dataset representation

[yuanqing-wang]

what features do we want for Dataset object?

now we have the following functionalities

• [ ] csv import
• [ ] batch
• [ ] split
• [ ] temporal splitting

we should consider adding the following:

• [ ] extra features (which level)
• [ ] different node representation input

[miretchin]

Based on our discussion earlier, elaborating on the "different node representation input":

1. Have dataset object annotate different inputs with different style of representing graphs. i.e., different preprocessing steps that yield different representations, like dataset.smiles_representation or some other representation.
2. Have datasets be typed. Typed according to what input representation it assumes, and then have a flag that can be passed forward to the model (and by extension, the models should be typed).

[dnguyen1196]

Potentially, we might want to do something like this paper which the input has both the graph representation + junction tree representation (I digged in the code and it's possible to process molecules into junction trees either on the fly or as part of preprocessing)

https://arxiv.org/pdf/2006.12179.pdf

[yuanqing-wang]

https://docs.google.com/document/d/1Yp4qZ-9U1kPQI3upwDzJrys8Gjr0F92O_8PVrK-UfqU/edit

design doc @miretchin @karalets @dnguyen1196