Deep Clustering and Dynamic Routing Based TSK Fuzzy System for Classification

Takagi-Sugeno-Kang (TSK) fuzzy system is a classical machine learning method with flexible modelling and good nonlinear approximation. However, as the data size and feature dimension increase, it has continually challenged the TSK fuzzy system to achieve satisfactory performance. Therefore, this paper proposes a TSK fuzzy system that combines deep clustering and the dynamic routing algorithm (DDTSK). Firstly, to accurately divide the complex input data space, a deep clustering based on sparse auto-encoder and K-means is proposed for determining the antecedent parameters. Then, a consequent parameters fine-tuning method based on dynamic routing is proposed to establish the correlation between input and output space. To further improve the classification performance, updating the output weights of the TSK fuzzy classifier via ridge regression algorithm. The experiments are conducted on 10 UCI machine learning datasets having different data sizes and feature dimensions. The experimental results show that the proposed optimization technique can further improve the accuracy and generalization of the TSK fuzzy system.