基于自适应粒子群算法的特征选择

Feature Selection Based on Adaptive Particle Swarm Optimization

在模式分类问题中,数据往往存在不相关或冗余的特征,从而影响分类的准确性。特征选择的提出,很好地解决了这一问题。特征选择的关键在于利用最少的特征获得最佳的分类效果。为了达到这一目的,一种基于自适应粒子群的特征选择的理论被提出。相比于原始的粒子群算法,在初始过程中引入混沌模型增加其初始粒子的多样性,在更新机制中引入自适应因子增加其全局搜索能力。同时将特征数目引入到适应度函数中,在迭代前期通过惩罚因子调节分类准确率和特征数目对于适应度函数的影响,在迭代中后期惩罚因子恒定,使特征数目对于适应度函数的影响趋于稳定。自适应粒子群算法具有很好的全局收敛性,能够避免陷入局部最优,尤其适合高维数据的降维问题。大量的理论分析和仿真实验的结果表明,与其他粒子群算法(PSO)的特征选择结果相比,在数据特征数目各异的情况下,该算法具有更好的分类效果,同时表明了所提算法的可行性以及优越性。

In pattern classification problems,there is often irrelevant or redundant features in data,thus affecting the accuracy of the classi- fication. Feature selection is proposed to be a good solution to this problem. The key of feature selection is to use the least feature for the best classification results. In order to achieve this object, a theory based on adaptive particle swarm feature selection is presented. Com- pared to the original particle swarm optimization, chaos model is introduced in the initial process of increasing its diversity of primary par- ticles, the introduction of adaptive factor to increase its global search capability in the update mechanism. At the same time the number of features will be introduced to the fitness function, in the early iterations adjustment classification accuracy and the number of features by penalizing factor for adapting to the impacts of the function, in the latter part of the penalty factor constant iteration, bringing the number of features of the fitness function tends to affect stable. Adaptive particle swarm algorithm has good global convergence and can avoid falling into local optimum,especially for lower-dimensional problem of high dimensional data. A large number of theoretical analysis and simulation results show that compared with other PSO feature of the election results,in the case where the number of different data charac- teristics, this algorithm has better classification results. Also it shows that the proposed algorithm is feasible and superior.