HKUDS/UrbanGPT

Question about GCN in ST_Enc.

Closed this issue · 2 comments

Hello, I read your wonderful paper but still have some question.

It seemes that the GCN is always unable in ST_Enc, does it conflict with the statement "spatio-temporal encoder is designed to be independent of graph structures when modeling spatial correlations" in the paper? Or there is any misunderstanding I made?

Snipaste_2024-07-12_12-11-34

Thanks!

Hello, thank you for your interest in our work. In many real-world scenarios, the graph structure is unknown; therefore, the designed spatio-temporal encoder does not rely on graph structures. We model spatial correlations using prompt instructions that contain spatial information, such as POIs.

Thank's for your explanation!