MinerSweeper Automates Business Process Discovery Pipelines Through Multi-Objective Metaheuristic Optimization

Researchers have developed and evaluated MinerSweeper, an open-source software tool designed to automate business process discovery pipelines using multi-objective evolutionary optimization. Process discovery extracts workflow models, such as Petri nets, from event logs recorded by enterprise systems to depict underlying operational behaviors. In conventional workflows, analysts manually select preprocessing filters, choose discovery algorithms, and configure internal hyperparameters. This manual approach creates a vast combinatorial search space that cannot be explored exhaustively. MinerSweeper reframes process discovery as an objective-driven search, orchestrating preprocessing, algorithmic selection, and parameter calibration into a unified multi-objective pipeline. By applying evolutionary optimization, the platform discovers non-dominated process models approximating the Pareto front across competing quality criteria, enabling stakeholders to evaluate balanced trade-offs rather than relying on arbitrary default settings.

This automated orchestration addresses a recognized gap within process mining. While hyperparameter optimization is foundational in machine learning, its application to business process discovery has remained underdeveloped. Standard open-source frameworks like ProM and PM4Py offer mature implementations of individual discovery algorithms but lack automated mechanisms to coordinate and evaluate full discovery pipelines. Prior efforts attempted partial automation through algorithm selection frameworks that overlooked internal hyperparameters, or examined hyperparameter tuning on limited logs without providing reusable tools. Other tools explored declarative models or integrated preprocessing with algorithm selection, yet encountered severe computational scaling limits. MinerSweeper resolves these limitations by treating preprocessing strategies, algorithm families, specific variants, and their associated continuous and categorical hyperparameters as an integrated search space evaluated against multiple conflicting quality criteria.

Automating this end-to-end workflow overcomes three primary challenges identified by the authors. First, discovered models must align with specific organizational requirements, which can be defined by weighting process quality metrics. Second, the space of possible pipeline configurations is too vast for exhaustive evaluation in realistic computational settings. Third, process discovery involves multiple conflicting quality metrics, where improving one dimension inevitably impairs another. Because analytical goals vary across organizations, a single universal reference model does not exist in process mining. MinerSweeper addresses these constraints by applying metaheuristic optimization to approximate the Pareto front of optimal solutions. These candidate configurations represent non-dominated solutions where no alternative pipeline improves a given quality metric without worsening at least one other, allowing practitioners to establish their evaluation reference model after the search.

Developed under the Design Science Research methodology, MinerSweeper operates as a modular, containerized application distributed across four microservices connected via REST interfaces. A client-side Frontend Service built with ReactJS and Vite delivers a web interface accessible locally on port 5173. The interface allows users to launch optimization tasks, monitor progress, inspect historical runs, and view discovered Petri nets. Optimization workflows are managed by the Optimisation Service, a Python application built on Flask and the jMetalPy library. This service formulates the multi-objective problem, coordinates asynchronous evolutionary execution, and assigns candidate pipelines to downstream mining workers. Persistent storage is provided by a PostgreSQL database, which records experiment metadata, final Pareto-optimal configurations, and intermediate population snapshots to support reproducibility, auditing, and multi-level result inspection.

The Optimisation Service employs the Non-dominated Sorting Genetic Algorithm III, or NSGA-III, as its core metaheuristic. The researchers selected NSGA-III because its reference-point mechanism effectively resolves many-objective problems involving four or more conflicting metrics. Pipeline configurations are encoded as flat continuous vectors representing decision variables for preprocessing operations, discovery algorithm families, specific variants, and active hyperparameters. To navigate this conditional space, MinerSweeper uses a hierarchical gating mechanism during decoding, parsing only the genes that govern active components while ignoring inactive variables. Internal safeguards adjust log-dependent thresholds to prevent filters from discarding entire event traces. When candidate configurations encounter plugin failures, memory exhaustion, or semantic incompatibilities, the system assigns a penalty score of zero, ensuring invalid solutions are heavily dominated and discarded without halting execution.

By default, the NSGA-III implementation uses a population of 100 individuals and terminates after 1,000 evaluations. Evolutionary operators are managed statically: binary tournament selection directs mating, Simulated Binary Crossover operates with a probability of 0.9 and distribution index of 20.0, and Polynomial Mutation runs with a distribution index of 20.0 and probability of one over the number of decision variables. Reference directions are systematically generated using eight partitions for four or more objectives and twelve partitions otherwise. To ensure exact replication across stochastic runs, the platform tracks random generator seeds sequentially while evaluating candidate configurations concurrently across worker threads. Persisting the random seed, experiment metadata, and metric selections within PostgreSQL ensures practitioners can replicate identical optimization trajectories across independent executions.

Candidate pipeline configurations are evaluated by the Mining Service, a lightweight REST application implemented in Java 8 using SparkJava. Java 8 ensures compatibility with ProM plugins and JAR dependencies running in headless mode. The service ingests event logs in standard XES format, executes specified preprocessing chains, applies the chosen discovery algorithm, and outputs process models as Petri nets in PNML format. The service computes designated quality metrics using either alignment-based or replay-based conformance checking. Because NSGA-III minimizes objectives, metrics intended for maximization are multiplied by negative one, producing negative values in raw inspection views. Standard quality dimensions are normalized between zero and one, whereas structural metrics retain their operational values and enter the optimization directly as formal mathematical objectives rather than descriptive post-hoc indicators.

