Buffer-based adaptive fuzzy classifier

In the age of a technological revolution, heterogeneous sources are generating streams of data at a high rate, and an online classification of these data can facilitate data mining and analysis. Among the available classifiers, fuzzy-system-based (FSB) classifiers provide remarkable contributions due to their antecedent-consequent rule base structure. The Mamdani and Takagi-Sugeno type structure always uses the identical antecedent portion with fuzzy sets, which are themselves specified by parameterized membership functions driven by logical AND/OR operations. These membership functions are discerned either by experts or from data. However, for online or stream data, using a predefined membership function is not ideal. Meanwhile, a data-cloud has the ability to adopt changes in stream data, which share the same properties as those of a cluster but does not have any predefined shapes or a particular radius; rather, data-cloud offer a more objective representation of real-time data. Moreover, most algorithms with FSB classifiers avoid the presence of temporarily irrelevant data points or data-clouds that can be relevant in the future. In this paper, we develop a novel data-cloud-based classification algorithm for stream data classification called buffer-based adaptive fuzzy classifier (BAFC). The offline training stage of this algorithm can identify data-cloud from a static dataset to construct the AnYa type fuzzy rule. This algorithm is also able to cope with the dynamic nature of stream data. At the online or one-pass training stage, BAFC updates its rule base by creating and merging data-cloud based on its potential area. This algorithm also introduces a recursive formula for calculating data-cloud density with a buffer that is used for storing temporarily irrelevant data clouds. BAFC also uses the online pruning system of data-clouds to address storage problems. This approach can solve the issues associated with the parameterization and redundant rule base for other types of stream data (e.g., sensor data, bank transaction, intruder detection, images and videos, and, stock market and disease prediction) classification algorithms. This two-stage algorithm is evaluated on several benchmark datasets, and the results prove its superiority over different well-established classifiers in terms of classification accuracy (90.82% for 6 datasets and 97.13% for the MNIST dataset), memory efficiency (twice higher than other classifiers), and efficiency in addressing high-dimensional problems.