, Bpi challenge 2012, 2012. URL:https: //data.4tu.nl/articles/_/12689204/1. doi:10.4121/UUID: 3926DB30-F712-4394-AEBC-75976070E91F

    2012 doi

  53. [61]

    Akiba, S

    T. Akiba, S. Sano, T. Yanase, T. Ohta, M. Koyama, Optuna: A next- generation hyperparameter optimization framework, in: The 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2019, pp. 2623–2631. 36

    2019

  54. [62]

    Z. Li, S. Arora, An exponential learning rate schedule for deep learning, arXiv preprint arXiv:1910.07454 (2019)

    2019 arXiv

  55. [63]

    L. N. Smith, Cyclical learning rates for training neural networks, in: 2017 IEEE winter conference on applications of computer vision (WACV), IEEE, 2017, pp. 464–472. Statement During the preparation of this work, the authors used ChatGPT (Ope- nAI) for language refinement and...

    2017

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.