Anomaly Detection in Microservices Using Ensemble Methods

Salvado, Nuno Miguel da Silva

http://hdl.handle.net/10400.6/19574

Use this identifier to reference this record.

Name:	Description:	Size:	Format:
11513_29166.pdf		3.62 MB	Adobe PDF	Download

Send Feedback

Authors

Salvado, Nuno Miguel da Silva

Advisor(s)

Freire, Mário Marques

Abstract(s)

Anomaly detection represents a critical factor in ensuring the reliability and resilience of microservice-based systems, where failures can rapidly propagate and compromise overall service availability. This dissertation investigates the application of classical Machine Learning (ML) algorithms and ensemble methods for anomaly detection in microservices, using the TraceRCA dataset as a representative benchmark. The work begins with a systematic literature review, which categorizes traditional and ML-based approaches to anomaly detection, identifying key research gaps and datasets. Building on this foundation, a complete experimental pipeline was developed, including preprocessing, feature engineering, and anomaly labeling, followed by the evaluation of multiple baseline classifiers such as Logistic Regression (LogReg), Support Vector Machine (SVM), Decision Tree (DT), K-Nearest Neighbors (KNN), Multilayer Perceptron (MLP), and Gaussian Naïve Bayes (GNB). To enhance predictive performance, ensemble techniques including Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Histogram-based Gradient Boosting (HGBM) were implemented and compared against baselines. The evaluation considered both predictive accuracy and resource efficiency, measuring metrics such as F1-score, precision, recall, accuracy, Receiver Operating Characteristic – Area Under the Curve (ROC-AUC), as well as execution time, Random Access Memory (RAM) consumption, and Central Processing Unit (CPU) utilization. The experimental results demonstrate that ensemble models consistently outperform baselines, with boosting-based methods (XGBoost, LightGBM, HGBM) achieving the highest predictive performance, while RF offered stable results with moderate computational overhead. These findings highlight the trade-offs between accuracy and efficiency, underlining the importance of careful algorithm selection according to deployment constraints. This research contributes by providing a comprehensive benchmark of ML and ensemble methods for anomaly detection in microservices, bridging the gap between predictive performance and practical applicability in real-world environments.