Privacy-Preserving Construction of Ellipsoidal Granular Descriptors Based on Horizontal Federated Gustafson–Kessel Algorithm

Abstract

This study is concerned with a realization of horizontal federated Gustafson–Kessel clustering algorithm and the ensuing construction of ellipsoidal information granules. As a fundamental component of Granular Computing, information granules play an important role in human-centric computing, such as human cognition and decision-making. Driven by the concerns of data privacy and confidentiality, it is of interest to investigate how to construct information granules on the basis of horizontally partitioned numeric data distributed across different sites using a privacy-preserving approach. To meet this challenge, federated learning has become an appealing solution to the problem of forming meaningful clusters (information granules) while ensuring data privacy and confidentiality. A two-development strategy is applied in the proposed algorithm: first, a collection of numeric representatives (prototypes) is obtained with the use of federated Gustafson–Kessel algorithm, which is able to reveal ellipsoidal shapes in the datasets and second, information granules are built through engaging the principle of justifiable granularity. A series of experimental studies demonstrate the effectiveness of the proposed federated Gustafson-Kessel algorithm in revealing the structure of the entire dataset. The formed ellipsoidal information granules help us gain a better insight into the topology of the overall dataset.