MCSRc : a type-0 fuzzy classifier to handle concept changes in data streams
Chen Wenhao
- xiii, 113 pages : illustrations (some color) ; 30 cm. + 1 CD-ROM
Faculty of Computing
Thesis (Master of Science) -- Universiti Malaysia Pahang – 2023
Includes bibliographical references
Classification problems are an integral part of data mining and artificial intelligence. With the growing demand for real-time classification tasks, various classifiers are turning to online modeling to process data streams. Data streams are always unpredictable and unstable, so as to handling concept change becomes an inevitable research challenge, because the problem of learning new concepts from unstable data streams directly determines whether a classifier has good capabilities of processing data streams. Among various classifiers, fuzzy rule-based systems have been considered as a good candidate due to their excellent interpretability and mathematical performance. Especially, as the third alternative structure to the traditional fuzzy rule-based systems (TS-Type, M-Type), the type-0 fuzzy rule-based system was developed for high dimensional and more complex problems. However, the development of the type-0 fuzzy rule-based system is slow because its highly integrated technology makes its own development more difficult. Meanwhile, the rapid development of data stream processing technology makes the type-0 fuzzy rule-based system face unprecedented pressure. This leads to the need to pay more attention to the problem of handling CC in data stream classification when developing type-0 fuzzy rule-based systems. To face the above problem, this research proposes a new type-0 fuzzy rule-based classifiers, namely Multi-Clouds-Single-Rule type-0 fuzzy classifier (MCSRc). The proposed MCSRc can handle CC better due to it integrates two techniques: Automatic Data Partitioning (ADP) technique and sliding window-based restriction strategy. ADP supports MCSRc to build new data clouds more naturally, and the sliding window-based limiting strategy supports MCSRc to retain more current data clouds and merge some old unstable data clouds. In the systematic evaluation, two types of data streams are utilized for revealing MCSRc’s overall and real-time performance. As the numerical evaluation results supported, the generalization performance of the proposed classifiers outperforms the performance of three existing advanced fuzzy classifiers. In sum, type-0 fuzzy classifiers’ overall performance is improved by about 17.77%, also means that the proposed MCSRc has higher stability, accuracy, and less dependence on the number of data clouds. Most importantly, MCSRc’s growing potentiality is improved by about 274.43%. This means that the proposed MCSRc has an excellent ability to stabilize itself and prevent performance degradation when dealing with various unstable data streams. It is worth noting that the learning process of the two proposed classifiers is in a one-pass way and learns from scratch without any offline part or chunk-by-chunk learning strategy.