Feature selection is a key pre-processing technique for classification which aims at removing irrelevant or redundant features from a given dataset. Generally speaking, feature selection can be considered as a multi-objective optimization problem, i.e, removing number of features and improving the classification accuracy. Genetic algorithms (GAs) have been widely used for feature selection problems. The crossover operator, as an important technique to search for new solutions in GAs, has a strong impact on the final optimization results. However, many crossover operators are problem-dependent and have different search abilities. Thus, it is a challenge to select the most efficient one to solve different feature selection problems, especially when the nature of feature selection problems is unknown in advance. In order to overcome this challenge, in this paper, a multi-objective binary genetic algorithm integrating an adaptive operator selection mechanism (MOBGA-AOS) is proposed. In MOBGA-AOS, five crossover operators with different search characteristics are used. Each of them is assigned a probability based on the performance in the evolution process. In different phases of evolution, the proper crossover operator is selected by roulette wheel selection according to the probabilities to produce new solutions for the next generation. The proposed algorithm is compared with five well-known evolutionary multi-objective algorithms on ten datasets. The experimental results reveal that MOBGA-AOS is capable of removing a large amount of features while ensuring a small classification error. Moreover, it obtains prominent advantages on large-scale datasets, which demonstrates that MOBGA-AOS is competent to solve high-dimensional feature selection problems.

特征选择是分类的关键预处理技术，旨在从给定数据集中去除不相关或冗余的特征。一般来说，特征选择可以看作是一个多目标优化问题，即去除特征数量，提高分类精度。遗传算法 (GA) 已广泛用于特征选择问题。交叉算子作为在遗传算法中寻找新解的重要技术，对最终的优化结果有很大的影响。然而，许多交叉算子都依赖于问题并且具有不同的搜索能力。因此，选择最有效的方法来解决不同的特征选择问题是一项挑战，尤其是当特征选择问题的性质事先未知时。为了克服这一挑战，在本文中，提出了一种集成自适应算子选择机制的多目标二元遗传算法（MOBGA-AOS）。在 MOBGA-AOS 中，使用了五个具有不同搜索特性的交叉算子。根据进化过程中的表现，它们中的每一个都被分配了一个概率。在进化的不同阶段，轮盘选择根据概率来选择合适的交叉算子，为下一代产生新的解决方案。在十个数据集上将所提出的算法与五种著名的进化多目标算法进行了比较。实验结果表明，MOBGA-AOS 能够在保证较小分类误差的同时去除大量特征。此外，它在大规模数据集上获得了突出的优势，