Generalize 1D sensor arrangements to arbitrary graph structure

albertotb / solar

Solar data from OAHU

4 stars 0 forks source link

Generalize 1D sensor arrangements to arbitrary graph structure #3

Open vicgalle opened 5 years ago

vicgalle commented 5 years ago

Current model only takes into account 1D interactions. It should be possible to define a graph (for instance, via Delaunay triangulation) and then replace the 1D convolution with a graph convolution.

[ ] Define triangulation and build graph conv net.
[ ] Think how to impute NaNs in the triangulation

(Should do the second point before since it is a requirement...)

vicgalle commented 5 years ago

Maybe we can use some sort of graph encoder as in here to learn embeddings robust to missing/faulty sensors.

But, then how can we address the question of "where to place B next sensors to maximize predictability"? (i.e., need to find a way to fusion with GPs)