Process-Aware Event Log Augmentation for Process Prediction under Labeled Data Scarcity

Sjoerd van Straten; Alessandro Padella; Marwan Hassani

cs.LGOct 17, 2026

Process-Aware Event Log Augmentation for Process Prediction under Labeled Data Scarcity

Authors: Sjoerd van Straten, Alessandro Padella, Marwan Hassani

Organizations: Eindhoven University of Technology, Eindhoven, The Netherlands · University of Padua, Padua, Italy · University of Wuppertal, Wuppertal, Germany

Abstract

Predictive Process Monitoring (PPM) forecasts future events of ongoing business processes based on event logs. However, deep learning PPM approaches are often limited by labeled data scarcity and the low variability of real-world logs. To address this, we introduce SiamSA-PPM, a self-supervised framework combining Siamese learning with statistically grounded data augmentation. It employs novel transformation methods that leverage control-flow semantics and frequent behavioral patterns to generate realistic, semantically valid trace variants, learning generalizable prefix representations without labeled supervision. Extensive experiments on real-world event logs with a wide range of variability, incorporating exhaustive hyperparameter tuning, demonstrate the value of this process-aware augmentation. Compared to state-of-the-art baselines, SiamSA-PPM achieves competitive or superior predictive performance in both next activity and final outcome prediction tasks. An open source implementation of our method is publicly available under https://github.com/SvStraten/SiamSA-PPM-Extension..

Figures & tables

Figure 1: The impact of class imbalance and data augmentation on predictive models using a loan application scenario. Training on a raw, severely imbalanced log (left) leads to biased predictions. The model achieves a high overall accuracy by favoring the majority class but fails to detect the critical minority class. By augmenting the training log to balance the distributions (right), the model can successfully learn the patterns of the rare cases, drastically improving performance on the minority class.

Figure 2: Random augmentation strategies can preserve semantics in image and text domains, where structural flexibility exists. However, in process data, such augmentations risk violating control-flow dependencies for instance, producing logically inconsistent or non-executable sequences.

Domain	Method (Authors)	Technique	Process Prediction	Image Classification	Text Classification	Representation Learning	Contrastive Learning	Siamese Learning	Random Augment.	Context- aware	Hallucin- ation
PPM	Tax et al. (2017) [ 26 ]	LSTM	✓
	Di Mauro et al. (2019) [ 10 ]	CNN	✓
	Pasquadibisceglie et al. (2019) [ 24 ]	CNN	✓
	Bukhsh et al. (2021) [ 3 ]	Transformer	✓
	Ni et al. (2022) [ 22 ]	Auto-Encoder	✓
	Taymouri et al. (2020) [ 27 ]	GAN	✓

Table 1: Overview of related methods across Predictive Process Monitoring (PPM), Natural Language Processing (NLP), and Computer Vision (CV). The light green shaded block in the top right highlights the primary research gap addressed in this work: while representation learning (including contrastive and Siamese architectures) and advanced data augmentation techniques are highly prevalent in NLP and CV, they remain underexplored in PPM. Crucially, no existing PPM method integrates contrastive or Siamese learning with event log augmentation.

Figure 3: Overview of our three novel transformation methods: StatisticalInsertion (green), StatisticalDeletion (red) and StatisticalReplacement (purple).

Figure 4: Overview of the pre-training architecture for SiamSA-PPM. Given an input prefix

x

, two augmented views

v

and

v^{\prime}

are generated using transformation functions

t

and

t^{\prime}

chosen following our augmentation strategy and passed through a shared Transformer encoder and projection Multi-Layer Perceptron (MLP). The online network produces a prediction

q_{\theta}(z_{\theta})

, while the target network (updated via EMA) produces

z^{\prime}_{\xi}

. A loss is computed between the prediction and the stop-gradient of the target to enable representation alignment.

	Dataset	#Events	#Cases	#Act.	Avg. Len.	#Variants
Target	Helpdesk	21,348	4,580	14	4.66	226
	BPIC13-c	6,660	1,487	7	4.48	327
	Sepsis	15,214	1,050	16	14.49	846
	BAC	212,721	32,429	15	6.56	189
Variable	BPIC12	262,200	13,087	36	20.04	4,366
	BPIC17	200,033	5,168	25	38.71	3,248

Table 2: Statistics of the event logs used in the experimental evaluation, divided into the target group and highly variable datasets. The split is determined based on the number of variants.

