【摘要】 特征选择是机器学习的重要分类任务，其结果直接影响后续学习算法性能。麻雀搜索算法(SSA)是近年来提出的一种基于麻雀智能行为的元启发式算法，它考虑麻雀的社会组织及其对环境的适应性求解连续优化问题，文中将SSA求解离散类的特征选择问题。首先采用混沌反向学习作为初始化策略，并结合差分进化算法改进更新解阶段，随后利用三种二进制方式将处理连续类问题的SSA转换为适应特征选择问题的二进制版本，并最终选择对分类性能提升最大的S型转移函数结合改进后的SSA,形成文中提出的一种包裹式的二进制麻雀特征选择算法BSFSA,为测试BSFSA的性能，使用k-最近邻作为分类器并选用21个UCI数据集，与7种先进的基于群体智能的包裹式特征选择算法和2种过滤式特征选择算法在分类精度和维度缩减率等方面进行比较，结果显示，BSFSA在18个数据集中取得最高分类精度，此外也取得了5个最高维度缩减率和5个次高维度缩减率。实验结果表明，BSFSA能够出色兼顾特征子集的分类精度与维度缩减能力，相较对比算法体现出一定优势。更多还原

【Abstract】 Feature selection is an important classification task in machine learning, and its results directly affect the performance of subsequent learning algorithms. The Sparrow Search Algorithm(SSA) is a meta-heuristic algorithm based on the intelligent behavior of sparrows proposed in recent years. It considers the social organization of sparrows and its adaptability to the environment to solve the continuous optimization problem. In this paper, SSA is used to solve the discrete feature selection problem. First, chaotic oppsite learning is used as the initialization strategy and combined with the differential evolution algorithm to improve the update solution stage, and then three binary methods are used to convert the SSA that deals with the continuous class problem into a binary version adapted to the feature selection problem, and finally choose sigmod transfer function to improve the classification performance. In order to test the performance of BSFSA, k-nearest neighbors are used as classifiers and 21 feature selection fields are selected. The UCI dataset is compared with seven advanced swarm intelligence-based wrapping feature selection algorithms and two filtering feature selection algorithms in terms of classification accuracy and dimension reduction rate. The comparison results show that BSFSA achieves the highest performance in 18 datasets In addition, the five highest dimensional reduction rates and the five second highest dimensional reduction rates have been obtained. The experimental results show that BSFSA can excellently take into account the classification accuracy and dimensional reduction ability of feature subsets. Compared with the nine comparison algorithms, it shows certain advantages.更多还原