Zare, Hadi2026-05-062026-05-062026https://hdl.handle.net/1828/23807In the field of Business Process Management (BPM), accurately predicting the next activity in an ongoing process is critical for improving operational efficiency, optimizing resource allocation, and enabling proactive decision-making. Although recent advances in machine learning (ML) and artificial intelligence (AI) have significantly improved predictive performance, several key challenges remain. These include: (i) the limited transparency of deep learning models when applied to datasets that may not require such complexity; (ii) the reliance on trial-and-error methods for selecting suitable models across diverse event logs; and (iii) the underutilization of attributes that can carry valuable information for prediction. Building on these identified gaps, this thesis is motivated by two central research questions: (1) How can predictive models be designed to effectively capture the complexity and variability inherent in modern event logs? and (2) How can organizations systematically determine the most suitable predictive models for their specific process characteristics? To address the first question, this thesis introduces a novel predictive architecture called the Dynamic Attribute-Wise Transformer (DAW-Transformer). The model enhances predictive capability by extending the standard transformer architecture through the integration of multi-head attention and a dynamic windowing mechanism tailored to each dataset. This design captures long-range dependencies across multiple event attributes, offering a richer and more detailed representation of process behavior and improving the model’s ability to generalize across heterogeneous logs. To address the second question, an Entropy-Driven Model Selection Framework is proposed. This framework employs process entropy as a quantitative indicator of event log complexity, enabling adaptive model selection that balances predictive accuracy and interpretability. By aligning model choice with dataset variability, it overcomes the limitations of existing approaches that apply a single predictive model indiscriminately across all process types, regardless of their structural diversity. The effectiveness of the proposed methods is validated through comprehensive experiments on six publicly available event logs across domains such as healthcare, logistics, and public administration. Results demonstrate that the DAW-Transformer achieves superior performance, particularly on high-entropy processes, where activities exhibit greater variability, while interpretable models such as Decision Trees perform competitively on low-entropy, structured processes. These findings highlight the value of aligning model complexity with process entropy and underscore entropy’s role as a guiding principle for model selection in predictive business process monitoring. In summary, this thesis contributes to: (1) improving interpretability and reducing trial-and-error model selection via an Entropy-Driven Model Selection Framework; (2) mitigating attribute underutilization through the DAW-Transformer; and (3) conducting comprehensive evaluations across diverse datasets. Together, these contributions enhance predictive accuracy, strengthen interpretability, and establish a data-driven foundation for adaptive model choice.enAvailable to the World Wide Webnext activity predictionbusiness process monitoringtransformersThesisZare, H., Abbasi, M., Ahang, M., & Najjaran, H. (2025). An innovative next activity prediction approach using process entropy and daw-transformer. arXiv preprint arXiv:2502.10573.