		SiamSA-PPM (ours)							Tax [ 26 ]		Di Mauro [ 10 ]		Pasquad. [ 24 ]		Bukhsh [ 3 ]
	Dataset	$\bm{\alpha}$	$\bm{\beta}$	$\bm{\gamma}$	$\bm{\delta}$	$\bm{\lambda}_{\textbf{max}}$	bs	$\bm{\eta}$	$\bm{units}$	$\bm{\eta}$	$\bm{filters}$	$\bm{\eta}$	$\bm{filters}$	$\bm{\eta}$	$\bm{emb}$	$\bm{\eta}$
Target	Helpdesk	$10^{-3}$	$10^{-3}$	$10^{-3}$	$10^{-4}$	4	256	0.05	150	0.0005	16	0.0005	16	0.001	32	0.0005
	BPIC13-c	$10^{-3}$	$10^{-4}$	$10^{-4}$	$10^{-3}$	4	256	0.05	100	0.0005	16	0.001	16	0.001	32	0.0005
	Sepsis	$10^{-3}$	$10^{-4}$	$10^{-4}$	$10^{-3}$	3	256	0.05	50	0.001	64	0.001	32	0.001	128	0.001
	BAC	$10^{-4}$	$10^{-4}$	$10^{-4}$	$10^{-4}$	3	256	0.05	50	0.001	32	0.0005	32	0.0005	128	0.001
Variable	BPIC12	$10^{-4}$	$10^{-4}$	$10^{-3}$	$10^{-4}$	4	256	0.05	100	0.001	64	0.001	32	0.0001	64	0.001

Table 3: Optimal hyperparameters selected for each event log after grid search on the validation set.

Figure 5: Increase in prefix and trace entropy as a function of augmentation factor, showing how data augmentation enhances variability across different event logs.

Dataset	Log Type	Method
Dataset	Log Type	Tax [ 26 ]	Di Mauro [ 10 ]	Pasquad. [ 24 ]	Bukhsh [ 3 ]
Helpdesk	Original Log	$\bm{70.63}\pm\bm{0.44}$	$\bm{78.85}\pm\bm{0.04}$	$\bm{78.81}\pm\bm{0.17}$	$78.45\pm 0.25$
	+ Random Augment. [ 15 ]	$69.21\pm 0.68$	$75.56\pm 0.43$	$78.64\pm 0.32$	$78.76\pm 0.17$
	+ Our Strategy	$68.79\pm 0.78$	$77.54\pm 0.56$	$78.71\pm 0.12$	$\bm{78.93}\pm\bm{0.10}$
BPIC13-c	Original Log	$\bm{55.18}\pm\bm{3.17}$	$56.85\pm 2.01$	$\bm{52.31}\pm\bm{0.17}$	$55.92\pm 2.23$
	+ Random Augment. [ 15 ]	$53.81\pm 3.73$	$57.89\pm 1.97$	$52.01\pm 2.56$	$54.97\pm 2.69$

Table 4: Accuracy of the baselines for next activity prediction using the original training log, the augmented training log following [ 15 ] and the augmented training log following our strategy (cf. Section 4.2 ). Bold denotes the highest accuracy per dataset and baseline.

Dataset	Tax [ 26 ]	Di Mauro [ 10 ]	Pasquad. [ 24 ]	Bukhsh [ 3 ]	SiamSA-PPM
Helpdesk	$70.63\pm 0.44$	$\bm{78.85}\pm\bm{0.04}$	$\mathit{78.81}\pm\mathit{0.17}$	$78.45\pm 0.25$	$\underline{78.60}\pm\underline{0.18}$
BPIC13-c	$55.18\pm 3.17$	$\mathit{56.85}\pm\mathit{2.01}$	$52.31\pm 0.17$	$\underline{55.92}\pm\underline{2.23}$	$\bm{57.06}\pm\bm{0.24}$
Sepsis	$52.69\pm 1.40$	$\mathit{59.44}\pm\mathit{1.93}$	$56.07\pm 1.72$	$\underline{59.16}\pm\underline{2.18}$	$\bm{60.62}\pm\bm{0.20}$
BAC	$80.32\pm 0.31$	$95.12\pm 0.20$	$\underline{95.19}\pm\underline{0.22}$	$\mathit{95.47}\pm\mathit{0.03}$	$\bm{95.49}\pm\bm{0.01}$

Table 5: Accuracy on the target group datasets (mean

\pm

std) for next activity prediction. Bold denotes the highest performance, italic and underline the second and third highest respectively.

