Predictive Process Monitoring (PPM) forecasts how ongoing organizational processes unfold, enabling information systems to move beyond execution support toward proactive analysis and monitoring. Although deep learning has improved prediction accuracy in PPM, most approaches follow a single-task learning (STL) setup, training a separate model per task. This increases maintenance effort and overlooks potential synergies. Multi-task learning (MTL), which jointly learns multiple prediction targets in one model, offers a promising alternative, yet its effectiveness in PPM remains underexplored. It remains unclear whether and under which settings MTL improves upon STL, which prediction tasks benefit most from joint learning, which task combinations are particularly synergistic, and if and how tasks should be balanced. To fill this gap, we present the first comprehensive empirical study of MTL for PPM, evaluating a variety of task combinations, neural architectures, and optimization methods. Overall, our results position MTL as a strong paradigm for PPM: we see substantial improvements in next-activity prediction and inherent mitigation of class imbalance using MTL, while task balancing is especially critical under low-capacity models.
Figures & tables
Tasks
Arch.
STL
Com.
MTO
Analysis
Tax et al. (2017)
2
L
✓
EW
Gherissi et al. (2022)
3
L
EW
Camargo et al. (2019)
3
L
✓
∼
EW
Taymouri et al. (2020)
2
L
EW
Hennig and Schmidt (2025)
3
T
EW
Wang et al. (2024)
3
T
static
Table 1: Comparison of existing PPM approaches that apply MTL. Arch.: L = LSTM, T = Transformer, C = CNN. STL indicates whether MTL is compared against its single-task baseline. Com. indicates whether all task combinations are evaluated. Analysis indicates whether the study analyzes when and why MTL is beneficial.
Figure 1: Multi-task setup with shared backbone (LSTM/CNN/Transformer) and task-specific heads.
Event log
Cases
Events
ACT
RES
Variants
ALT
P2P
608
9119
21
27
70
21.5
Production
225
4503
24
41
217
20.6
Helpdesk
4580
21348
14
22
226
40.9
Sepsis
1050
15214
16
26
846
28.5
BPIC20DD
10500
56437
17
7
99
11.5
BPIC20ID
6449
72151
34
8
753
86.5
Table 2: Characteristics of the employed event logs: ACT = distinct activities, RES = distinct resources, ALT = average lead time of process instances (days).
Figure 2: MTL performance by model architecture, measured by Δm .
Model
NAP
NTP
RTP
LSTM
0.004
0.250
0.055
CNN
0.004
0.570
0.734
Transformer
0.008
0.008
0.742
Table 3: Comparison of MTL and STL using the Wilcoxon signed-rank test. P-values are based on all task combinations and MTO methods.
Log
MTO
NAP
NTP
RTP
Δm↓
ACC ↑
macro-F1 ↑
MAE ↓
MAE ↓
P2P
STL
0.73 (0.01)
0.65 (0.01)
1.07 (0.08)
4.47 (0.11)
/
EW
0.86 (0.00)
0.78 (0.01)
1.08 (0.07)
4.73 (0.15)
-4.19
GradDrop
0.87 (0.01)
0.78 (0.01)
1.03 (0.02)
4.41 (0.09)
-8.05
IMTL
0.85 (0.01)
0.78 (0.01)
1.07 (0.08)
5.13 (0.52)
-1.09
Production
STL
0.44 (0.01)
0.21 (0.01)
0.99 (0.16)
4.72 (0.29)
/
Table 4: Results for LSTM. For each log, we show the average results over three random seeds (and the standard deviation) for STL, EW, and the best/worst MTO method according to Δm . MAE reported in days.
Figure 3: Comparison of NAP’s train and validation loss for STL and MTL (with EW) on two datasets. The lines represent the train loss and the small shaded area captures the gap to the validation loss, respectively.
Figure 4: Average model uncertainty per dataset on the respective test set, measured by the normalized entropy. Results are based on LSTM models with EW.
Figure 5: Per-class F1-score comparison between MTL and STL models using CNN architecture for BPIC12C. The MTL model employs the NAP+NTP+RTP task combination. Activity classes are sorted by frequency.
Figure 6: Impact of task combinations on performance. Results are based on EW with the LSTM model.
Figure 7: Box plots showing Δm per MTO method, aggregated over all architectures and task combinations.
Figure 8: Δm performance for the large (L) and small (S) variants of the CNN model. We show the best and worst performing MTO methods as well as EW.
Finding
Recommendation
MTL mainly improves NAP, only minimal impact on temporal tasks.
Use MTL particularly when activity prediction is central.
MTL improves performance on infrequent activities.
Consider MTL when rare but relevant activities matter.
Equal weighting is often competitive.
Start with simple EW before using more complex MTO methods.
Task balancing matters under limited model capacity.
Use adaptive weighting when models are small or data is complex.
Task-wise effects can differ.
Monitor each task separately rather than relying only on aggregate scores.
Table 5: Key findings and practical recommendations.
Predictive process monitoring (PPM) leverages event logs to forecast the future of running process instances, for instance, predicting the next activity, the remaining time until case completion, or the time to the next event. While PPM research in recent years has been dominated by deep sequence models trained from scratch, such as Long Short-Term Memory (LSTM) models, foundation-model approaches---particularly large language models (LLMs)---are increasingly explored for PPM. At the same time, tabular foundation models with in-context learning capabilities offer a promising alternative but have not yet been systematically benchmarked for PPM. Thus, it remains unclear whether classical sequence-based models remain competitive in this evolving landscape. This paper compares the three modeling paradigms both conceptually and empirically through a controlled benchmark across multiple datasets and prediction tasks. The results show that sequence models consistently perform best for next activity prediction, whereas tabular foundation models are competitive on temporal tasks, with LLMs usually lagging behind despite higher cost.
Lennart Fertig, Lukas Kirchdorfer, Tobias Sesterhenn
University of Mannheim, Mannheim, Germany · SAP Signavio, Walldorf, Germany · Technical University of Clausthal, Clausthal, Germany
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.
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 (PPM) aims at predicting at runtime and as early as possible the future states of a process execution. Common tasks include predicting the next event, the time to completion of a trace, and outcomes. Existing approaches typically consider an event from the perspective of the executed activities along with their timestamps and case identifiers. This leads to the disadvantage that in real-life settings, there is much more information recorded in the event log that is not captured or completely ignored when performing prediction tasks. We introduce GNN4PPM, an approach that predicts all next events along with their complete data payload at once. We represent event information in a heterogeneous knowledge graph that captures the event log as an RDF semantics, and train the embeddings with a Relational Graph Convolutional Network (R-GCN). Our approach is promising in comparison to existing solutions, and experiments with state-of-the-art solutions prove the accuracy and applicability of GNN4PPM in complex settings.
Ana Costa, Johannes Mäkelburg, Luise Pufahl
Technical University of Munich, School of Computation, Information and Technology, Heilbronn, Germany