Jochen De Weerdt

Large language models have shown strong results in predictive process monitoring, but their computational cost raises questions about their practical value compared with smaller sequence models trained from scratch. This paper studies whether parameter-efficient fine-tuning (PEFT) and quantization can provide a suitable balance between predictive performance and computational cost in multi-task predictive process monitoring. We evaluate recurrent networks, Transformers, PEFT-based LLMs, and a process-specific foundation model across six real-world event logs and three tasks: next-activity, next-resource, and remaining-time prediction. We assess predictive performance, hyperparameter sensitivity, peak GPU memory usage, and the total runtime of different model-selection procedures. The results show that PEFT-based LLMs are mainly sensitive to the learning rate, whereas LoRA rank, target modules, numerical precision, and most backbone choices have smaller effects. This supports replacing broad PEFT searches with a focused learning-rate search. Quantization reduces peak GPU memory usage by up to 70.9\% with limited predictive loss in most settings. A focused PEFT procedure can also be competitive in total runtime with broad hyperparameter searches over models trained from scratch. Finally, task-level results show that aggregate scores in multi-task may hide effects related to class imbalance and resource-label structure. Overall, quantized PEFT offers a practical and competitive approach for adapting pretrained LLMs to multi-task predictive process monitoring.

Vector space representations play an important role in process mining tasks such as clustering, predictive monitoring, and anomaly detection. Existing representational approaches, however, are usually centered on predefined notions of similarity, often associated with control-flow behavior or fixed graph structures, limiting the explicit incorporation of domain knowledge into the representation construction process. This paper proposes FlexSimCase2Vec, a flexible similarity-based representational scheme for process mining that constructs case neighborhoods from user-defined similarity perspectives and supports weighting mechanisms to regulate the influence of different perspectives on the resulting vector space. The generated neighborhood structures are then used to learn vector representations of process cases. The proposed approach was evaluated through trace clustering and anomaly detection studies using both real-world and synthetic event logs. Experimental results show that FlexSimCase2Vec achieves competitive performance in control-flow-oriented scenarios while enabling the incorporation of multiple similarity perspectives without substantially deteriorating vector space organization. The results also indicate that the weighting mechanism allows explicit control over the trade-off between different notions of similarity according to the analytical objectives. These findings suggest that FlexSimCase2Vec provides a flexible and interpretable alternative for vector space representation in process mining, particularly in scenarios requiring the incorporation of domain knowledge and multiple process perspectives.

Vector space representations play an important role in process mining tasks such as clustering, predictive monitoring, and anomaly detection. Existing representational approaches, however, are usually centered on predefined notions of similarity, often associated with control-flow behavior or fixed graph structures, limiting the explicit incorporation of domain knowledge into the representation construction process. This paper proposes FlexSimCase2Vec, a flexible similarity-based representational scheme for process mining that constructs case neighborhoods from user-defined similarity perspectives and supports weighting mechanisms to regulate the influence of different perspectives on the resulting vector space. The generated neighborhood structures are then used to learn vector representations of process cases. The proposed approach was evaluated through trace clustering and anomaly detection studies using both real-world and synthetic event logs. Experimental results show that FlexSimCase2Vec achieves competitive performance in control-flow-oriented scenarios while enabling the incorporation of multiple similarity perspectives without substantially deteriorating vector space organization. The results also indicate that the weighting mechanism allows explicit control over the trade-off between different notions of similarity according to the analytical objectives. These findings suggest that FlexSimCase2Vec provides a flexible and interpretable alternative for vector space representation in process mining, particularly in scenarios requiring the incorporation of domain knowledge and multiple process perspectives.

Process Model Forecasting (PMF) aims to predict how the control-flow structure of a process evolves over time by modeling the temporal dynamics of directly-follows (DF) relations, complementing predictive process monitoring that focuses on single-case prefixes. Prior benchmarks show that machine learning and deep learning models provide only modest gains over statistical baselines, mainly due to the sparsity and heterogeneity of the DF time series. We investigate Time Series Foundation Models (TSFMs), large pre-trained models for generic time series, as an alternative for PMF. Using DF time series derived from real-life event logs, we compare zero-shot use of TSFMs, without additional training, with fine-tuned variants adapted on PMF-specific data. TSFMs generally achieve lower forecasting errors (MAE and RMSE) than traditional and specialized models trained from scratch on the same logs, indicating effective transfer of temporal structure from non-process domains. While fine-tuning can further improve accuracy, the gains are often small and may disappear on smaller or more complex datasets, so zero-shot use remains a strong default. Our study highlights the generalization capability and data efficiency of TSFMs for process-related time series and, to the best of our knowledge, provides the first systematic evaluation of temporal foundation models for PMF.