Fix the code that aligns returned data with the output schema

[karlhigley]

• match_representations was implemented a bit weirdly and seems to have had corner cases we weren't tracking. This updates the implementation to be less clever and more straightforward, which also makes it work better.

[github-actions]

Documentation preview

https://nvidia-merlin.github.io/systems/review/pr-335

[oliverholworthy]

match_representations was implemented a bit weirdly and seems to have had corner cases we weren't tracking. This updates the implementation to be less clever and more straightforward, which also makes it work better.

By 'work better', is this about filtering out columns that are not present in the schema from the data passed to match representations?

[karlhigley]

By 'work better', is this about filtering out columns that are not present in the schema from the data passed to match representations?

It wasn't intended to do that, but I suppose it may have that effect. TBH, the first implementation was a bit clever and hard to think about, so instead of trying to fix it we just started over from scratch and wrote it in a fashion that was easier to understand. Once we did that, tests started passing so we called it a win. (In particular, the test for Radek's TF transformer next-item prediction example notebook.)