Dataset	Tax [ 26 ]	Di Mauro [ 10 ]	Pasquad. [ 24 ]	Bukhsh [ 3 ]	SiamSA-PPM
BPIC12	$70.85\pm 0.79$	$\mathit{83.20}\pm\mathit{1.00}$	$\underline{83.14}\pm\underline{0.27}$	$\bm{83.47}\pm\bm{0.31}$	$79.48\pm 0.05$
BPIC17	$75.25\pm 0.11$	$\bm{88.77}\pm\bm{0.13}$	$\underline{88.48}\pm\underline{0.23}$	$\mathit{88.63}\pm\mathit{0.12}$	$85.65\pm 0.06$
BPIC15-1	$24.83\pm 0.44$	$\bm{47.15}\pm\bm{0.55}$	$\mathit{41.90}\pm\mathit{0.46}$	$\underline{38.54}\pm\underline{1.53}$	$37.11\pm 0.13$
BPIC13-i	$55.02\pm 0.51$	$\bm{59.39}\pm\bm{0.11}$	$\underline{59.10}\pm\underline{3.79}$	$58.69\pm 0.12$	$\mathit{59.27}\pm\mathit{0.08}$

Table 6: Accuracy on the highly variable datasets (mean

\pm

std) for next activity prediction. Bold denotes the highest performance, italic and underline the second and third highest respectively.

	Target	Tax [ 26 ]	Di Mauro [ 10 ]	Pasquad. [ 24 ]	Bukhsh [ 3 ]	SiamSA-PPM
BPIC12	Approved	$65.63\pm 3.40$	$56.42\pm 0.25$	$\mathit{74.73}\pm\mathit{0.58}$	$\bm{75.43}\pm\bm{0.22}$	$\underline{73.58}\pm\underline{0.14}$
	Declined	$75.31\pm 1.50$	$69.70\pm 0.64$	$\underline{79.50}\pm\underline{1.15}$	$\bm{80.86}\pm\bm{0.04}$	$\mathit{80.04}\pm\mathit{0.02}$
	Cancelled	$63.36\pm 2.48$	$\bm{75.76}\pm\bm{0.01}$	$74.47\pm 3.02$	$\mathit{75.63}\pm\mathit{1.13}$	$\underline{75.39}\pm\underline{0.06}$
Sepsis	Release-A	$\bm{84.15}\pm\bm{2.27}$	$77.39\pm 0.98$	$\underline{81.95}\pm\underline{0.48}$	$\mathit{83.42}\pm\mathit{0.83}$	$80.71\pm 0.11$
	Release-B	$92.47\pm 1.75$	$91.17\pm 2.42$	$\underline{93.28}\pm\underline{0.43}$	$\mathit{94.06}\pm\mathit{0.65}$	$\bm{94.21}\pm\bm{0.08}$
	Release-C	$90.58\pm 0.11$	$90.10\pm 0.49$	$\mathit{90.80}\pm\mathit{0.06}$	$\bm{90.82}\pm\bm{0.00}$	$\underline{90.59}\pm\underline{0.00}$

Table 7: Accuracy (mean

\pm

std) for final outcome prediction. Bold denotes the highest performance, italic and underline the second and third highest respectively.

Figure 6: Next activity transitions following the A_PARTLYSUBMITTED event in the BPIC12 dataset. The absolute transition counts of the test set (20%) establish the ground truth.

Figure 7: Comparison between events directly following A_PARTLYSUBMITTED in the original train log and the two augmented logs, highlighting the semantic-violating noise injected by the random augmentation method [ 15 ] compared to our strategy.

Figure 8: Next activity transitions following the Service closure Request with network responsibility event in the BAC dataset. The absolute transition counts of the test set (20%) establish the ground truth.

Figure 9: Comparison between events directly following Service closure Request with network responsibility in the original train log and the two augmented logs.

Figure 10: Relative inference time (normalized to the slowest method per dataset) across all benchmark datasets.

