Iterative Stepwise Selection and Threshold for Learning
Causes in Time Series
Jianxin Yin,
Shaopeng Wang,
Wanlu Deng,
Ya Hu,
Zhi Geng - Pekin University, China
When we explore the causal relationship among time series variables, we first remove the
potential seasonal term then we deal with the problem in the feature selection framework.
For a time series with seasonal term, we use several sequences of sin(t) and cos(t) functions
with different frequencies to design a 'pseudo' design matrix, and the seasonal term is
removed by getting the regression residual of the original series on this 'pseudo' design
matrix. An iterative stepwise subset selection and threshold method are then applied.
The cut-value for the threshold is selected by an EBIC criterion. Some simulations are
performed to assess our method. In the PROMO task of the Potluck challenge, we apply
our method and obtain a specificity of above 77% while keep the sensitivity of around 89%
on the PROMO task.