Advanced feature selection approach with Halton-based enhanced adaptive dung beetle algorithm

Abstract The dung beetle optimizer (DBO) is a simple structure with minimal hyperparameters and a population-based optimization algorithm that mimics the foraging behaviors of dung beetles. However, DBO has several limitations, including slow convergence and local optimum susceptibility, especially with multimodal or combinatorial functions. This article presents the adaptive dung beetle algorithm (AQDBO), which builds upon enhanced solution quality (ESQ) and a multi-strategy hybrid methodology. First, the Halton sequence is employed during the initialization generation, which results in a better population distribution. It also helps to reduce the chances that the AQDBO prematurely converges. Second, an adaptive convergence factor is proposed, prioritizing exploring in the early stages and local exploitation afterward. Third, an improved exploration strategy is suggested to boost the global search capacity of AQDBO, and finally, an ESQ strategy diversifies the optimal global solution to escape from suboptimal regions. The experiments tested the AQDBO over 51 benchmark functions from the CEC’17, CEC’20, and CEC’22. The results were compared with multiple optimization algorithms, and non-parametric tests were performed to verify the performance of the proposed algorithm. Additionally, we have created a binary version of the AQDBO for real applications to solve the feature selection problem in data classification. The binary AQDBO performs well over 15 datasets from the UCI repository with different degrees of complexity. Furthermore, the AQDBO achieved average accuracy in the range of 0.7108 to 0.9995 in the FS experiments, with the average number of selected features ranging from 1 to 688.07, covering datasets from low to high dimensionality.