pith. sign in

arxiv: 2606.18726 · v1 · pith:LG3C6VKWnew · submitted 2026-06-17 · 💻 cs.LG · cs.AI

Graph Grounded Cross Attention Transformer Neural Network for Structurally Constrained Full Event Sequence Generation in Predictive Process Monitoring

Fang Wang , Ernesto Damiani This is my paper

Pith reviewed 2026-06-26 21:23 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords predictive process monitoringevent sequence generationgraph attention transformerstructural constraintsfull sequence predictionconstrained decoding
0
0 comments X

The pith

GGATN generates full event sequences respecting process topology and attributes in a single non-autoregressive pass.

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

The paper introduces GGATN to solve the unified task of generating complete event sequences in predictive process monitoring while enforcing transition feasibility, temporal order, termination, and attribute consistency. It grounds generation in a global process graph used as structured activity memory, combines transformer self-attention for sequence context with graph-grounded cross-attention to inject topology, and produces activities, timestamps, length, and attributes together before applying Viterbi-style constrained decoding. Experiments across six benchmark logs report stronger similarity metrics and control-flow fidelity than local-instruction LLM baselines, with zero hallucinated activities and zero sequence-level attribute inconsistencies. Ablation studies identify the global graph encoder as a stable structural prior, and interpretability analysis traces how graph structure, sequence context, feedback, and decoding interact.

Core claim

GGATN uses a global process graph as structured activity memory, contextualizes sequence positions through Transformer self-attention, and injects process topology through graph-grounded cross-attention; unlike autoregressive decoding it generates activities, timestamps, length, and event- and sequence-level attributes in one pass, followed by Viterbi-style graph-constrained decoding for feasible paths and explicit termination.

What carries the argument

Graph Grounded Cross Attention that treats the global process graph as external memory and performs cross-attention between sequence positions and graph nodes to enforce topology during generation.

If this is right

  • Sequence similarity, Damerau-Levenshtein similarity, bigram control-flow similarity, and duration distribution all improve over prompted LLM baselines on the six evaluated logs.
  • Zero hallucinated activities and zero sequence-level attribute inconsistency are maintained across the reported experiments.
  • Ablation removing the global graph encoder degrades performance, confirming it functions as a stable structural prior.
  • Interpretability shows graph structure, sequence context, refinement feedback, and constrained decoding jointly determine output paths.

Where Pith is reading between the lines

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

  • The single-pass non-autoregressive design may lower latency for real-time process monitoring dashboards compared with iterative LLM prompting.
  • The same graph-grounding pattern could be tested on other constrained sequence tasks such as workflow scheduling or clinical pathway generation where topology must be respected.
  • If the upfront graph is mined from noisy logs, an online version that updates the graph encoder during generation might be needed to preserve the reported zero-inconsistency property.

Load-bearing premise

An accurate global process graph is available in advance and graph-grounded cross-attention can inject its topology without distorting sequence context or needing post-hoc fixes.

What would settle it

Running the model on a log where the supplied global process graph is deliberately incomplete or contains spurious edges and checking whether hallucinated activities or attribute inconsistencies appear at non-zero rates.

Figures

Figures reproduced from arXiv: 2606.18726 by Ernesto Damiani, Fang Wang.

