A Hypersphere Information Granule-Based Fuzzy Classifier Embedded With Fuzzy Cognitive Maps for Classification of Imbalanced Data

In this article, a hypersphere information granule-based fuzzy classifier integrated with Fuzzy Cognitive Maps (FCM), named FCM-IGFC, is proposed for the classification of imbalanced data. The proposed FCM-IGFC is structured by sequentially linking a Takagi-Sugeno-Kang (TSK) fuzzy system (FCM-TSK) – which is embedded with FCMs – and a rule-based fuzzy system (IGFS) based on hypersphere information granules. The FCM-TSK leverages the inference capabilities of the FCM to allow for the movement of data within the original space, while the IG-FS creates a mapping between samples and majority and minority classes using information granules. In this study, we introduced an innovative bottom-up granulation method and an overlap elimination technique for constructing hypersphere information granules. These methods facilitate the creation of information granules that accurately represent the structure of classes, even when dealing with im- balanced data. Moreover, the stacked structure of the FCM-IGFC offers data transfer capabilities. This helps reduce the complexity of distributions such as small, disjointed clusters and irregular class boundaries, with the support of the FCM, thereby making it easier to use information granules to describe the class structure. A series of experiments conducted on 12 publicly available datasets demonstrated that the performance of FCM-IGFC significantly surpasses that of existing granule-based fuzzy classifiers. Additionally, it is competitive with top-tier classifiers that incorporate sampling methods.

I. Introduction

Classification of imbalanced data is a common occurrence in real-world applications, including disease diagnosis [1], [2], nature disaster prediction [3], material strength prediction [4] and financial crisis prediction [5]. In an imbalanced dataset, the quantity of samples that represent the minority class is significantly fewer than those representing the majority class. Despite this disparity, the minority class often carries substantial importance in numerous scenarios. However, due to the imbalance in prior probabilities, the minority class tends to be overlooked in studies concerning imbalanced data. This issue is particularly noticeable with traditional rule-based fuzzy classifiers, which are originally designed with the presumption of a balanced class distribution. These classifiers thus encounter considerable challenges when tasked with accurately classifying imbalanced data.