APBRmetrics

I don't know if this has been discussed before, but it seems that there could be a place on this board for discussion of new ways to model players, either possible ideas, or new methods to test things out.

For instance, including some small number of interaction terms in a Plus minus regresssion. My idea would be to use the interactions that are highly positive/highly negative from another paper(Maybe the Big 2's and big 3's paper from last Sloan conference) and include a dummy whenever the player types were a certain combination of players.

Impossible to include every interaction term, but maybe the ones that there's a good reason to include could be chucked in there.

Any ideas on how to improve this? A method that could be used?

Maybe one could try to include first level interactive terms for the top 3-5 guys on the team.

Has anyone applied machine learning or some other advanced technique to the APM process?