Training Strategy	Test Accuracy (%)				Average
	Helpdesk	BPIC13-c	Sepsis	BAC
Supervised only	$78.50\pm 0.13$	$55.67\pm 0.21$	$60.87\pm 0.53$	$91.50\pm 0.10$	71.64
+ Pre-training (random strategy)	$78.58\pm 0.23$	$56.09\pm 0.28$	$60.13\pm 0.90$	$95.43\pm 0.06$	72.56
+ Pre-training (our strategy)	$78.60\pm 0.18$	$57.06\pm 0.24$	$60.62\pm 0.20$	$95.49\pm 0.01$	$\mathbf{72.94}$

Table 8: Ablation study for the three training strategies over the target datasets.

Figure 11: Accuracy comparison between a deep learning Transformer model (left) and shallow Random Forest Classifier (right) on the BPIC15-1 dataset, under varying raw training subset sizes (20%, 40% and 60%). Each strategy (no augmentation, random augmentation and our augmentation method) uses the same base data, but the latter two are upsampled to the original 80% training size. The red dashed line denotes the baseline of training on the full 80% raw training data.

Predictive Process Monitoring

Predictive process monitoring uses partial event histories to forecast how processes will unfold and help people act before cases finish.

See all 25 papers

D-TAIA: Domain-Aware LLM Adaptation for Multi-Task Predictive Process Monitoring
Aug 28, 2026Sjoerd van Straten, Christine Jacob, Marwan HassaniPredictive Process Monitoring Domain Adaptation
Department of Mathematics and Computer Science, Eindhoven University of Technology, The Netherlands · Department of Mathematics and Computer Science, University of Wuppertal, Germany
Revisiting Predictive Process Monitoring in the Age of Foundation Models: A Comparative Study of Sequence, Tabular, and LLM Approaches
Jul 30, 2026Lennart Fertig, Lukas Kirchdorfer, Tobias SesterhennTabular Foundation Models Predictive Process Monitoring
University of Mannheim, Mannheim, Germany · SAP Signavio, Walldorf, Germany · Technical University of Clausthal, Clausthal, Germany
On the Potential of Multi-Task Learning in Predictive Process Monitoring
Sep 11, 2026Lukas Kirchdorfer, Keyvan Amiri Elyasi, Heiner StuckenschmidtPredictive Process Monitoring Multi-Task Learning
University of Mannheim

Aug 28, 2026J/K move · Enter open · S save

D-TAIA: Domain-Aware LLM Adaptation for Multi-Task Predictive Process Monitoring

Sjoerd van Straten, Christine Jacob, Marwan Hassani

Department of Mathematics and Computer Science, Eindhoven University of Technology, The Netherlands · Department of Mathematics and Computer Science, University of Wuppertal, Germany

Predictive Process Monitoring Domain Adaptation Multi-Task Learning

Open paper AskPDF arXiv

Predictive Process Monitoring (PPM) enables organizations to forecast future process behavior, such as the next activity and remaining time of ongoing cases. In practice, three conditions cause existing methods to degrade, namely data scarcity, high process entropy and distributional shift. While Foundation Models (FMs), especially Large Language Models (LLMs), offer a new paradigm through broad sequential reasoning, adapting them to multi-task PPM under these conditions remains an open challenge. Existing FM-based approaches either lack mechanisms for handling distributional shift or rely on direct regression heads that can be structurally misaligned with continuous time prediction tasks. This paper introduces D-TAIA (Domain-aware Training and Attention-based Inference Architecture), a framework for a joint next activity and remaining time prediction task via parameter-efficient fine-tuning of an FM backbone. Our approach combines domain-aware triplet loss (DATL) pre-training with FAISS-based nearest neighbor retrieval for remaining time prediction, and adopts the TAIA inference strategy to preserve pre-trained sequential reasoning during fine-tuning. Evaluated across four real-world event logs, D-TAIA consistently shows SOTA or competitive performance compared to a fine-tuned LLM and a recurrent neural network baseline. Ablation studies confirm that techniques from NLP and computer vision can be transferred effectively to PPM with only a 10M-parameter backbone, though component contributions vary by dataset entropy.

Abstract

Predictive Process Monitoring

Explore similar work

D-TAIA: Domain-Aware LLM Adaptation for Multi-Task Predictive Process Monitoring

Revisiting Predictive Process Monitoring in the Age of Foundation Models: A Comparative Study of Sequence, Tabular, and LLM Approaches

On the Potential of Multi-Task Learning in Predictive Process Monitoring