I have condtructed a simulation of approximately 20 agents travelling through a 29 by 35 square grid.

I recorded the position of each agents on the grid for each timestep, and figured I could do something like the examples below:

https://keras.io/examples/timeseries/timeseries_traffic_forecasting/

https://stellargraph.readthedocs.io/en/stable/demos/time-series/gcn-lstm-time-series.html

The goal is to predict the occupancy of a node in the future timeframe.

Based on what I understood, I constructed a 1015(29*35) by 1015 adjacency matrix and 1015 by (total timestep length, up to 40000) matrix to represent the occupancy of a node. Since those input matrices were large, putting those pytorch ran out of memory. I was thinking of using sparse matrices instead, but I would like to know if such thing is possible for GNN, and I want to know if there is a better approach.

I’m fairly new to neural networks so I greatly appreciate your help!