Abstract: In recent years, deep TSK (Takagi-Sugeno-Kang) fuzzy classifiers based on stacked ensemble structures have become a key focus in the research of TSK fuzzy classifiers. Compared with traditional single TSK fuzzy classifiers, deep TSK fuzzy classifiers not only exhibit enhanced generalization capability but also offer better interpretability. However, when applying deep TSK fuzzy classifiers to epileptic electroencephalogram（EEG）signal detection, the following two challenges should be addressed: (1) How to improve the existing deep structures to ensure accurate detection of epileptic EEG signals while accelerating model construction and enhancing model interpretability (e.g., the less fuzzy rules and providing two types of interpretabilities); (2) How to leverage human cognitive behavior to further enhance the generalization capability of deep TSK fuzzy classifiers. To address the above two challenges, an adversarial hybrid TSK fuzzy classifier (AH-TSK) for epileptic EEG signal detection is proposed, which primarily consists of two aspects: For the challenge (1), based on the existing deep stacked ensemble structures, by introducing the wide ensemble structure, a novel deep-and-wide-based hybrid ensemble structure is proposed to integrate a linear sub-model called sparse representation-based linear classifier (i.e., SRLc) and several nonlinear sub-models called adversarial TSK fuzzy sub-model (i.e., A-TSK); For the challenge (2), inspired by two cognitive behaviors “from global coarse to local refinement” and “knowledge abandonment”, a new adversarial training method is proposed. It first trains a linear model (SRLc) on all original samples of EEG dataset to classify nonlinear distributed samples. Then, an adversarial training method by introducing “knowledge abandonment” is proposed to train several A-TSKs parallelly on the obtained nonlinear distributed samples of the EEG dataset. Finally, the final output of AH-TSK can be obtained by adopting the proposed nearest-labelled strategy on the outputs of an SRLc and all A-TSKs. Experimental results on the EEG datasets confirm that compared with the comparative methods, AH-TSK exhibits enhanced generalization capability, faster running speed, and better interpretability. Moreover, AH-TSK provides more types of interpretabilities, i.e., linguistic and attribute-significance-based interpretability.

Key words: hybrid Takagi-Sugeno-Kang (TSK) fuzzy classifier, epileptic electroencephalogram (EEG) signals detection, adversarial training method, attribute-significance-based interpretability, linguistic interpretability

摘要： 近年来，基于栈式集成结构的深度TSK（Takagi-Sugeno-Kang）模糊分类器已成为TSK模糊分类器研究热点之一，与传统单一的TSK模糊分类器相比，深度TSK模糊分类器不仅具有增强的泛化能力，而且具有较好的可解释性。然而，当深度TSK模糊分类器应用于癫痫脑电图（EEG）信号检测时，需要解决如下两个挑战：（1）如何改进现有的深度结构，在保证癫痫EEG信号检测精度的基础上，加快模型的构建速度并同时提高模型的可解释性（更少的规则数和提供两种类型的可解释性）；（2）如何利用人类认知行为，进一步提升深度TSK模糊分类器的泛化能力。为了解决上述两个挑战，提出面向癫痫EEG信号检测的对抗混合TSK模糊分类器（AH-TSK）。针对挑战（1），在现有深度栈式集成结构的基础上，引入宽度集成结构，从而提出一种新型的基于深度和宽度的混合集成结构，集成单个线性子模型（SRLc）和多个非线性子模型（A-TSK）；针对挑战（2），基于“从全局粗略到局部精细化”和“知识遗弃”这两种认知行为，提出了一种新的对抗训练方法。该方法先在EEG数据集的所有原始样本上训练线性模型（SRLc），以分类非线性分布的样本；在得到的非线性部分上，引入“知识遗弃”的对抗策略，并行训练多个A-TSK；通过使用最近标签策略，对SRLc和所有A-TSK的输出进行集成得到最终输出。实验结果表明，与对比方法相比，AH-TSK具有增强的泛化能力、较快的运行速度以及较好的可解释性，此外能够提供更多类型的可解释性（语义和基于特征重要性的可解释性）。

关键词: 混合TSK模糊分类器, 癫痫脑电图（EEG）信号检测, 对抗训练方法, 基于特征重要性的可解释性, 语义可解释性