On the Potential of Multi-Task Learning in Predictive Process Monitoring
Lukas Kirchdorfer, Keyvan Amiri Elyasi, Heiner Stuckenschmidt
cs.LG
Quick verdict
Claims MTL strongly improves next-activity prediction in PPM; broad benchmark largely convincing, but temporal-task gains are statistically null.
Summary
This empirical study asks whether multi-task learning (MTL) outperforms the prevailing single-task setup in predictive process monitoring (PPM). The authors benchmark three architectures (LSTM, CNN, Transformer), all four combinations of three tasks (next activity, next timestamp, remaining time prediction), and thirteen multi-task optimization methods across nine public event logs, training 1,404 multi-task models per architecture over three seeds. Results show MTL beats single-task models in the majority of runs, with gains concentrated almost entirely in next-activity prediction: accuracy rises by roughly 10 percentage points on most logs (e.g., Sepsis 48%→60%, BPIC20ID +24 points), macro-F1 improves 10–15 points on all datasets, and Wilcoxon tests confirm significance for NAP on all architectures while temporal tasks show mostly null effects. Task combinations barely matter, and all thirteen optimization methods perform similarly; however, a smaller-CNN ablation shows equal weighting fails under low capacity, favoring inverse-loss weighting (UW-O). Limitations: significance tests pool non-independent configurations without multiple-comparison correction, the leakage-robust split results live only in the repository, and no comparison is run against dedicated class-imbalance methods (focal loss, cost-sensitive learning) despite the headline imbalance claim.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
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
- MTL outperforms STL in the majority of 1,404-per-architecture experiments, with Δm first and third quartiles below zero for all three architectures (Figure 2).
- The overall MTL gain is driven by next-activity prediction, significant at the 1% level for all architectures (Table 3), with ~10-point accuracy gains on most logs (Table 4).
- MTL inherently mitigates class imbalance, improving macro-F1 on all nine logs, often by 10–15 points (Section 5.2, Figure 5).
- Task combination choice and MTO method choice have minimal impact at sufficient capacity; under low capacity, equal weighting fails and inverse loss weighting (UW-O) is the practical compromise (Figures 7–8).
- MTL's advantage stems from inductive transfer of temporal supervision, not overfitting prevention (Figure 3 loss analysis).
Strengths
Unusually broad, systematic benchmark
9 public logs × 3 architectures × 4 task combinations × 13 MTO methods × 3 seeds = 1,404 multi-task models per architecture (Section 5.1), covering all pairwise and full task combinations — no prior PPM MTL study (Table 1) comes close.
Claims calibrated by statistics, not overreached
Table 3's Wilcoxon tests honestly show NTP/RTP mostly non-significant (RTP p=0.055–0.742); the authors explicitly attribute overall gains to NAP instead of claiming uniform improvement.
Mechanistic analyses beyond leaderboard numbers
Actionable class-imbalance finding
High reproducibility posture
Weaknesses
Pooled, uncorrected significance testing
Δm aggregate can mask per-task degradation
Class-imbalance claim lacks the decisive baseline
Capacity conclusion anchored to one architecture
Key robustness evidence kept out of the paper
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper claims to be the first systematic, large-scale empirical evaluation of multi-task learning (MTL) for predictive process monitoring (PPM). The scope is precisely bounded: three PPM tasks — next activity prediction (NAP), next timestamp prediction (NTP), remaining time prediction (RTP) — evaluated in all four multi-task combinations, across three architectures (LSTM, CNN, Transformer), thirteen multi-task optimization (MTO) methods, and nine public event logs (Table 2: P2P, Production, Helpdesk, Sepsis, BPIC20DD, BPIC20ID, BPIC12C, BPIC15-1, BPIC13I), yielding 1,404 MTL models per architecture at three seeds. Four research questions structure the study (RQ1: MTL vs STL; RQ2: per-task impact; RQ3: task combinations; RQ4: MTO methods). Headline claims: MTL outperforms STL in the majority of experiments; gains concentrate in NAP (~10 accuracy points on most logs); class imbalance is inherently mitigated; task balancing matters only under low capacity; inductive transfer, not overfitting prevention, is the mechanism.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 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
Does multi-task learning improve predictive process monitoring?
According to this study, yes in most settings: a single jointly trained model beat separate single-task models in the majority of 1,404 experiments per architecture, mainly by raising next-activity prediction accuracy by roughly 10 percentage points, while next-time and remaining-time predictions were largely unchanged.
Which multi-task optimization method should I use for PPM?
The paper finds all 13 tested methods perform similarly when models are sufficiently large, so equal weighting is a fine default. Under small model capacity, inverse loss weighting (UW-O) and Nash-MTL performed best, while equal weighting ranked among the worst.
Does MTL help with class imbalance in next-activity prediction?
The authors report macro-F1 gains of 10–15 points across all nine event logs, with rare activities improving most, suggesting MTL mitigates class imbalance without specialized loss functions — though the paper does not compare against dedicated imbalance methods such as focal loss.
Is the code and data for this paper available?
Yes. The implementation, the nine event logs (originally from data.4tu.nl), and supplementary results are in a public GitHub repository: github.com/lukaskirchdorfer/MTL4PPM.
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.
- The provided text ends at the reference list; the appendix holding the MTO-method statistical test (referenced as 'see Appendix' in Section 5.4) is not part of the extraction.
- Figure panel values are extraction-garbled (e.g., Figure 2 axis text '125 100 75 50 25 0 25 m'), so figures were judged via the prose descriptions; only table values were quoted numerically.
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.