• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Friday, November 20, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Technology

Swarm Intelligence Framework Clusters Event Logs to Improve Discovered Process Models

by
November 20, 2026
in Technology
Reading Time: 7 mins read
0
Swarm Intelligence Framework Clusters Event Logs to Improve Discovered Process Models

Swarm Intelligence Framework Clusters Event Logs to Improve Discovered Process Models

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

Modern organizations increasingly rely on process mining techniques to extract operational insights from recorded event logs, translating timestamps, case identifiers, and activity labels into formal process models. However, real-world operational environments frequently generate heterogeneous execution sequences characterized by alternative pathways, rework cycles, concurrent operations, and exceptional routines. When analysts attempt to discover a single end-to-end process model from such complex logs, the resulting diagrams often collapse into incomprehensible tangles of interconnected nodes and transitions, colloquially known as spaghetti models. To address this structural challenge, researchers have introduced PSOCluster, a behavior-oriented trace-clustering framework that couples an interpretable declarative representation with centroid-aware particle swarm optimization. By partitioning an event log into coherent sublogs before applying process discovery algorithms, the framework enables the discovery of focused, highly conformant process models. In an extensive empirical investigation using an operational laboratory dataset of 18,891 process traces, the proposed method achieved an average size-weighted process-model F1-score of 0.98 across 30 independent experimental runs, establishing a robust link between internal clustering mechanics and downstream model interpretability.

The foundational challenge in trace clustering lies in capturing the underlying behavior of execution sequences without drowning in redundant or high-dimensional data. Standard frequency-based vectors count how often activities occur but completely disregard the execution order, meaning that two traces with identical activities executed in reverse sequences appear identical. Conversely, advanced sequence-alignment and n-gram methods preserve order but can impose severe computational overhead on large logs while yielding numerical distance scores that lack clear semantic explanations. PSOCluster resolves this tension by adopting a declarative representation rooted in Declare-style behavioral templates. Rather than relying on rigid string matching, the framework evaluates traces against four distinct families of temporal rules: unary templates describing whether an activity exists, occurs at most once, or serves as the start or end event; order templates specifying response, precedence, and general succession relationships; immediate-order templates capturing direct succession through chain response and chain precedence; and coexistence templates checking mutual occurrence regardless of distance. These pairwise templates are evaluated under non-vacuous semantics, ensuring that satisfaction values reflect genuine behavioral occurrences rather than trivial vacuous truths.

To prevent combinatorial explosion when generating candidate rules across large activity vocabularies, the authors established a structured three-tier feature selection pipeline that runs entirely prior to optimization. Initially, the system instantiates candidate rules across the observed activities, which produced 2,128 candidate behavioral dimensions for the primary 19-activity laboratory dataset. An empirical support filter immediately discards rules satisfied by fewer than 10 percent or more than 99.5 percent of the traces, removing both exceedingly rare anomalies and near-constant patterns that offer zero discriminative power. The surviving rules are then evaluated using a Bernoulli entropy calculation that prioritizes dimensions with balanced binary distributions and high feature variance. Finally, a greedy sequential pruning step examines rules in descending entropy order, measuring the Jaccard similarity of their satisfaction columns and discarding any candidate that exhibits a similarity of 0.98 or greater with previously retained rules. In the laboratory experiment, this pipeline compressed 2,128 initial rules into just 67 informative behavioral dimensions, accomplishing a 96.85 percent dimensionality reduction while preserving multi-perspective behavioral coverage across activity presence, ordering, immediate transitions, and coexistence.

Once the behavioral feature space is fixed, the clustering challenge is formulated as an optimization problem solved through continuous particle swarm optimization. In conventional swarm clustering, candidate solutions either encode data-point assignments or optimize cluster centers using a single partition-level fitness score that treats the entire collection of centroids as an indivisible unit. PSOCluster departs from this paradigm by implementing a dual-memory architecture that explicitly decouples partition-level cognitive guidance from centroid-specific social learning. Each particle in the swarm represents a complete candidate configuration of K continuous centroids in the selected behavioral space. While each particle retains a personal-best memory based strictly on the global quality of its entire partition, the swarm simultaneously maintains an archive of centroid-indexed social memories. Each entry in this social archive is updated independently whenever an individual cluster achieves a higher local quality score than previously recorded for that centroid index, regardless of whether the overall partition containing it was successful. This design allows high-performing cluster centers to guide the search even when they emerge from suboptimal global configurations.

