Adaptive grid-meshed-buffer clustering algorithm for outlier detection in evolving data stream

• 第一著者: Abdulateef, Alaa Fareed
• フォーマット: 学位論文
• 言語: 英語; 英語; 英語
• 出版事項: 2023
• 主題: QA Mathematics
• オンライン･アクセス: https://etd.uum.edu.my/11343/1/permission%20to%20deposit-not%20allow-s902134_0001.pdf; https://etd.uum.edu.my/11343/2/s902134_01.pdf; https://etd.uum.edu.my/11343/3/s902134_02.pdf; https://etd.uum.edu.my/11343/

As the number of connected devices rises, real-time processing of data streams has garnered significant attention and interest within the scientific community. Clustering, known for its versatility in real-time data stream processing and independence from labeled instances, is a suitable method for analyzing evolving data streams. Existing clustering algorithms for outlier detection encounter significant challenges due to insufficient data pre-processing methods and the absence of a suitable data summarization framework for effective data stream clustering. This research introduces Adaptive Grid-Meshed-Buffer Stream Clustering Algorithm (AGMB), that addresses these weaknesses and improves outlier detection. The AGMB algorithm is built upon three algorithms: 1) Grid-Multi-Buffer Stream Clustering (GMBSC), 2) Cautious Grid-Multi-Buffer Stream Clustering (C-GMBSC) and 3) Adaptive-Grid-Multi-Buffer Stream Clustering (A-GMBSC). The GMBSC addresses inadequate data pre-processing, grid projection, and buffering issues while the C-GBMSC includes an outlier elimination strategy to cautiously eliminate detected outliers’ data points before it turns into normal data. The A-GMBSC is designed by adding an adaptive density threshold which maintains a cluster model of normal data. The effectiveness of the final outcome, which is the AGMB, is validated through an experimental evaluation conducted on eight datasets that are synthetic and real-life datasets. Evaluation was made based on various evaluation metrics related to outlier detection and clustering quality. The results indicate that the AGMB algorithm outperformed existing benchmark algorithms in terms of predefined evaluation criteria with an overall 72% accuracy compared to benchmark algorithms which is 11 % only. Hence, the empirical evidence highlights the superiority and practical relevance of the proposed algorithm in tackling outlier detection in evolving data streams. This may be useful for real-world applications such as surveillance systems based on IoT and customer behaviour analytics systems.