nardus/stnn

Packaged version of Delasalles et al.'s spatio-temporal neural network model

Spatio-Temporal Neural Networks for Space-Time Series Forecasting and Relation Discovery

Packaged version of Delasalles et al.'s stnn model, with modifications to allow tuning with the ray.tune library.

Original implementation:

The reference implementation is at edouardelasalles/stnn/, and is described in:

• Ziat A, Delasalles E, Denoyer L, Gallinari P. 2017. Spatio-Temporal Neural Networks for Space-Time Series Forecasting and Relations Discovery2017 IEEE International Conference on Data Mining (ICDM). Presented at the 2017 IEEE International Conference on Data Mining (ICDM). pp. 705–714. doi:10.1109/ICDM.2017.80
• Delasalles E, Ziat A, Denoyer L, Gallinari P. 2019. Spatio-temporal neural networks for space-time data modeling and relation discovery. Knowl Inf Syst 61:1241–1267. doi:10.1007/s10115-018-1291-x

Commands for reproducing synthetic experiments:

Heat Diffusion

STNN

python stnn/train_stnn.py --dataset heat --outputdir output_heat --manualSeed 2021 --xp stnn

STNN-R(efine)

python stnn/train_stnn.py --dataset heat --outputdir output_heat --manualSeed 5718 --xp stnn_r --mode refine --patience 800 --l1_rel 1e-8

STNN-D(iscovery)

python stnn/train_stnn.py --dataset heat --outputdir output_heat --manualSeed 9690 --xp stnn_d --mode discover --patience 1000 --l1_rel 3e-6

Modulated Heat Diffusion

STNN

python stnn/train_stnn.py --dataset heat_m --outputdir output_heat_m --manualSeed 679 --xp stnn

STNN-R(efine)

python stnn/train_stnn.py --dataset heat_m --outputdir output_heat_m --manualSeed 3488 --xp stnn_r --mode refine --l1_rel 1e-5

STNN-D(iscovery)

python stnn/train_stnn_.py --dataset heat_m --outputdir output_m --xp test --manualSeed 7664 --mode discover --patience 500 --l1_rel 3e-6

Data format

The file heat.csv contains the raw temperature data. The 200 rows correspond to the 200 timestep, and the 41 columns are the 41 space points. The file heat_relations.csv contains the spatial relation between the 41 space points. It is a 41 by 41 adjacency matrix A, where A(i, j) = 1 means that series i is a direct neighbor of series j in space, and is 0 otherwise.