Figure 1
Figure 1. Figure 1: Overview of GGATN for Structurally Constrained Full Event Sequence Generation 4.2. Global Graph Attention Encoder To capture global transition structure and temporal dynamics across the event log, we construct a global process graph and encode activity nodes into structure aware latent representations using an edge aware GAT encoder. Wang and Damiani: Preprint submitted to Elsevier Page 6 of 35 [PITH_FULL… view at source ↗
Figure 2
Figure 2. Figure 2: Ablation Analysis of GAT Encoder Training Regimes across Six Datasets Wang and Damiani: Preprint submitted to Elsevier Page 20 of 35 [PITH_FULL_IMAGE:figures/full_fig_p020_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ablation Analysis of GAT Attention Structure Shifts for Four Training Regimes across Helpdesk, BPI20, and Sepsis Datasets We compare four GGATN variants. In the frozen setting, used as the main model, the graph encoder is fixed while the remaining GGATN modules are trained. In the fully trainable setting, the graph encoder and the remaining modules are optimized jointly from the beginning, allowing activit… view at source ↗
Figure 4
Figure 4. Figure 4: Ablation Analysis of GAT Attention Structure Shifts for Four Training Regimes across BPI13C, BPI13I and BPI17 Datasets Figures 3 and 4 provide qualitative evidence of how graph attention weights are redistributed under frozen, staged, and fully trainable regimes. As expected, the frozen model shows zero embedding drift in [PITH_FULL_IMAGE:figures/full_fig_p022_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Dual Stage Attention Analysis Panel of Exactly Match Sample (BPI20) and Non Exactly Match Sample (Sepsis) Wang and Damiani: Preprint submitted to Elsevier Page 24 of 35 [PITH_FULL_IMAGE:figures/full_fig_p024_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Stage Wise Interpretability Analysis of GGATN 10.4. Structured Decoding and Transition Correction [PITH_FULL_IMAGE:figures/full_fig_p027_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Dual Stage Attention Analysis Panel of Representative Samples from Helpdesk Wang and Damiani: Preprint submitted to Elsevier Page 35 of 35 [PITH_FULL_IMAGE:figures/full_fig_p035_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Dual Stage Attention Analysis Panel of Representative Samples from BPI13C and BPI13I Wang and Damiani: Preprint submitted to Elsevier Page 36 of 35 [PITH_FULL_IMAGE:figures/full_fig_p036_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Dual Stage Attention Analysis Panel of Representative Samples from BPI17 Wang and Damiani: Preprint submitted to Elsevier Page 37 of 35 [PITH_FULL_IMAGE:figures/full_fig_p037_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Graph Grounded Cross Attention to Transition Admissibility and Target Activity for Helpdesk, Sepsis, BPI13C and BPI17 0.4 0.2 0.0 0.2 0.4 0.6 ptrue Change in true activity probability after refinement B1 B2 B3 B4 B5 Relative sequence position bin 0.10 0.05 0.00 0.05 0.10 0.15 0.20 0.25 0.30 Jensen Shannon divergence Distributional shift between provisional and final predictions Provisionally misclassified… view at source ↗
Figure 11
Figure 11. Figure 11: Refinement Based Activity Distribution Reshaping for BPI13I, Sepsis, BPI20 and BPI17 Wang and Damiani: Preprint submitted to Elsevier Page 38 of 35 [PITH_FULL_IMAGE:figures/full_fig_p038_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Transition Correction under Structured Decoding for Helpdesk, BPI13I, and BPI13C Wang and Damiani: Preprint submitted to Elsevier Page 39 of 35 [PITH_FULL_IMAGE:figures/full_fig_p039_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Transition Correction under Structured Decoding for Sepsis and BPI20 Wang and Damiani: Preprint submitted to Elsevier Page 40 of 35 [PITH_FULL_IMAGE:figures/full_fig_p040_13.png] view at source ↗
read the original abstract

Structurally constrained event sequence generation remains challenging because generated paths must preserve transition feasibility, temporal order, termination, and attribute consistency. In predictive process monitoring (PPM), this challenge appears as full event sequence generation, whereas existing work mainly addresses component tasks such as next activity, remaining time, outcome, and attribute prediction. This paper proposes the Graph Grounded Cross Attention Transformer Neural Network (GGATN) for this unified PPM task. GGATN uses a global process graph as structured activity memory, contextualizes sequence positions through Transformer self attention, and injects process topology through graph grounded cross attention. Unlike autoregressive decoding, GGATN generates activities, timestamps, length, and event level and sequence level attributes in a single pass, followed by Viterbi style graph constrained decoding for feasible paths and explicit termination. Experiments on six benchmark event logs show more reliable generation quality than local instruction prompted LLM baselines. GGATN achieves strong performance on sequence similarity, Damerau Levenshtein similarity, bigram based control flow similarity, and duration distribution, while maintaining zero hallucinated activities and zero sequence level attribute inconsistency. Ablation analyses confirm the global graph encoder as a stable structural prior. Interpretability analyses show how graph structure, sequence context, feedback refinement, and constrained decoding shape generation.

