سامانه همایش های دانشگاه آزاد اسلامی واحد نجف آباد

صفحه اصلی / سومین کنفرانس بین المللی مهندسی و علوم کامپیوتر و اولین کنفرانس ملی سامانه های هوشمند، کلان داده و رایانش پیشرفته

A Lightweight and Interpretable Approach for Anomaly Detection in Distributed Traces

نویسندگان :

Sobhan Kiani¹

1- دانشگاه صنعتی امیرکبیر (پلی‌تکنیک تهران)

کلمات کلیدی :

Distributed Tracing،Anomaly Detection،Graph Analysis،Online Learning

چکیده :

Distributed tracing is essential for observing microservice-based systems, yet anomaly detection remains challenging due to high-dimensional traces, scarce labels, and strict latency constraints. Existing deep learning methods often achieve high accuracy but are computationally expensive, hard to interpret, and poorly suited for online deployment. This paper proposes TAA, a lightweight and interpretable anomaly detection framework for distributed tracing data. TAA models each trace as a graph of service interactions and extracts domain-informed structural and performance features to produce fixed-size representations for supervised learning. The framework supports both offline batch training and online incremental learning, enabling low-latency, per-trace anomaly detection. Evaluated on the Train-Ticket benchmark with injected faults and limited labeled data, TAA outperforms deep, unsupervised, and semi-supervised baselines in F1-score while reducing processing time. The online variant maintains stable performance and adapts effectively as new labels become available, making TAA well suited for real-time production monitoring.