Akademik Veri Yönetim Sistemi
2025 Innovations in Intelligent Systems and Applications Conference, ASYU 2025, Bursa, Türkiye, 10 - 12 Eylül 2025, (Tam Metin Bildiri)
Detection of anomalies is crucial for maintaining operational integrity, however, existing state-of-the-art solutions require high-level computational resources, which hinders their deployment in constrained environments. In this paper, we introduce LeanTrace, a novel, resource-conscious framework for business process anomaly detection on a trace-level. LeanTrace systemically evaluates five embedding generation strategies: Word2Vec, FastText, GloVe, Node2Vec, and DeepWalk across dimensionalities ranging between 16 to 256. The methodology combines these embeddings with five distinct anomaly detectors: Isolation Forest, Local Outlier Factor, Elliptic Envelope, One-Class Support Vector Machine, and a cluster based anomaly detector configured to use K-Means, K-Medoids, and CLARA clustering algorithms. Through comprehensive testing on one public and one private dataset, LeanTrace configured with Fast-Text and Local Outlier Factor achieves a performance of up to a 0.69 F1-score and 0.83 AUC, where the CLARA based anomaly detector yields recall scores of up to 0.95. Furthermore, running on a low power ARM Cortex-A58, LeanTrace takes less than a second to encode over 10.000 traces and completes predictions as fast as 0.25 seconds. By balancing detection accuracy and minimal footprint, LeanTrace paves the way for resource-constrained deployment.