当前位置：网站首页>SMA TE: Semi-Supervised Spatio-Temporal RepresentationLearning on Multivariate Time Series

SMA TE: Semi-Supervised Spatio-Temporal RepresentationLearning on Multivariate Time Series

2022-07-27 11:47:00 【Xiao Chen who wants money】

This is an article based on Tapnet Revised articles .

Innovation points ：

1、 Spatial information of time series is added （ That is to say variable Axis ）

2、 Semi supervised learning

3、 Yes embedding learning visualization

4、 And 13 Supervised learning and 4 Half supervised learning as baseline

（ Personally feel , Compare with Tapnet, The main difference lies in the addition of spatial information , And with AE frame ）

A:MTS You need to consider 2 Parts of , Part of it is time dependence $p(x_{t'}|x_{t}) (t'>t)$ And spatial correlation . Previous papers have only considered 2 individual variable Correlation between corresponding points . But this paper thinks , We should consider 2 individual variable One m Correlation between windows . That is to say $s_{t}$ The relevance of , among $s__{t}=\{x_{t-m/2,t+m/2}^i ,{x_{t-m/2,t+m/2}^j} \}$ . This brings about a spatial correlation matrix of time neighbors and spatial variables given the spatial state .

B： The definition of semi supervised learning , $D_{l}$ Is a label sample , $D_{u}$ There is no label sample .

SMATE:

SMATE Yes 3 The key part , One is spatiotemporal dynamics encoder, Sequence decoder, And semi supervised three-step regularization of the embedded layer .

SMB：

Simply put, it is to pass the sequence through the spatial module , Add spatial information , Personally feel , Be similar to SE modular , Only a parameter is added in the paper M, To adjust the correlation in a window .

Three step regularization （three-step regularization）：

1: Find the average value of each class . Is to put all categories 1 Of embedding vector Initialize clustering to $c_{k}$ ,k Is the number of categories .

2、 Adjust the centroid position of labeled data , It's the mean , Calculate the distance from each instance to the center of mass , Then calculate the new weight Wki

Then update the centroid

3、 Adjust the centroid position with unlabeled data , Calculate the specific gravity of the distance from the unlabeled to the center of mass , Then calculate the new weight