Github user jkbradley commented on the pull request:
https://github.com/apache/spark/pull/4047#issuecomment-70146828
@witgo I agree that there are 2 different use regimes for LDA:
interpretable topics and featurization. The current implementation follows
pretty much every other graph-based implementation Iâve seen:
* 1 vertex per document + 1 vertex per term
* Each vertex stores a vector of length # topics.
* On each iteration, each doc vertex must communicate its vector to any
connected term vertices (and likewise for term vertices), via map-reduce stages
over triplets.
I have not heard of methods which can avoid this amount of communication
for LDA. Iâm sure the implementation can be optimized, so please make
comments here or JIRAs afterwards about that. For modified models, it might be
possible to communicate less: sparsity-inducing priors, hierarchical models,
etc.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
