基于SDA与SVR混合模型的迁移学习预测算法

摘要

为了解决大数据时代下小样本数据预测精度不高的问题,提出一种基于堆栈降噪自编码(SDA)与支持向量回归机(SVR)的混合模型.该方法采用源域大样本数据对堆栈降噪自编码和支持向量回归机混合模型进行迁移预训练,再利用目标域小样本数据微调混合模型.堆栈降噪自编码器具有良好的通用深层特征自主抽取能力,能够发掘源领域与目标领域相似任务间的共有特征知识,该知识能够辅助支持向量回归机在高维噪声小样本数据集上的预测.在多种数据集上的实验结果证明了该方法的有效性.%To improve the prediction accuracy of small sample in the era of big data,this paper introduced a novel hybrid model based on stacked denoising auto-encoder(SDA) and support vector regression (SVR).The hybrid model is pretrained by using a large number of source domain data,and then it is fine-tuned by a small amount of target domain data.The method takes the advantage of SDA,extracting common features autonomously on related but different target domain data.By transferring these prior knowledge,the hybrid model can provide a relatively accurate prediction result on high-dimensional and noisy small sample.Experimental results on extensive datasets demonstrate the effectiveness of the proposed model.