Fuzzy Ordered $c$-Means Clustering and Least Angle Regression for Fuzzy Rule-Based Classifier: Study for Imbalanced Data

Abstract

This article introduces a new classifier design method that is based on a modification of the traditional fuzzy clustering. First, a new fuzzy ordered c-means clustering is proposed. This method can be considered as a generalization of the concept of the conditional fuzzy clustering by introducing ordering and weighting distances from data to cluster prototypes. As a result, a more local impact of data on created groups and increased repulsive force between group prototypes are obtained. The proposed method provides a better representation of the data classes, in particular for classes with small cardinality in the training set (imbalanced data). A special initialization of the prototypes is also introduced. Next, the proposed clustering method is used to construct the premises of if-then rules of a fuzzy classifier. The conclusions of the rules are obtained by the least angle regression algorithm, which selects only those rules, that maximize the generalization ability of a classifier. Each if-then rule is represented in easily interpretable Mamdani-Assilian form. Finally, an extensive experimental analysis on 89 benchmark balanced and imbalanced datasets is performed to demonstrate the validity of the introduced classifier. Its competitiveness to state-of-the-art classifiers, with respect to both performance and interpretability, is shown as well.

A Method for Analyzing the Performance Impact of Imbalanced Binary Data on Machine Learning Models

“…This study employed non-parametric testing to analyze and compare whether there are significant differences in the performance stability of machine learning models on imbalanced datasets. These tests have been used in several empirical studies and are highly recommended in the field of machine learning and data mining [20,30] to confirm experimental results. The non-parametric test procedure consists of three steps.…”

“…Table 2 summarizes the 48 imbalanced datasets with different IR, including the total number of instances, numbers of features, class names and number of instances belonging to the minority. To obtain multiple binary imbalanced data, we refer to the method in similar imbalanced data classification studies [18][19][20], and transform multiclass imbalanced data in the UCI and Data Hub into binary imbalanced data by combining one or more classes. As shown in Table 2, although some imbalanced datasets are the same imbalanced data, with different versions, they are different.…”