A Semantic-based Hoist Mutation Operator for Evolutionary Feature Construction in Regression [Hot off the Press]

Hengzhe Zhang, Qi Chen, Bing Xue, Wolfgang Banzhaf, Mengjie Zhang

Abstract: This Hot-off-the-Press paper summarizes our recently published work, "A Semantic-based Hoist Mutation Operator for Evolutionary Feature Construction in Regression" [9] published in IEEE Transactions on Evolutionary Computation. Our study introduces a semantic-based hoist mutation operator to control tree bloat and reduce tree sizes in genetic programming (GP) based evolutionary feature construction algorithms. The proposed operator identifies the most informative subtree with the largest cosine similarity to the target semantics and then hoists the subtree to the root as a new GP tree. This process reduces the tree sizes without compromising learning capability. The proposed operator is supported by the probably approximately correct (PAC) learning theory, ensuring that it does not degrade the generalization upper bound of GP models. By employing a hashing-based redundancy checking strategy, the proposed method outperforms seven bloat control methods on 98 datasets in reducing model size while maintaining test performance at the same level. These findings demonstrate the capability of using semantic hoist mutation for bloat control in GP-based evolutionary feature construction.