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 proposes the Graph Grounded Cross Attention Transformer Neural Network (GGATN) for full event sequence generation in predictive process monitoring. GGATN encodes a global process graph as structured activity memory, uses transformer self-attention to contextualize sequence positions, and injects topology via graph-grounded cross-attention. It performs single-pass generation of activities, timestamps, length, and attributes, followed by Viterbi-style graph-constrained decoding for feasible paths and termination. Experiments on six benchmark event logs report superior sequence similarity, Damerau-Levenshtein similarity, bigram control-flow similarity, and duration distribution compared to local instruction-prompted LLM baselines, with zero hallucinated activities and zero sequence-level attribute inconsistency. Ablations confirm the global graph encoder as a stable structural prior, and interpretability analyses examine the roles of graph structure, sequence context, feedback refinement, and constrained decoding.

Significance. If the central performance claims hold after clarifying the contribution of the neural components versus post-processing, the work offers a unified architecture for structurally constrained sequence generation that integrates external graph priors with attention mechanisms. The reported ablation on the global graph encoder and the interpretability analyses provide concrete evidence of the structural prior's role. The approach addresses a gap between component-wise PPM tasks and full-sequence generation, though applicability depends on the availability of an accurate upfront process graph.

major comments (2)
  1. [Abstract] Abstract: The reported 'zero hallucinated activities and zero sequence level attribute inconsistency' are stated after describing 'Viterbi style graph constrained decoding for feasible paths and explicit termination.' It is not specified whether these metrics are measured on raw GGATN outputs or only after the constrained decoder (which has direct access to the global graph). This distinction is load-bearing for the claim that GGATN 'achieves strong performance ... while maintaining zero' and for the comparison to 'local instruction prompted LLM baselines,' which receive no equivalent global-graph post-processing.
  2. [§4 (Experiments)] §4 (Experiments) and ablation analyses: The paper does not report whether the LLM baselines were given access to the same global process graph or subjected to equivalent constrained decoding. Without this, the performance gap cannot be attributed specifically to the graph-grounded cross-attention mechanism versus the shared use of the global graph as a hard constraint.
minor comments (2)
  1. [Abstract] The abstract and method description should explicitly state the data splits, hyperparameter search procedure, and whether statistical significance tests were performed across the six logs.
  2. [Method] Notation for the graph-grounded cross-attention (e.g., how the graph encoder output is projected into the transformer layers) should be defined with an equation or pseudocode for reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight key points for improving the clarity of our claims regarding the contributions of the GGATN architecture versus post-processing. We address each major comment below and will revise the manuscript to resolve the ambiguities.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The reported 'zero hallucinated activities and zero sequence level attribute inconsistency' are stated after describing 'Viterbi style graph constrained decoding for feasible paths and explicit termination.' It is not specified whether these metrics are measured on raw GGATN outputs or only after the constrained decoder (which has direct access to the global graph). This distinction is load-bearing for the claim that GGATN 'achieves strong performance ... while maintaining zero' and for the comparison to 'local instruction prompted LLM baselines,' which receive no equivalent global-graph post-processing.

    Authors: We agree this distinction requires explicit clarification. The reported zero hallucinated activities and zero sequence-level attribute inconsistency refer to the final outputs after Viterbi-style graph-constrained decoding, which enforces feasibility using the global process graph. Raw single-pass GGATN outputs (prior to decoding) can include infeasible transitions that are corrected by the decoder. We will revise the abstract, method description, and experimental reporting to state this explicitly, including that the LLM baselines receive no equivalent post-processing. revision: yes

  2. Referee: [§4 (Experiments)] §4 (Experiments) and ablation analyses: The paper does not report whether the LLM baselines were given access to the same global process graph or subjected to equivalent constrained decoding. Without this, the performance gap cannot be attributed specifically to the graph-grounded cross-attention mechanism versus the shared use of the global graph as a hard constraint.

    Authors: The LLM baselines are purely local instruction-prompted models without access to the global process graph and without any constrained decoding; their outputs are unconstrained and can include hallucinations. The performance gap therefore reflects both the graph-grounded cross-attention in GGATN and the subsequent constrained decoding. We will add explicit statements in §4 and the ablation section clarifying that baselines lack the global graph and post-processing. A fully isolated ablation (e.g., LLMs with graph prompting) is outside the current experimental scope but could be noted as future work. revision: yes

