I am using GCN for traffic forecasting domain and would like to ask about the usage of a weighted undirected adjacency matrices. It would be a spatio temporal model with GCN modelling the spatial information whilst GRU or LSTM modelling the temporal.
As each node (traffic location) may not necessarily have similar values, but fluctuate similarly (similar peak and non peak times), the normalisation technique of the vanilla GCN may cause the the model to under-predict the values of locations with typically less amount of traffic.
I'm still new to the world of GNNs but if you could give me some direction and some people have directed me to this paper/technique. Wondering if you have some ideas in implementing this? I am thinking that a weighted adjacency instead of adjacency with 1, and 0 could be used. But not sure if this would break the model..
I am using GCN for traffic forecasting domain and would like to ask about the usage of a weighted undirected adjacency matrices. It would be a spatio temporal model with GCN modelling the spatial information whilst GRU or LSTM modelling the temporal.
As each node (traffic location) may not necessarily have similar values, but fluctuate similarly (similar peak and non peak times), the normalisation technique of the vanilla GCN may cause the the model to under-predict the values of locations with typically less amount of traffic.
I'm still new to the world of GNNs but if you could give me some direction and some people have directed me to this paper/technique. Wondering if you have some ideas in implementing this? I am thinking that a weighted adjacency instead of adjacency with 1, and 0 could be used. But not sure if this would break the model..
Thanks for any response