Multi-objective genetic programming for binary classification with adaptive thresholds and a generalization-optimizing fitness function

Abstract

Genetic programming (GP) has been widely applied to classifier construction due to its flexible representation and powerful feature construction capabilities. Existing studies have proposed various fitness functions to improve GP-based classifiers, but most of them rely on a fixed decision threshold. However, when dealing with imbalanced classification problems, a fixed threshold often biases the model toward the majority class, thereby compromising overall performance. To address this issue, in this paper, we propose a novel multi-objective GP framework for constructing binary classifiers with adaptive threshold adjustment. During evolution, the method employs Youden’s Index to dynamically adjust the threshold of each individual, enabling the classifiers to better fit the underlying data distribution. In addition, we introduce a new class separation metric, dist_t, to quantify the clarity of class boundaries and enhance the generalization ability of the evolved models. The framework jointly optimizes three objectives: minority class accuracy, majority class accuracy, and the proposed dist_t metric. Experiments on 14 imbalanced datasets demonstrate that our method significantly outperforms conventional single-objective GP with fixed thresholds. Further results also confirm the positive impact of the proposed dist_t metric on classification performance. Compared to seven existing GP algorithms and five traditional machine learning classifiers, our approach achieves superior overall performance and better generalization ability.