基于流形正则化的多元时间序列半监督回归

Manifold regularization based semi-supervised regression on multivariate time series

针对多元时间序列半监督回归只考虑样本间空间关系信息而忽略了样本间时域信息的问题,提出了一种考虑样本间时域信息的半监督回归算法(ST-LapRLSR).在时域光滑性假设下,构造了一种能更好地反映样本间内蕴几何结构的正则化项.在建立图拉普拉斯的过程中,将样本点间的时序关系引入到边的权重计算中,并在流形正则化框架下利用该正则化项进行半监督回归,最后通过表示定理进行求解.在公共数据集和煤矿多传感器数据上进行了实验,结果表明:在多元时间序列半监督回归中,与只考虑样本空间关系信息的最小二乘正则化算法(LapRLSR)相比,ST-LapRLSR能同时利用样本的时空信息,准确率得到了提高.

Traditional semi-supervised regression on multivariate time series only takes account of the spatial information of samples,and the temporal information is always neglected.To solve the problem,a semi-supervised regression algorithm ST-LapRLSR is proposed which takes account of the temporal information of samples.For time series an assumption of temporal smooth is proposed,and based on this assumption,a regularization item that could reflect more underline information of samples is constructed.During constructing graph Laplacian,the temporal relation of samples is used in computing edge weight.Semi-supervised regression under manifold regularization framework using the proposed regularization item is carried out,and solved by the Representer theorem.The experiments take on public dataset and multivariate sensor time series data of mine,and the results show that,in semi-supervised regression on multivariate time series,ST-LapRLSR which uses temporal and spatial information of samples simultaneously achieves better accuracy contrast to LapRLSR which only considers the spatial information of samples.