Prototype-based manifold regularization approach with concept drift adaptation in data stream classification
Not available
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Doctoral thesis |
| Language: | English |
| Published: |
Universiti Teknologi Malaysia
2025
|
| Subjects: | |
| Online Access: | https://utmik.utm.my/handle/123456789/40331 |
| Abstract | Abstract here |
| _version_ | 1854975122249613312 |
|---|---|
| author | Muhammad Zafran Muhammad Zaly Shah |
| author2 | Anazida Zainal, supervisor |
| author_facet | Anazida Zainal, supervisor Muhammad Zafran Muhammad Zaly Shah |
| author_sort | Muhammad Zafran Muhammad Zaly Shah |
| description | Not available |
| format | Doctoral thesis |
| id | utm-123456789-40331 |
| institution | Universiti Teknologi Malaysia |
| language | English |
| publishDate | 2025 |
| publisher | Universiti Teknologi Malaysia |
| record_format | dspace |
| record_pdf | Abstract |
| spelling | utm-123456789-403312025-08-21T14:42:05Z Prototype-based manifold regularization approach with concept drift adaptation in data stream classification Muhammad Zafran Muhammad Zaly Shah Anazida Zainal, supervisor Computer engineering Not available Due to the large volume of data being generated continuously, various approaches in the field of data stream mining have been proposed to allow data to be processed in real-time. However, most of these are supervised learning approaches that are difficult to train when labelled data is scarce, which is often the case in data stream mining. Labelling each arriving data in a data stream is impractical due to the rapid arrival of data. In addition, due to the temporality nature of data streams, the model must be updated regularly to mitigate the threat of concept drifts that can harm the model’s prediction reliability. To address the labelled data scarcity issue, a manifold regularization approach (MRA) was proposed via the Semi-Supervised – Extreme Learning Machine (SS-ELM) which can be trained using a partially labelled dataset, making an assumption on the similarity between the labelled and unlabelled data. Unfortunately, the manifold regularization approach still falls short of being practical for data stream mining. It requires determining the Radial Basis Function (RBF) kernel width parameter that can only be determined via a hyperparameter search that requires many labelled data. The correct RBF kernel width parameter setting is important to describe the similarity between each data so the model can fully exploit the information from the labelled data to maximize the prediction performance. This study proposed combining the manifold regularization approach with a prototype-based approach to eliminate the requirement of a hyperparameter search to make the manifold regularization approach more practical for data stream mining. The prototype-based approach assisted the manifold regularization approach by describing the similarity between each data instead of relying on the RBF kernel. Based on the experiments, the proposed approach performed similarly to other comparable approaches without a hyperparameter search. Furthermore, the proposed approach was more practical and did not require a hyperparameter search or any additional labelled data. In addition, the combination of the manifold regularization and prototype-based approach enabled the detection and adaptation to concept drift when labelled data was scarce by allowing concept drifts caused by decision boundary shifts to be detected without the error rate approach. Many labelled data were required, and the model was further adapted with varying concept drift severity by identifying which underlying model in the ensemble was the most affected by the concept drift. Hence, the contribution of this study is two-fold. Firstly, it allows a manifold regularization approach to be deployed without having to search for a specific hyperparameter. Secondly, it detects and adapts to concept drifts without using an error rate approach that requires many labelled data. zulraizam UTM 217 p. Thesis (Doctor of Philosophy (Computer Science)) - Universiti Teknologi Malaysia, 2023 2025-03-06T09:51:27Z 2025-03-06T09:51:27Z 2023 Doctoral thesis https://utmik.utm.my/handle/123456789/40331 vital:152733 valet-20230813-09483 ENG Closed Access UTM Complete Unpublished Completion application/pdf Universiti Teknologi Malaysia |
| spellingShingle | Computer engineering Muhammad Zafran Muhammad Zaly Shah Prototype-based manifold regularization approach with concept drift adaptation in data stream classification |
| thesis_level | PhD |
| title | Prototype-based manifold regularization approach with concept drift adaptation in data stream classification |
| title_full | Prototype-based manifold regularization approach with concept drift adaptation in data stream classification |
| title_fullStr | Prototype-based manifold regularization approach with concept drift adaptation in data stream classification |
| title_full_unstemmed | Prototype-based manifold regularization approach with concept drift adaptation in data stream classification |
| title_short | Prototype-based manifold regularization approach with concept drift adaptation in data stream classification |
| title_sort | prototype based manifold regularization approach with concept drift adaptation in data stream classification |
| topic | Computer engineering |
| url | https://utmik.utm.my/handle/123456789/40331 |
| work_keys_str_mv | AT muhammadzafranmuhammadzalyshah prototypebasedmanifoldregularizationapproachwithconceptdriftadaptationindatastreamclassification |