基于黑寡妇算法的特征选择方法研究

特征选择通过去除无关和冗余特征提高学习算法性能，本质是组合优化问题。黑寡妇算法是模拟黑寡妇蜘蛛生命周期的元启发式算法，在收敛速度、适应度值优化等方面具有诸多优势。针对黑寡妇算法不能进行特征选择的问题，设计五种优化策略：二进制策略、“或门”策略、种群限制策略、快速生殖策略以及适应度优先策略，提出黑寡妇特征选择算法（black widow optimization feature selection algorithm，BWOFS）和生殖调控黑寡妇特征选择算法（procreation controlled black widow optimization feature selection algorithm，PCBWOFS），从特征空间中搜索有效特征子集。在多个分类、回归公共数据集上验证新方法，实验结果表明，相较其他对比方法（全集、AMB、SFS、SFFS、FSFOA），BWOFS和PCBWOFS能找到预测精度更高的特征子集，可提供有竞争力、有前景的结果，而且与BWOFS相比，PCBWOFS计算量更小，性能更好。

Research on Feature Selection Method Based on Black Widow Optimization Algorithm

Feature selection is essentially a combinatorial optimization problem to improve the performance of learning algorithms by removing irrelevant and redundant features. Black widow optimization algorithm（BWO） is a new meta-heuristic algorithm of simulating the life cycle of black widow spiders. This algorithm has many advantages in terms of convergence speed and fitness value optimization. To counter the problem that black widow optimization algorithm cannot select feature, five kinds of optimization strategies are designed, including binary strategy, “or” strategy, population limit strategy, rapid reproduction strategy and fitness priority strategy, black widow optimization feature selection algorithm（BWOFS） and procreation controlled black widow optimization feature selection algorithm（PCBWOFS） are proposed to search effective feature subsets from feature space. New methods is validated in multiple classified and regression common datasets. The experimental results show that compared with other comparison methods（full set, AMB, SFS, SFFS and FSFOA）, BWOFS and PCBWOFS can find feature subsets with higher prediction accuracy, and can provide competitive and promising results. Moreover, compared with BWOFS, PCBWOFS has smaller computation and better performance.