CoSMo: A Framework to Instantiate Conditioned Process Simulation Models

Rafael Seidi Oyamada, Gabriel Marques Tavares, Sylvio Barbon Junior, Paolo Ceravolo

TL;DR

This paper introduces CoSMo, a novel framework that uses a conditioned recurrent neural network to simulate business processes while adhering to user-defined constraints. This approach addresses the limitations of traditional deep learning models in process simulation by allowing for more control and "what-if" analysis, moving beyond purely stochastic generation.

What Problem Does This Paper Address?

Process simulation is crucial for understanding and improving business processes, enabling tasks like compliance checking, performance analysis, and "what-if" scenario evaluation. Existing methods fall into categories like Data-driven Simulation (DDS) and Deep Learning (DL) based approaches. DDS models are interpretable but can oversimplify, while DL models excel at generalization but struggle with stochasticity and integrating external knowledge for controlled simulations, particularly for "what-if" analyses. This paper aims to bridge this gap by enabling DL-based process simulation with user-defined constraints, offering greater control and flexibility.

Methods

The core of CoSMo is a novel conditioned recurrent neural network architecture. This architecture extends a standard RNN by introducing a learnable parameter K that explicitly models the relationship between external constraint vectors (e.g., declarative rules) and the sequential event data. The framework involves:

  1. Preprocessing: Extracting declarative rules (using the DECLARE language) from event logs to serve as constraints.
  2. Conditioned Recurrent Network: A modified RNN that takes both event sequences and constraint vectors as input. The constraint vector is integrated at each time step, allowing the network to learn how these constraints influence the generation of subsequent events.
  3. CoSMo Framework: This framework orchestrates the entire process, from constraint extraction and network training to conditioned simulation. It allows users to specify constraints, which can be modified (e.g., by inverting rules) to perform "what-if" analyses.

Datasets

The authors evaluated CoSMo using five distinct real-world event logs from the 4TU.ResearchData repository:

  • BPI12 (complete)
  • BPI13 (problem version)
  • BPI17 (travel permit subset)
  • BPI20
  • SEPSIS

These datasets vary in size and characteristics (number of unique activities, events, cases, and average case length) to ensure a comprehensive evaluation.

Key Insights and Ideas

  • Bridging Stochasticity and Control: The primary insight is that deep learning models, while powerful for sequence generation, can be made controllable by explicitly incorporating external knowledge (constraints). This moves beyond purely stochastic simulation towards more purposeful and interpretable generation.
  • DECLARE as a Constraint Language: The paper demonstrates the effectiveness of using the DECLARE language, a declarative approach to process modeling, as a rich source of constraints for deep learning-based simulation. This leverages existing process mining techniques for a new application.
  • Generalization Beyond Training: CoSMo's ability to achieve high precision even when simulating with inverted constraints (rules that were not fulfilled in the training data) highlights its generalization capabilities. This is crucial for "what-if" scenarios where desired behaviors might deviate from historical data.
  • Template-Specific Challenges: The research reveals that certain DECLARE templates (e.g., Positive Relations) are more complex and lead to higher dimensionality in the constraint vector, making them harder for the model to learn and generalize.

Key Results

  • High Precision with Original Constraints: CoSMo achieved a median precision of 99% when simulating with the original set of DECLARE constraints, demonstrating its ability to accurately learn and replay existing process semantics.
  • Strong Generalization with Inverted Constraints: The model achieved a median precision of 81% when simulating with inverted constraints, indicating a significant capability to adapt to new, potentially unseen, rule sets. This is a key indicator of its "what-if" analysis potential.
  • Vanilla Baseline Performance: The vanilla simulation (without constraints) showed a median precision of only 58%, underscoring the limitations of purely stochastic generation and the benefit of incorporating constraints.
  • Data-flow Performance: The remaining time (RT) prediction, a data-flow aspect, showed comparable performance between CoSMo and the vanilla approach after initial convergence, suggesting that adding constraints doesn't significantly degrade this aspect of simulation.

Limitations and Future Work

  • Rule Correlation: The paper notes that correlations among DECLARE rules (e.g., existence and absence) can lead to redundant information in the constraint vector. Future work could involve preprocessing steps to eliminate these correlations for improved efficiency and accuracy.
  • Arbitrary Rule Selection: The selection of activities and rules for simulation was somewhat arbitrary (based on frequency quartiles). Future work should involve domain experts to ensure the relevance and impact of chosen rules in practical scenarios.
  • Broader Constraint Types: The current study focuses on declarative rules. Future research should explore a wider range of constraints, including time-related constraints and intra-case feature availability (e.g., resources).
  • Real-world Applicability: While methodologically validated, further exploration of CoSMo's applicability in real-world settings is needed.

Summary updated

BibTeX
@inproceedings{oyamada2024cosmo,
  author = {Rafael Seidi Oyamada and Gabriel Marques Tavares and Sylvio Barbon Junior and Paolo Ceravolo},
  title = {{CoSMo: A Framework to Instantiate Conditioned Process Simulation Models}},
  booktitle = {Bpm},
  year = {2024},
  doi = {10.1007/978-3-031-70396-6\_19},
}
Figure