基于有序编码的核极限学习顺序回归模型

Ordered Code-based Kernel Extreme Learning Machine for Ordinal Regression

顺序回归是机器学习领域中介于分类和回归之间的有监督问题。在实际中,许多带有序关系标签的问题都可以被建模成顺序回归问题,因此顺序回归受到众多学者的关注。基于极限学习机(ELM)的算法能有效避免因迭代过程陷入的局部最优解,减少训练时间,但基于极限学习机的算法在顺序回归问题上的研究较少。该文将核极限学习机与纠错输出编码相结合,提出了一种基于有序编码的核极限学习顺序回归模型。该模型有效解决了如何在顺序回归中取得良好的特征映射以及如何避免传统极限学习机中隐层节点个数依赖于人工设置的问题。为验证提出模型的有效性,该文在多个顺序回归数据集上进行了测试,测试结果表明,相比于传统ELM模型,该文提出的模型在准确率上平均提升了10.8%,在数据集上预测表现最优,而且获得了最短的训练时间,从而验证了模型的有效性。

Ordinal regression is one of the supervised learning issues,which resides between classification and regression in machine learning fields.There exist many real problems in practice,which can be modeled as ordinal regression problems due to the ordering information between labels.Therefore ordinal regression has received increasing interest by many researchers recently.The Extreme Learning Machine（ELM）-based algorithms are easy to train without iterative algorithm and they can avoid the local optimal solution; meanwhile they reduce the training time compared with other learning algorithms.However,the ELM-based algorithms which are applied to ordinal regression have not been exploited much.This paper proposes a new ordered code-based kernel extreme learning ordinal regression machine to fill this gap,which combines the kernel ELM and error correcting output codes effectively.The model overcomes the problems of how to get high quality feature mappings in ordinal regression and how to avoid setting the number of hidden nodes by manual.To validate the effectiveness of this model,numerous experiments are conducted on a lot of datasets.The experimental results show that the model can improve the accuracy by 10.8% on average compared with traditional ELM-based algorithms and achieve the stateof-the-art performance with the least time.