Improved Sequential Fuzzy Indexed Search Trees for Fast Classification

Machine learning has been a significant part of computer science in the last few decades. The application of fuzzy inference systems has become more and more prevalent as well, as they can not only use expert knowledge conveniently, but also handle uncertainty well, which has made them very useful in various fields of healthcare and medicine. In previous works, the authors have proposed various new classifiers (so-called Sequential Fuzzy Indexing Tables and Sequential Fuzzy Indexed Search Trees) that can realize a fast fuzzy inference by using indexing tables that store pre-calculated fuzzy values, with the drawback of a trade-off in memory space. In this paper, an improved version of the latter classifier is presented along with a new training algorithm in order to achieve fast, fuzzy inference-based classification that can provide roughly the same level of classification performance as the basic classifier and its training algorithm, while greatly reducing its memory cost. The idea behind the new training algorithm is that instead of focusing on a single training sample at the time, the values of all samples are regarded in for one attribute at a time, and by grouping their values, the fuzzy sets that make up the fuzzy membership functions can be built. The performance of the proposed method is evaluated on benchmark datasets.