search

GestaltCogTeam / GinAR

Code for our SIGKDD'24 paper GinAR: An End-To-End Multivariate Time Series Forecasting Model Suitable for Variable Missing
22 stars 4 forks source link

Dataset construction #2

Open chenywu opened 6 months ago

chenywu commented 6 months ago

Hi Chengqing,

Congratulations for accepted by KDD'24.

I am interested to know how did you construct the dataset with missing values?

ChengqingYu commented 5 months ago

Hi Chengqing,

Congratulations for accepted by KDD'24.

I am interested to know how did you construct the dataset with missing values?

Hi, sorry for my late late reply.

The setting of missing values is introduced in Section 3.1 (Preliminaries) of the paper. We randomly generate M numbers proportionally, and for the input feature X (which consists of N time series, with M being smaller than N), we convert the values of the corresponding M variables among the N variables to zero.

Setting missing variables to zero is based on methods discussed in time series imputation-related papers such as GRIN.