Abstract: People tend to adopt a few representative features to describe a rule, ignoring very minor, redundant information. The classical interval type-2 TSK (Takagi-Sugeno-Kang) fuzzy system adopts full data feature space at the antecedent and consequent of each rule. For high-dimensional data, it is easy to increase the complexity and reduce the interpretability. This paper proposes the interval type-2 fuzzy subspace zero-order TSK system for this problem. It uses fuzzy subspace clustering and grid partition to generate sparse and structured rule centers at the antecedent, and simplifies the consequent to be zero-order. So an interval type-2 fuzzy system using semantic more concise rules can be built. The experimental results on synthetic datasets and real datasets show that the proposed method owns good classification performance with better interpretability.

Key words:  interval type-2 fuzzy system, TSK system, subspace, classification, interpretability

摘要： 人们倾向于使用少量的有代表性的特征来描述一条规则，而忽略极为次要的冗余的信息。经典的区间二型TSK（Takagi-Sugeno-Kang）模糊系统，在规则前件和后件部分会使用完整的数据特征空间，对于高维数据而言，易导致系统的复杂度增加和可解释性的损失。针对于此，提出了区间二型模糊子空间0阶TSK系统。在规则前件部分，使用模糊子空间聚类和网格划分相结合的方法生成稀疏的规整的规则中心，在规则后件部分，使用简化的0阶形式，从而得到规则语义更为简洁的区间二型模糊系统。在模拟和真实数据上的实验结果表明该方法分类效果良好，可解释性更好。

关键词: 区间二型模糊系统, TSK系统, 子空间, 分类, 可解释性