Hierarchical fuzzy regression tree: A new gradient boosting approach to design a TSK fuzzy model

This paper proposes a novel gradient-boosting-based ensemble system with a fuzzy regression tree (FRT) as its base component for regression tasks. FRT first initializes the rule space as a whole. In each division round, the rule space is supervised to be divided into smaller subspaces using a novel guaranteed membership division method. The consequent parameters are optimized using fuzzy weighted regularized L2 least squares. The FRT corresponding to the minimum MSE is kept as the initial model for the next division round. The above operation is repeated until the FRT reaches the maximum division rounds. Using gradient boosting to train subsequent FRTs to build hierarchical fuzzy regression tree (HFRT). The HFRT can guarantee the high accuracy while ensuring the low complexity of computation, particularly suitable for solving high-dimensional problems. HFRT performed ablation analysis experiments and compared with state-of-the-art algorithms on 28 data sets. The results confirm the effectiveness of the HFRT.