Evaluating candidate partitions during the search requires balancing geometric cluster separation against the preservation of dense local behavioral patterns. The framework accomplishes this through an adaptive Silhouette–Density objective function that governs particle fitness. For every trace in an event log, the algorithm precomputes a fixed-radius neighborhood in the binary feature space; using a radius of 0.9, neighbors correspond precisely to cases sharing identical rule-satisfaction profiles. During optimization, each cluster’s local density is quantified by the average proportion of within-cluster neighbors and normalized relative to the densest cluster in the candidate partition. This normalized density score is dynamically combined with the classic partition silhouette coefficient through an iteration-dependent weighting parameter. Starting with an initial silhouette weight of 0.6 and density weight of 0.4, the objective progressively increases the silhouette contribution up to 0.9 in the final evaluation round. Consequently, the swarm prioritizes discovering dense, recurrent execution patterns during early exploratory iterations before shifting focus toward refining sharp cluster separation and resolving ambiguous decision boundaries.

To maintain exploration diversity and prevent the swarm from stagnating in local optima, the optimization engine incorporates three complementary search mechanics. First, an inertia weight decreases linearly across the iteration budget from 0.8 down to 0.3, granting particles greater velocity and trajectory momentum during early search phases while encouraging fine-grained spatial exploitation as convergence approaches. Second, a coordinate-wise Gaussian mutation operator perturbs each centroid coordinate independently with an 8 percent probability, introducing minor stochastic adjustments that assist particles in escaping narrow geometric valleys. Third, the framework incorporates an explicit centroid repulsion mechanism to eliminate duplicate representatives. Following position updates and mutations, all centroid pairs are scanned; if any two centroids fall within a Euclidean distance of 0.1, the second centroid is relocated toward the data trace located farthest from any currently active centroid, accompanied by a small stochastic perturbation, and its velocity is reset to zero. This repulsion step prevents redundant centroids from collapsing into the same behavioral pocket and ensures broad spatial coverage.

The primary experimental evaluation was conducted on an operational healthcare dataset comprising 18,891 traces and approximately 7.2 events per case collected over six months across roughly 40 Laboratory Information Systems. Using a reference configuration of 10 particles, 50 evaluation rounds, six centroids, and symmetric acceleration coefficients of 1.8, the algorithm converged reliably from data-supported initial points to well-separated configurations. Diagnostic tracking demonstrated that mean pairwise inter-centroid distance expanded from an initial 4.49 to exploratory peaks exceeding 13.6 before stabilizing at 8.41, while mean centroid speed decreased from an early peak of 3.05 down to 0.88. Downstream process models were generated independently for each cluster using PM4Py’s Heuristics Miner with a 0.90 dependency threshold, and their quality was measured using unrounded token-based replay fitness and precision. Across 30 independent random seeds, PSOCluster maintained remarkable stability, achieving a mean size-weighted process-model F1-score of 0.98 with a standard deviation of 0.04 and a range between 0.91 and 0.99.

To establish competitive performance, the authors benchmarked PSOCluster against established conventional algorithms and adapted granularity-based methods under an identical 67-dimensional behavioral representation and evaluation pipeline. Conventional fixed-cluster baselines assigning all traces produced inferior process models: Spectral Clustering reached a weighted F1-score of 0.8729, Gaussian Mixture Models attained 0.8708, and standard K-Means achieved 0.8562. Furthermore, Spectral Clustering yielded severe partition imbalance, grouping 13,561 traces into one dominant cluster while isolating merely 125 traces in another. Emerging granularity-oriented techniques adapted for the study similarly struggled to translate geometric organization into process conformance, with an adapted Stacked Density Granulation pipeline scoring 0.8572 and a Regenerated Granular-Ball pipeline reaching 0.8137. Density-based algorithms such as HDBSCAN and DBSCAN exhibited a pronounced trade-off between conformance and data coverage; while HDBSCAN reached an apparent F1-score of 0.9557 on retained traces, it classified 3,075 traces (16.28 percent of the log) as unclustered noise, driving its coverage-adjusted F1-score down to 0.8001. In contrast, PSOCluster maintained complete log coverage across six balanced clusters without discarding outlier cases.

A systematic component-wise ablation analysis isolated the specific performance contributions of each algorithmic module across 30 random seeds. Stripping the density component to rely solely on silhouette optimization caused the steepest performance decline, dropping the mean weighted F1-score from 0.98 to 0.86 with increased variance. Replacing the centroid-specific social archive with a standard shared global best vector reduced performance to 0.89, confirming that retaining cluster-level guidance independently of global partition success is essential for guiding the swarm. Eliminating adaptive silhouette-density weighting degraded the score to 0.90, while disabling centroid repulsion and Gaussian mutation yielded scores of 0.93 and 0.95, respectively. Omitting entropy ranking and Jaccard pruning reduced scores to 0.94 and 0.92, verifying that representation filtering remains vital. Post-clustering inspection of external attributes confirmed that discovered clusters corresponded closely to operational laboratory pathways (such as Blood, Saliva, Biological, Biopsy, and Cerebrospinal Fluid streams) and distinct clinical categories (including HIV, Hepatitis, COVID-19, and Neonatal Sepsis), producing Petri nets that clearly differentiated linear diagnostic routines from complex looping workflows.

