search

starling-lab / BoostSRL

BoostSRL: "Boosting for Statistical Relational Learning." A gradient-boosting based approach for learning different types of SRL models.
https://starling.utdallas.edu
GNU General Public License v3.0
32 stars 21 forks source link

default inference algorithm in Boosted models #40

Open SeongwooLimKR opened 3 years ago

SeongwooLimKR commented 3 years ago

Dear all,

I have question about the inference algorithm in Boosted RDN/MLN.

In the papers, "Gradient-based Boosting for Statistical Relational Learning: The Relational Dependency Network Case" and "Gradient-based boosting for statistical relational learning: the Markov logic network and missing data cases", the authors wrote that the MC-SAT was used for the inference.

However, when I see the code in development branch, it seems like that the boosted RDN/MLN use the Gibbs sampling for inference.

Could you tell me what inference method is used for boosted RDN/MLN?

If the models use gibbs for inference, then could you explain why gibbs is used rather than MC-SAT?

Thank you, -Seongwoo

SeongwooLimKR commented 3 years ago

I clicked closing the issue by mistake. I am sorry for disturbing you.

boost-starai commented 3 years ago

Thanks for your interest. I feel that there is a misunderstanding.

MCSAT was used for Alchemy. Not for our software. The exact context is:

“We employed the default settings of Alchemy (Kok et al, 2010) for weight learning on all the datasets, unless mentioned otherwise. We set the multipleDatabases flag to true for weight learning. For inference, we used MC-SAT sampling with 1 million sampling steps or 24 hours whichever occurs earlier”

As you can see, these talk about Alchemy. We only use Gibbs Sampling for our work since it seems to suffice.

Thanks Sriraam

From: Seongwoo Lim [mailto:notifications@github.com] Sent: Thursday, October 22, 2020 5:54 AM To: starling-lab/BoostSRL BoostSRL@noreply.github.com Cc: Subscribed subscribed@noreply.github.com Subject: [External] [starling-lab/BoostSRL] default inference algorithm in Boosted models (#40)

This message was sent from a non-IU address. Please exercise caution when clicking links or opening attachments from external sources.

Dear all,

I have question about the inference algorithm in Boosted RDN/MLN.

In the papers, "Gradient-based Boosting for Statistical Relational Learning: The Relational Dependency Network Case" and "Gradient-based boosting for statistical relational learning: the Markov logic network and missing data cases", the authors wrote that the MC-SAT was used for the inference.

However, when I see the code in development branch, it seems like that the boosted RDN/MLN use the Gibbs sampling for inference.

Could you tell me what inference method is used for boosted RDN/MLN?

If the models use gibbs for inference, then could you explain why gibbs is used rather than MC-SAT?

Thank you, -Seongwoo

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHubhttps://github.com/starling-lab/BoostSRL/issues/40, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AF4KTVPQUDMCSZELZXORPCDSMAFKZANCNFSM4S3AY2RQ.