Rule learning

[zachary...@gmail.com]

I'm new to PSL and there is still a lot that I don't get but I wanted to ask one question. I get how you can learn the weights but can you learn the rules from data data as well? It seems a bit hand-wavy when it says you create rules based on an my intuition of the data. What if I pick a poor set of rule? Will it simply fail to generated inferences based on my poor rule set?

[vgod...@ucsc.edu]

Hi Zachary, 

what you're talking about is called structure learning - where rules are learnt from the data. Unfortunately it isn't supported in the current version of PSL.
To answer your other question, yes the quality of inference is largely based on the rules that you create.

You can find further information on structure learning here - https://linqs.soe.ucsc.edu/sites/default/files/papers/VEmbar-StarAI2018.pdf

- Vihang

[zachary...@gmail.com]

Thank you for the response. I'm going to look into how to integrate psl and this project https://github.com/AKSW/Mandolin

It says that it does inference but it's difficult to tell how or where it is done but it *does* produce rules and weights.

[zachary...@gmail.com]

Looks like inference is based on MLN.