Circularity Check

0 steps flagged

No load-bearing circularity detected; performance attributed to full pipeline with external graph prior and post-processing

full rationale

The provided abstract and context describe GGATN as generating outputs in a single pass followed by explicit Viterbi-style graph constrained decoding, with the global process graph presented as an upfront external structural prior. No equations, self-citations, or derivation steps are quoted that reduce the reported zero-inconsistency metrics or similarity scores to quantities defined by construction from fitted parameters on the same data. Ablation analyses are mentioned but not shown to create self-referential loops. This aligns with a minor score for normal self-referential language in method description without forcing the central claims.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies insufficient technical detail to enumerate free parameters, axioms, or invented entities; no equations or modeling choices are shown.

pith-pipeline@v0.9.1-grok · 5762 in / 1081 out tokens · 25329 ms · 2026-06-26T21:23:37.505084+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

35 extracted references · 13 canonical work pages · 8 internal anchors

  1. [1]

    arXiv preprint arXiv:2006.05205

    On the bottleneck of graph neural networks and its practical implications. arXiv preprint arXiv:2006.05205 . Bahdanau, D., Cho, K., Bengio, Y.,

    work page arXiv 2006

  2. [2]

    Neural Machine Translation by Jointly Learning to Align and Translate

    Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473 . Beck, D., Haffari, G., Cohn, T.,

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Graph-to-sequence learning using gated graph neural networks, in: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 273–283. Bronstein,M.M.,Bruna,J.,Cohen,T.,Veličković,P.,2021. Geometricdeeplearning:Grids,groups,graphs,geodesics,andgauges. arXivpreprint arXiv:2104.13478 . Brown, T., Mann,...

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  4. [4]

    Advances in neural information processing systems 33, 1877–1901

    Language models are few-shot learners. Advances in neural information processing systems 33, 1877–1901. Bukhsh, Z.A., Saeed, A., Dijkman, R.M.,

    1901

  5. [5]

    arXiv preprint arXiv:2104.00721

    Processtransformer: Predictive business process monitoring with transformer network. arXiv preprint arXiv:2104.00721 . Camargo,M.,Dumas,M.,González-Rojas,O.,2019. Learningaccuratelstmmodelsofbusinessprocesses,in:InternationalConferenceonBusiness Process Management, Springer. pp. 286–302. Cao, Y., Han, S., Gao, Z., Ding, Z., Xie, X., Zhou, S.K.,

    work page arXiv 2019

  6. [6]

    Knowledge-Based Systems 254, 109603

    Multi-task prediction method of business process based on bert and transfer learning. Knowledge-Based Systems 254, 109603. Cho,K.,VanMerriënboer,B.,Gulçehre,Ç.,Bahdanau,D.,Bougares,F.,Schwenk,H.,Bengio,Y.,2014. Learningphraserepresentationsusingrnn encoder–decoderforstatisticalmachinetranslation,in:Proceedingsofthe2014conferenceonempiricalmethodsinnatural...

    2014

  7. [7]

    Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling

    Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555 . Dissegna, S., Di Francescomarino, C.,

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [9]

    Remaining cycle time prediction with graph neural networks for predictive process monitoring, in: Proceedings of the 2023 8th international conference on machine learning technologies, pp. 95–101. Dwivedi, V.P., Bresson, X.,

    2023

  9. [10]

    arXiv:2012.09699

    A generalization of transformer networks to graphs. arXiv preprint arXiv:2012.09699 . Elman, J.L.,

    work page arXiv 2012

  10. [11]

    4186–4196

    Using local knowledge graph construction to scale seq2seq models to multi-document inputs, in: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pp. 4186–4196. Geng, S., Cooper, H., Moskal, M., Jenkins, S., Berman, J., Ranchin, N....

    2019

  11. [12]

    arXiv e-prints , arXiv–2501

    Generating structured outputs from language models: Benchmark and studies. arXiv e-prints , arXiv–2501. Gilmer,J.,Schoenholz,S.S.,Riley,P.F.,Vinyals,O.,Dahl,G.E.,2017. Neuralmessagepassingforquantumchemistry,in:Internationalconference on machine learning, Pmlr. pp. 1263–1272. Guo, N., Liu, C., Li, C., Zeng, Q., Ouyang, C., Liu, Q., Lu, X.,

    2017

  12. [13]

    12045–12072

    Beyond traditional benchmarks: Analyzing behaviors of open llms on data-to-text generation, in: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 12045–12072. Khan,A.,Le,H.,Do,K.,Tran,T.,Ghose,A.,Dam,H.,Sindhgatta,R.,2021.Deepprocess:supportingbusinessprocessexecutionusingamann-based recom...

