COMPARATIVE ANALYSIS OF SEQUENCE-BASED AND GRAPH-BASED DEEP LEARNING FOR ANOMALY DETECTION IN MICROSERVICE TRACES

Yesset Zhussupov; Ainur Zhumadillayeva; Shona Shinassylov; Ainur Amirtayeva

doi:10.26577/jpcsit4220265

Авторы

Yesset Zhussupov L.N. Gumilyov Eurasian national university, Astana, Kazakhstan https://orcid.org/0009-0001-0946-8094 (неаутентифицированный)
Ainur Zhumadillayeva L.N. Gumilyov Eurasian national university, Astana, Kazakhstan https://orcid.org/0000-0003-1042-0415 (неаутентифицированный)
Shona Shinassylov Al-Farabi Kazakh National University, Almaty, Kazakhstan https://orcid.org/0009-0002-5491-7255 (неаутентифицированный)
Ainur Amirtayeva Al-Farabi Kazakh National University, Almaty, Kazakhstan https://orcid.org/0009-0008-4579-287X (неаутентифицированный)

DOI:

https://doi.org/10.26577/jpcsit4220265

Ключевые слова:

microservice architecture, anomaly detection, graph neural networks, LSTM, distributed tracing, deep learning, observability

Аннотация

Переход от монолитных к микросервисным архитектурам значительно увеличил сложность мониторинга системы. Распределенная трассировка стала важным инструментом для мониторинга межсервисных взаимодействий, однако традиционные методы обнаружения аномалий часто не способны выявить сложные структурные взаимосвязи, присущие этим системам. В этой статье представлен всесторонний сравнительный анализ подходов к машинному обучению на основе последовательностей, в частности, сетей долговременной кратковременной памяти (LSTM), с методами глубокого обучения на основе графов, такими как графовые нейронные сети (GNN), для обнаружения аномалий в трассировках микросервисов. Мы оцениваем эти парадигмы по трем критически важным параметрам: структурная осведомленность, точность обнаружения и вычислительная эффективность. Наш анализ, основанный на современных фреймворках, включая DeepTraLog и TraceVAE, демонстрирует, что модели на основе графов обеспечивают более высокую структурную точность и достигают более высоких показателей точности-полноты, сохраняя при этом масштабируемость для многомерных телеметрических данных. Результаты показывают, что подходы на основе GNN достигают показателей F1-меры до 0,954, значительно превосходя альтернативные подходы на основе последовательностей, которые обычно достигают показателей F1-меры около 0,741. Эти результаты имеют важное значение для проектирования платформ мониторинга следующего поколения для облачных приложений.

Скачивания

Данные по скачиваниям пока не доступны.

Биографии авторов

Yesset Zhussupov , L.N. Gumilyov Eurasian national university, Astana, Kazakhstan

Yesset Zhussupov is a PhD student in Software Engineering at L.N. Gumilyov Eurasian National University (Almaty, Kazakhstan; yesset.zhussupov@gmail.com). He received his MSc in Software Engineering from the same university in 2022. Mr. Zhussupov serves as Chief Technology Officer at an AFSA-licensed brokerage organization, where he architects systems integrating big data and machine learning technologies. His research focuses on neural network applications in finance and economics, with particular emphasis on market forecasting and predictive modeling. His recent work includes the paper "Study of the Possibilities of Using Deep Artificial Intelligence in Forecasting the Green Paper Market in Kazakhstan." In 2015, he received the Intel Excellence Award for his work applying neural networks to computer vision–based emotion recognition.
Ainur Zhumadillayeva, L.N. Gumilyov Eurasian national university, Astana, Kazakhstan

PhD Dr. Ainur Zhumadillayeva is an Associate Professor of Associate Professor of Computer and Software Engineering at L. N. Gumilyov Eurasian National University (Astana, Kazakhstan). She received her PhD degree in the specialty *05.13.18 – Mathematical Modeling, Numerical Methods, and Software Packages* from the Institute of Mathematics of the Ministry of Education and Science of the Republic of Kazakhstan (Almaty) in 2010. Ainur Zhumadillayeva is the author and coauthor of educational publications and scientific monographs and has published more than 60 scientific and educational-methodical works in Kazakhstan and abroad, including publications indexed in international databases such as Scopus and Clarivate Analytics. Her research interests include mathematical modeling, numerical methods, machine learning, and artificial neural networks. She is a Member of the Institute of Electrical and Electronics Engineers (IEEE).
Shona Shinassylov, Al-Farabi Kazakh National University, Almaty, Kazakhstan

Shona Shinassylov is a Researcher and Lecturer at the Faculty of Information Technology and Artificial Intelligence, Al-Farabi Kazakh National University, Almaty, Kazakhstan. He received his Master's degree in Information Technology from Al-Farabi Kazakh National University and completed his doctoral studies in 2025. His research interests include digital twins, artificial intelligence, machine learning, deep reinforcement learning, smart buildings, the Internet of Things (IoT), energy management systems, and intelligent decision-support systems. His current research focuses on the design and development of virtual prototypes of smart building energy management systems based on digital twin technologies, machine learning methods, and multi-objective reinforcement learning.
Ainur Amirtayeva, Al-Farabi Kazakh National University, Almaty, Kazakhstan

Ainur Amirtayeva is a Master's student in the dual-degree program between Al-Farabi Kazakh National University (KazNU) and Northwestern Polytechnical University (NWPU) (Almaty, Kazakhstan; amirtayevaainur@gmail.com). She received her Bachelor's degree in Information Security Systems from Al-Farabi Kazakh National University. Her academic and professional interests include cybersecurity, artificial intelligence, machine learning, deep learning, neural networks, parallel and distributed programming, phishing detection, computer vision, network traffic analysis, intrusion detection systems, and the application of AI in healthcare. Her research focuses on the development of intelligent information security systems, the application of neural networks, and the use of parallel computing techniques to improve machine learning algorithms.