Abstract: Aiming at the problem of classification accuracy fluctuation caused by input layer weight and threshold random selection of Multi-output Extreme Learning Machine（MELM） classification model, a multi-classification model of extreme learning machine based on improved Flower Pollination Algorithm（CS-ACFPA） is proposed（CS-ACFPA-MELM）. Firstly, the adaptive strategy and Tent strategy are used to optimize the optimization method of Flower Pollination Algorithm（FPA）. Then a cost-sensitive fitness function is constructed to make the FPA better match the output of the MELM model. Finally, the improved FPA and the cost-sensitive fitness function are used to optimize the input weight and threshold of the extreme learning machine to improve the classification performance of the MELM model. In the contrast experiment, the effectiveness of the CS-ACFPA algorithm for the improvement of the MELM model is verified, and the advantages of the CS-ACFPA-MELM model on large-scale samples and the applicability of small samples are demonstrated.

Key words: classification model, extreme learning machine, pollen algorithm, cost sensitive, chaotic search

摘要： 针对多输出极限学习机（MELM）分类模型输入层权值和阈值随机选取导致的分类精度波动问题，提出一种基于改进花粉算法（CS-ACFPA）的极限学习机多分类模型（CS-ACFPA-MELM）。利用自适应算子和Tent策略优化花粉算法的寻优方式，构造一种基于代价敏感的适应度函数，使花粉算法能够更好地匹配MELM模型的输出，最后使用改进的花粉算法和基于代价敏感的适应度函数优化极限学习机的输入权值和阈值，以提高MELM模型的的分类性能。通过对比实验验证了CS-ACFPA算法对MELM模型改进的有效性，并且体现了CS-ACFPA-MELM模型在大规模样本上的优势以及小样本上的适用性。

关键词: 分类模型, 极限学习机, 花粉算法, 代价敏感, 混沌搜索