GNN4PPM: Multi-Target Predictive Process Monitoring with Relational Graph Convolutional Networks
Ana Costa, Johannes Mäkelburg, Luise Pufahl
cs.LG
Quick verdict
Claims multi-target next-event prediction via RDF graphs + R-GCN; best or competitive NAP on 6 BPIC logs, but validation-only evaluation weakens the evidence.
Summary
The paper introduces GNN4PPM, a multi-target predictive process monitoring approach that predicts the next event together with its complete data payload — activity, timestamp, and all event attributes — in a single model. Event logs are transformed into heterogeneous RDF knowledge graphs via YARRRML/RML mappings, with case, event, and attribute-value nodes plus attribute-level directly-follows edges. RDF2Vec biased random walks with CBOW produce initial node embeddings (tuned by silhouette score), and a three-layer R-GCN with basis decomposition performs relation-specific message passing. Training combines inverse-frequency-weighted cross-entropy for categorical attributes, MSE on log-transformed timestamps, and Huber loss for numerical attributes. On six public BPIC logs, GNN4PPM achieves the best next-activity accuracy on BPIC12A (88.8), BPIC12WC (83.2), and BPIC13O (78.1), remaining competitive elsewhere (80.8 on BPIC12W, where PROPHET reaches 89.2). Multi-target results split sharply: low-cardinality attributes score high (resource 99.58 on BPIC20R) while high-cardinality ones score poorly (resource 14.83 on BPIC12WC); timestamp MAE ranges 0.32–0.77 in normalized log space. The ablation shows removing timestamp or role attributes mostly hurts accuracy. Key limitations: all metrics are computed on the same validation set used for hyperparameter tuning, with no independent test set; no baseline exists for the full multi-target task; significance tests are absent; and comparability is mixed, since some baselines were re-run and others taken from a prior benchmark.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A single model can predict the next event's complete data payload (activity, timestamp, all attributes) simultaneously, eliminating per-target retraining (Sec. 1, 4).
- Encoding event logs as heterogeneous RDF knowledge graphs with attribute-level directly-follows edges lets an R-GCN exploit inter-attribute and temporal context that sequential prefix embeddings discard (Sec. 1, 4.1).
- GNN4PPM achieves the best next-activity accuracy on 3 of 6 public BPIC logs (BPIC12A 88.8, BPIC12WC 83.2, BPIC13O 78.1) and remains competitive on the rest (Table 2).
- Contextual attributes actively improve prediction: removing timestamp, resource, or org_role mostly decreases accuracy (Sec. 5.5, Fig. 4).
Strengths
Genuine multi-target novelty for PPM
First combination (per authors, Sec. 2) of RDF-ized event logs with an R-GCN to predict all next-event attributes jointly; attribute-level directly-follows triples (Eq. 3) explicitly model cross-attribute temporal dependencies that sequential prefix encoders lose.
Principled handling of heterogeneous targets
Sec. 4.3 assigns type-appropriate losses: inverse-frequency-weighted cross-entropy for imbalanced categorical attributes, MSE on log-transformed timestamps to compress heavy-tailed gaps, Huber loss for noisy numerical attributes; missing targets are masked from gradients.
Broad benchmark with released artifacts
Real ablation of attribute contributions
Leakage-aware embedding selection and honest limitations
Weaknesses
Evaluation on the tuning set
Core multi-target claim has no baseline
No statistical rigor
Ambiguous inference-time graph construction
Per-attribute results partly uninterpretable
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
GNN4PPM — Preliminary Review
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is GNN4PPM?
GNN4PPM is a predictive process monitoring method that represents an event log as an RDF knowledge graph and uses a Relational Graph Convolutional Network (R-GCN) with RDF2Vec initial embeddings to predict, in one single model, the next event together with all its attributes — the activity, the timestamp, and every recorded event attribute such as resource or role.
How is GNN4PPM different from LSTM-based next-activity prediction?
LSTM approaches encode a case prefix as a sequence, usually of activity labels only, and are trained per target. GNN4PPM instead builds a heterogeneous graph of cases, events, and attribute values with attribute-level directly-follows relations, embeds it with RDF2Vec, and trains an R-GCN with per-attribute losses, so contextual attributes inform prediction without retraining a separate model per target.
What datasets were used and how well does GNN4PPM perform?
Six public BPI Challenge logs (BPIC12 Application/Work/combined, BPIC13 Open problems, BPIC20 Travel permits and Requests for payment). The paper reports the best next-activity accuracy on BPIC12A (88.8), BPIC12WC (83.2), and BPIC13O (78.1), and competitive results elsewhere; note these are validation-set figures from a split also used for tuning.
Is code or data available for GNN4PPM?
Yes. The authors publish source code on GitHub (ana-luisa-costa/GNN4PPM) and additional result files on figshare; the six event logs themselves are public via the 4TU Research Data repository.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1, 2, and 4 are referenced but their plotted content is only partially recoverable from extracted text; ablation deltas were taken from the body-text numbers.
- Tables 2–3 were parsed from plain-text extraction; individual cells carry minor transcription risk (e.g., body text says resource 99.56 for BPIC20R while Table 3 shows 99.58).
- Review covers arXiv v1 (13 Sep 2026); no peer-review history is available.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.