Double Level Fuzzy Adversarial Clustering

While conventional Fuzzy C-means (FCM) clustering effectively learns fuzzy membership degrees (FMDs) indicating cluster belongingness probabilities, it faces critical limitations. Single Euclidean distance metric inadequately captures intrinsic sample structures. Moreover, non-fuzzy membership degrees (non-FMDs)---representing multi-label learning perspectives that quantify the likelihood of samples belonging to multiple clusters---have not been sufficiently leveraged to improve FMD estimation. To this end, this paper proposes a novel double level fuzzy adversarial (DLFA) clustering approach that exploits the interplay between FMD and non-FMD through a dual adversarial learning framework. At the fuzzy learning level, FMD and non-FMD establish an adversarial relationship by performing complementary fuzzy clustering operations. At the classifier learning level, they serve as outputs of distinct classifiers to guide adversarial training. This double level adversarial mechanism enforces mutual competition between the two learning streams, ultimately enhancing FCM's effectiveness. To capture richer sample structure, we introduce spectral rotation and scaling to integrate spectral information into FMD, enabling complementary structural representation. Furthermore, recognizing that non-FMD exhibits multi-peak characteristics akin to fuzzy multi-labels, we extend FCM by incorporating FMD results into the sample-prototype distance calculation. This strategy emphasizes samples with higher non-FMD values, thereby suppressing incorrect cluster assignments in FMD while amplifying correct ones, yielding more accurate membership estimation. Extensive experiments on 28 benchmark datasets demonstrate that DLFA achieves state-of-the-art clustering performance.