To evaluate methodological portability across diverse business domains, the authors executed a cross-dataset study across four publicly available benchmark logs: the BPI Challenge 2012 Financial log, the BPI Challenge 2020 Domestic Declarations log, a hospital Sepsis Cases log, and an IT Incident Management log. In each case, the rule extraction and feature pruning pipeline was applied independently, reducing candidate spaces by 93.4 to 99.2 percent down to compact representations of 13 to 100 features. PSOCluster achieved the highest size-weighted process-model F1-score on every evaluated benchmark, achieving 0.9567 on Financial, 0.9931 on Domestic Declarations, 0.9107 on Sepsis, and 0.9661 on Incident Management, yielding an unweighted cross-dataset mean of 0.9537. Despite these encouraging results, the authors noted key technical limitations: the framework requires a predefined cluster count K, utilizes Euclidean nearest-centroid assignments that cannot capture arbitrary non-convex geometries, and was evaluated using independent log reconstructions rather than direct transfer of pre-trained models. Nevertheless, the study demonstrates that combining interpretable declarative behavioral encoding with centroid-aware swarm search effectively resolves heterogeneous event logs into highly conformant, transparent process models.

Subject of Research: Behavior-oriented trace clustering using centroid-aware particle swarm optimization in process mining

Article Title: A centroid aware swarm intelligence approach for behavior oriented trace clustering

Article References: Sarabi, A., Sheikhalishahi, M., & Tajvidi, M. (2026). A centroid aware swarm intelligence approach for behavior oriented trace clustering. Discover Artificial Intelligence, 6(1), Article 1915. https://doi.org/10.1007/s44163-026-02778-9

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02778-9

Keywords: process mining, trace clustering, particle swarm optimization, Declare rules, behavioral representation, event logs, process discovery, conformance checking, replay fitness, precision, F1-score, Heuristics Miner, Jaccard pruning, Bernoulli entropy, silhouette coefficient, neighborhood density, laboratory information systems, operational workflows

News Source: Denise Maddox. (November 20, 2026). Swarm Intelligence Framework Clusters Event Logs to Improve Discovered Process Models. Scienmag.

Tags: behavioral representationBernoulli entropyconformance checkingDeclare rulesevent logsF1 scoreHeuristics MinerJaccard pruningLaboratory information systemsneighborhood densityoperational workflowsparticle swarm optimizationprecisionprocess discoveryprocess miningreplay fitnesssilhouette coefficienttrace clustering
Share12Tweet7Share2ShareShareShare1

Related Posts

Cosmological Grassmannian Simplifies Calculations for Cosmological Collider Physics and Bootstrap Correlators

Cosmological Grassmannian Simplifies Calculations for Cosmological Collider Physics and Bootstrap Correlators

November 20, 2026
Cooling Simulations Reveal How Thermal Stresses Initiate Cracks in Multiphase Materials

Cooling Simulations Reveal How Thermal Stresses Initiate Cracks in Multiphase Materials

November 20, 2026

Banach Fixed Point Theorem Extended to Fractional Integral Contractions

November 20, 2026

Researchers Extend Positive Geometries to Branch Cuts and String Scattering Amplitudes

November 20, 2026

POPULAR NEWS

  • Gene Knockouts and Buffer Systems Boost Bacillus subtilis Spore Display Yields

    Gene Knockouts and Buffer Systems Boost Bacillus subtilis Spore Display Yields

    29 shares
    Share 12 Tweet 7
  • Narrative Review Integrates Imaging, Biomechanics, and Movement in Lumbar Spondylolysis

    29 shares
    Share 12 Tweet 7
  • Bifrontal Brain Stimulation Improves Working Memory in Schizophrenia Without Sex Differences

    29 shares
    Share 12 Tweet 7
  • Robotic and Conventional Knee Replacements Show Comparable One-Year Joint Infection Rates

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Warm Incremental Sheet Forming Produces Accurate Biodegradable Cranial Implants from Magnesium

Cosmological Grassmannian Simplifies Calculations for Cosmological Collider Physics and Bootstrap Correlators

Machine Learning Accurately Identifies Regional Patterns and Predictors of Childhood Diarrhea

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 89 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.