    2021

  13. [14]

    Semi-Supervised Classification with Graph Convolutional Networks

    Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 . Koncel-Kedziorski, R., Bekal, D., Luan, Y., Lapata, M., Hajishirzi, H.,

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [15]

    2284–2293

    Text generation from knowledge graphs with graph transformers, in: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Wang and Damiani:Preprint submitted to ElsevierPage 29 of 35 GGATN Technologies, Volume 1 (Long and Short Papers), pp. 2284–2293. Kratsch, W.,Manderscheid, J., ...

    2019

  15. [16]

    20004–20026

    Exposing numeracy gaps: A benchmark to evaluate fundamental numerical abilities in large language models, in: Findings of the Association for Computational Linguistics: ACL 2025, pp. 20004–20026. Li, Y., Tarlow, D., Brockschmidt, M., Zemel, R.,

    2025

  16. [17]

    Gated Graph Sequence Neural Networks

    Gated graph sequence neural networks. arXiv preprint arXiv:1511.05493 . Lin, L., Wen, L., Wang, J.,

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [18]

    Mm-pred: A deep predictive model for multi-attribute event sequence, in: Proceedings of the 2019 SIAM international conference on data mining, SIAM. pp. 118–126. Mannhardt, F.,

    2019

  18. [20]

    Lstm networks for data-aware remaining time prediction of business process instances, in: 2017 IEEE symposium series on computational intelligence (SSCI), IEEE. pp. 1–7. Nguyen, A., Chatterjee, S., Weinzierl, S., Schwinn, L., Matzner, M., Eskofier, B.,

    2017

  19. [21]

    Time matters: Time-aware lstms for predictive business process monitoring, in: International Conference on Process Mining, Springer. pp. 112–123. Pasquadibisceglie,V.,Appice,A.,Castellano,G.,Malerba,D.,2019. Usingconvolutionalneuralnetworksforpredictiveprocessanalytics,in:2019 international conference on process mining (ICPM), IEEE. pp. 129–136. Pasquadib...

    2019

  20. [23]

    Radford,A.,Kim,J.W.,Hallacy,C.,Ramesh,A.,Goh,G.,Agarwal,S.,Sastry,G.,Askell,A.,Mishkin,P.,Clark,J.,etal.,2021.Learningtransferable visual models from natural language supervision, in: International conference on machine learning, PmLR. pp. 8748–8763. Raffel,C.,Shazeer,N.,Roberts,A.,Lee,K.,Narang,S.,Matena,M.,Zhou,Y.,Li,W.,Liu,P.J.,2020. Exploringthelimits...

    2021

  21. [24]

    IEEE Transactions on Knowledge and Data Engineering 36, 137–151

    Embedding graph convolutional networks in recurrent neural networks for predictive monitoring. IEEE Transactions on Knowledge and Data Engineering 36, 137–151. Rampášek,L.,Galkin,M.,Dwivedi,V.P.,Luu,A.T.,Wolf,G.,Beaini,D.,2022.Recipeforageneral,powerful,scalablegraphtransformer.Advances in Neural Information Processing Systems 35, 14501–14515. Rivera Lazo...

    2022

  22. [25]

    Multi-attribute transformers for sequence prediction in business process management, in: International Conference on Discovery Science, Springer. pp. 184–194. Scarselli,F.,Gori,M.,Tsoi,A.C.,Hagenbuchner,M.,Monfardini,G.,2008. Thegraphneuralnetworkmodel. IEEEtransactionsonneuralnetworks 20, 61–80. Schmidt, F.,

    2008

  23. [28]

    Sequencetosequencelearningwithneuralnetworks

    Sutskever,I.,Vinyals,O.,Le,Q.V.,2014. Sequencetosequencelearningwithneuralnetworks. Advancesinneuralinformationprocessingsystems

    2014

  24. [29]

    Predictivebusinessprocessmonitoringwithlstmneuralnetworks,in:Internationalconference on advanced information systems engineering, Springer

    Tax,N.,Verenich,I.,LaRosa,M.,Dumas,M.,2017. Predictivebusinessprocessmonitoringwithlstmneuralnetworks,in:Internationalconference on advanced information systems engineering, Springer. pp. 477–492. Taymouri, F., Rosa, M.L., Erfani, S., Bozorgi, Z.D., Verenich, I.,

    2017

  25. [30]