MinerSweeper assesses discovered models across four standard process mining dimensions: fitness, precision, simplicity, and generalization. Fitness measures the proportion of observed log behavior reproduced by the model, precision verifies the model avoids extraneous behavior, simplicity quantifies structural comprehensibility, and generalization tests the model’s ability to accommodate future behavior without overfitting. Following optimization, the platform uses an Augmented Scalarization Function to streamline model selection. This mechanism allows practitioners to assign weights to each quality dimension according to operational priorities, transitioning from the multi-dimensional Pareto front to a single preferred business process model. Users can inspect solutions grouped by identical metric values, visualize the Pareto front via parallel coordinates, and export complete experimental datasets, including run metadata and evolutionary snapshots, as standard CSV files.

The platform was evaluated empirically across six publicly available event logs representing diverse operational contexts without manual preprocessing. The benchmark collection included three datasets from the Business Process Intelligence Challenge: Open (819 cases, 2,351 events, 3 activities, 13.2 percent variants), Closed (1,487 cases, 6,660 events, 4 activities, 12.3 percent variants), and Incidents (7,554 cases, 65,533 events, 4 activities, 20.0 percent variants). The evaluation also incorporated three larger datasets: the Hospital log (100,000 cases, 451,359 events, 18 activities, 1.0 percent variants), the Road Traffic Fine Management Process log (150,370 cases, 561,470 events, 11 activities, 0.2 percent variants), and the Sepsis log (1,050 cases, 15,214 events, 16 activities, 80.6 percent variants). Across ten independent runs per dataset, optimization showed stable convergence across all logs.

Convergence was assessed using hypervolume against a reference point of zeros across all four quality dimensions. Hypervolume scores showed low variance across ten independent runs, yielding mean values of 0.9309 for Open, 0.9732 for Closed, 0.9731 for Incidents, 0.9719 for Hospital, 0.9726 for RTFMP, and 0.9050 for Sepsis. Execution times reflected log scale, with average runtimes ranging from 6.14 seconds for Open and 10.46 seconds for Closed to 498.52 seconds for Incidents, 122.98 seconds for Sepsis, 3,733.42 seconds for Hospital, and 6,263.62 seconds for RTFMP. While hypervolume remained consistent across runs, execution times varied substantially on logs like Sepsis (standard deviation of 135.08 seconds) and Incidents (standard deviation of 487.49 seconds), reflecting the variable computational expense of candidate pipelines sampled during stochastic exploration.

Comparative benchmarking against the default Inductive Miner baseline demonstrated the advantages of multi-objective optimization. While the baseline produced imbalanced models, MinerSweeper identified configurations optimizing specific metrics. On the Open dataset, the baseline scored 1.000 fitness, 0.943 precision, 0.692 simplicity, and 0.933 generalization, whereas MinerSweeper found models achieving 1.000 precision and simplicity with 0.963 generalization. For the Hospital log, where the baseline reached 1.000 fitness but only 0.563 precision and 0.626 simplicity, MinerSweeper identified configurations achieving 1.000 precision, 1.000 simplicity, and 0.996 generalization. On the high-variant Sepsis log, the baseline produced 0.433 precision and 0.622 simplicity, while MinerSweeper generated solutions achieving 1.000 precision or 1.000 simplicity with 0.966 precision, demonstrating substantial improvements over default single-model configurations.

Parallelization experiments evaluating one to sixteen local worker threads demonstrated that concurrency benefits scale with problem size. For simple logs like Open and Closed, speed gains were marginal due to thread-coordination overhead. Conversely, evaluating the Hospital dataset dropped from over 35 minutes with a single worker to approximately 16 minutes with eight workers, exceeding a two-fold speedup. However, scaling to sixteen workers caused memory saturation within the Java Virtual Machine. Because each worker instantiates independent ProM plugins and process trees, high worker counts on dense logs create severe heap pressure, triggering garbage-collection pauses and memory swapping that degrade performance. The authors noted that future research must address the computational demands of expanding search spaces, incorporate broader structural metrics, and refine model selection to balance technical quality with domain interpretability.

Subject of Research: Multi-objective optimization of business process discovery pipelines

Article Title: MinerSweeper: Multi-objective pipeline optimisation for business process discovery

Article References: Jiménez-Jiménez, R., Ramos-Gutiérrez, B., Gómez-López, M. T., Varela-Vaca, Á. J., & Nebro, A. J. (2026). MinerSweeper: Multi-objective pipeline optimisation for business process discovery. SoftwareX, 36, Article 103187. https://doi.org/10.1016/j.softx.2026.103187

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103187

Keywords: process mining, business process discovery, multi-objective optimization, NSGA-III, hyperparameter tuning, Petri nets, software engineering, design science research, MinerSweeper, Multi-objective, pipeline, optimisation