Anomaly Detection in Time Series Data Using Generative Modeling
Project Overview
Detecting anomalies in tabular time-series data is challenging because each dataset can have different patterns and behaviors. A single model often struggles to perform well across different types of data. To address this, this study proposes a flexible ensemble-based framework that combines the strengths of several models to improve anomaly detection.
Our approach uses meta-learning to guide the selection of the best combination of models for a given dataset. We extract meta-features—summaries of the dataset’s structure and characteristics—to help the system decide which models to use and how much to rely on each one.
The model pool includes a variety of techniques from machine learning, deep learning, and generative models. For example, Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs) are included with hybrid designs to better capture both high-dimensional features and time-based patterns.
To combine the outputs of the different models, we use a method called stacking, which learns how to weight the predictions based on past performance. This allows the framework to adapt better than traditional averaging or voting methods.
Overall, this project aims to build a robust and adaptable system for detecting anomalies in time-series data across many domains, helping to improve accuracy and reduce the need for manual tuning.
Project Team
Dr. Priyanga Dilini Talagala, Department of Computational Mathematics, University of Moratuwa
Ms.Theepana Govintharajah, Faculty of Information Technology, University of Moratuwa
Mr. Gowsigan Kanagalingam, Faculty of Information Technology, University of Moratuwa
Mr. Pavadaran Pathmaranjan, Faculty of Information Technology, University of Moratuwa
Outputs
Publicaitons
- Govintharajah, T., Pathmaranjan, P., Kanagalingam, G., & Talagala, P. D. (2024, December). Generalized Meta Framework for Forecasting. In 2024 9th International Conference on Information Technology Research (ICITR) (pp. 1-6). IEEE. DOI: 10.1109/ICITR64794.2024.10857715
Presentations
- Generalized Meta Framework for Forecasting
Date: December 5th to 6th, 2024
Event: 9th International Conference on Information Technology Research (ICITR)
Location: University of Moratuwa, Sri Lanka
Awards
- Best Paper Award, Data Science and Data Driven Applications Track, 9th International Conference on Information Technology Research (ICITR) 2024, Faculty of Information Technology, University of Moratuwa
