Abstract: Feature selection technology can effectively solve the curse of dimensionality problem, and many search strategies have been applied to feature selection problems. Aiming at the low search ability of harmony feature selection algorithm, an improved Harmony Search Feature Selection algorithm based on self-adaptation Global Pitch Adjusting（HSFS-GPA） is proposed to enhance the exploitation ability of solution space. The distance between feature sets is introduced into the feature selection problem. In the algorithm search process, a new harmony is generated by combining with the global information. The distance between the candidate harmony and the current optimal harmony or the worst harmony is changed. At the same time, competition selection mechanism is established to improve solution precision and enhance the ability of escaping local optima, the information of the worst harmony in harmony memory is updated at each iteration. HSFS-GPA is compared with the original harmony feature selection algorithm, particle swarm optimization algorithm and genetic algorithm. The size of the feature subset selected by HSFS-GPA is reduced by 15% than that of the original harmony search algorithm, and the average of the subset evaluation is increased to 0.98. The experimental result shows that HSFS-GPA can search for a better feature subset under the same condition.

Key words: harmony search, feature selection, meta heuristics, dimension reduction, big data

摘要： 特征选择技术能有效解决维数灾难问题，许多搜索策略已经被应用到特征选择问题中。针对和声特征选择算法搜索能力低下的问题，提出了一种基于全局自适应调距的和声特征选择算法（HSFS-GPA）。将特征集的距离定义引入到特征选择问题中，在算法搜索过程中结合全局信息对随机产生的新和声进行调整，以一定概率减小候选和声与当前最优和声的距离来加快算法搜索速度，或减少候选和声与最差和声的距离以避免陷入局部最优；同时，采用竞争选择方案随时更新和声库全局信息，改进和声库的更新机制提高算法搜索质量。将HSFS-GPA与原始和声特征选择算法、粒子群算法和遗传算法进行对比实验，HSFS-GPA所选特征子集的大小比原始和声算法减少15%，子集评价值平均提高到0.98。实验结果表明，HSFS-GPA能在相同的条件下搜索到更优质的特征子集。

关键词: 和声搜索, 特征选择, 启发式算法, 维数约减, 大数据