Ask HN: Machine Learning for Relational Data

[riphay]

I've been dabbling in machine learning and trying to find a way to apply it to some of the data at work. I work in a custom fabrication / construction company and I'm trying to build a regression tool with Tensorflow to predict fabrication time at a piece-level using time data at a higher "project-level." A simplified model of what I'm looking at:

Project-level data: Special requirements [Feature], Time taken by various departments [welding, painting, shipping, etc.] [LABELS]

Piece-level Data: Type of piece [F], Weight [F], Length [F], # of attachments [F], [Desired output time] (does not exist historically)

Attachment-level Data: Type [F], Length [F], Weight [F]

All of the classification and regression examples I've been able to find are for "flat" non-relational data. Some of the features can be flattened (such as requirements), crucially the existing labels cannot be flattened without making very broad assumptions (i.e. distributing the time among the pieces by their weight).

Are there any examples anyone is aware of that could help me build a framework for this problem? Is ML even the right approach for this? Thank you for your help.

(edited for clarity)

[davelnewton]

I'm not sure ML would be required, depending on the data you have, but maybe.

In any case, the easiest thing to do is to create flat data from your existing relational data.