    LLaMA: Open and Efficient Foundation Language Models

    Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 . Wang and Damiani:Preprint submitted to ElsevierPage 30 of 35 GGATN Van Dongen, B.F.,

    work page internal anchor Pith review Pith/arXiv arXiv

  26. [31]

    4TU.ResearchData

    BPI Challenge 2020: Prepaid Travel Cost (Event Log). URL:https://doi.org/10.4121/uuid: 52fb97d4-4588-43c9-9d04-3604d4613b51, doi:10.4121/uuid:52fb97d4-4588-43c9-9d04-3604d4613b51. dataset, Version

    work page doi:10.4121/uuid: 2020

  27. [32]

    Graph Attention Networks

    Graph attention networks. arXiv preprint arXiv:1710.10903 . Verenich, I., Dumas, M., Rosa, M.L., Maggi, F.M., Teinemaa, I.,

    work page internal anchor Pith review Pith/arXiv arXiv

  28. [33]

    ACM Transactions on Intelligent Systems and Technology (TIST) 10, 1–34

    Survey and cross-benchmark comparison of remaining time prediction methods in business process monitoring. ACM Transactions on Intelligent Systems and Technology (TIST) 10, 1–34. Wang, F., Ceravolo, P., Damiani, E., 2025a. Comprehensive attribute encoding and dynamic lstm hypermodels for outcome oriented predictive business process monitoring. arXiv prepr...

    work page arXiv

  29. [34]

    Expert Systems with Applications , 130320

    Time-aware and transition-semantic graph neural networks for interpretable predictive business process monitoring. Expert Systems with Applications , 130320. Wang, F., Kosca, L., Kosca, A., Gacesa, M., Damiani, E., 2025d. Auto-ml graph neural network hypermodels for outcome prediction in event- sequence data, in: 2025 IEEE 19th International Conference on...

    2025

  30. [35]

    Outcome-oriented predictive process monitoring with attention-based bidirectional lstm neural networks, in: 2019 IEEE international conference on web services (ICWS), IEEE. pp. 360–367. Wang, Y., Zhao, Y.,

    2019

  31. [36]

    6389–6415

    Tram: Benchmarking temporal reasoning for large language models, in: Findings of the Association for Computational Linguistics: ACL 2024, pp. 6389–6415. Weinzierl,S.,2021. Exploringgatedgraphsequenceneuralnetworksforpredictingnextprocessactivities,in:Internationalconferenceonbusiness process management, Springer. pp. 30–42. Weinzierl, S., Dunzer, S., Zilk...

    2024

  32. [37]

    Sutran: an encoder-decoder transformer for full-context-aware suffix prediction of business processes, in: 2024 6th International Conference on Process Mining (ICPM), IEEE. pp. 17–24. Xu,K.,Wu,L.,Wang,Z.,Feng,Y.,Witbrock,M.,Sheinin,V.,2018. Graph2seq:Graphtosequencelearningwithattention-basedneuralnetworks. arXiv preprint arXiv:1804.00823 . Yin, J., Qiu, ...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  33. [38]

    Zhou, H., Zhang, S., Peng, J., Zhang, S., Li, J., Xiong, H., Zhang, W., 2021a

    Do transformers really perform badly for graph representation? Advances in neural information processing systems 34, 28877–28888. Zhou, H., Zhang, S., Peng, J., Zhang, S., Li, J., Xiong, H., Zhang, W., 2021a. Informer: Beyond efficient transformer for long sequence time-series forecasting, in: Proceedings of the AAAI conference on artificial intelligence,...

    2021

  34. [39]

    5459–5468

    Modeling graph structure in transformer for better amr-to-text generation, in: Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP), pp. 5459–5468. Appendix A. Complete Results by Dataset The appendix tables follow the abbreviations d...

    2019

  35. [40]

    The main generation metricsarecoverage,SS,DL,bigramJSD,anddurationWD.Datasetspecificcolumnsreportactivity,temporal,event level, and sequence level attribute metrics

    Briefly,Mdenotes model;Gis the main GGATN model with a frozen graph attention encoder,G_jis the fully joint training variant, andG_s5/G_s10are staged unfreezing variants.L(4k),L(32k), andM(4k)denote the Llama and Mistral baselines. The main generation metricsarecoverage,SS,DL,bigramJSD,anddurationWD.Datasetspecificcolumnsreportactivity,temporal,event leve...

    1923