Multivariate time-series forecasting for GCN-LSTM

stellargraph / stellargraph

StellarGraph - Machine Learning on Graphs

https://stellargraph.readthedocs.io/

Apache License 2.0

2.93k stars 431 forks source link

Multivariate time-series forecasting for GCN-LSTM #1527

Closed timpitman closed 4 years ago

timpitman commented 4 years ago

Description

Can we use Keras's existing multi-variate LSTM?

User Story

As a ...

I want to [accomplish some things]

So that [achieve a larger goal]

Done Checklist

[ ] Produced code for required functionality
[ ] Tests written and coverage checked
[ ] Code review performed
[ ] Documentation on Google Docs (if applicable)
[ ] Documentation in repo
[ ] Version number reflects new status
[ ] CHANGELOG.md updated
[ ] Team demo

huonw commented 4 years ago

One approach here might be generalising the transposition for graph convolution to include some reshaping. For this, the input would have shape something like (batch dim, num nodes, time steps, variate dimension), and the be reshaped to run graph convolution on time steps * variate dimension values, and LSTM on num nodes * variate dimension values:

https://github.com/stellargraph/stellargraph/blob/ee3bd69435ba0a7b630dd0ae566f095c99167e66/stellargraph/layer/gcn_lstm.py#L178