Time series data: prediction

Multivariate time series (Multivariate time series MTS) Of Prediction is widely used in meteorology, transportation and other fields .

Due to data collection 、 Transmission and storage limitations , In the real world MTS Data usually contains missing values , Make use of existing MTS prediction model （ Such as linear regression and cyclic neural network ） Become infeasible .

Although many efforts have been made to solve this problem , But most of them only rely on local dependencies to estimate missing values , While ignoring the global time dynamics .

When the deletion rate is high , Or when the data has continuous missing values , Local dependence / Patterns will become less useful ; Therefore, exploring the global pattern at the same time can alleviate such problems .

therefore , Joint modeling of local and global time dynamics for MTS Prediction is very useful .

Therefore, this paper studies the missing values by jointly exploring the local and global time dynamics MTS New problems in prediction .

This paper proposes a new framework LGnet, It uses memory networks to explore global patterns from local perspective estimation .

In this paper, antagonistic training is further introduced to enhance the modeling of global time distribution .

Experimental results on real-world datasets show that LGnet For missing values MTS The effectiveness of prediction and its robustness under various missing rates .

Reference material ：
Time series prediction @Joint modeling of local and global temporal dynamics for multivariate time series forecast…