[jira] [Commented] (MADLIB-1351) Add stopping criteria on perplexity to LDA

["Scott Hajek (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/MADLIB-1351?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16873498#comment-16873498 ] 

Scott Hajek commented on MADLIB-1351:
-------------------------------------

I like the proposed API for stopping criterion on LDA. I think *evaluate_every* and *perp_tol*  are sufficient, and the default values are reasonable. I might suggest calling the latter by a slightly less abbreviated name for clarity, like *perplex_tol* or *perplexity_tol*. The brevity isn't as necessary in MADlib since you don't/can't specify with keyword arguments in SQL function calls anyway. It's mainly for documentation.

 

Although I guess [named notation|[https://www.postgresql.org/docs/9.0/sql-syntax-calling-funcs.html]] will be (or maybe already is) available once Greenplum catches up with later versions of Postgres. In that case, maybe a good compromise would be *perplex_tol*. 

> Add stopping criteria on perplexity to LDA
> ------------------------------------------
>
>                 Key: MADLIB-1351
>                 URL: https://issues.apache.org/jira/browse/MADLIB-1351
>             Project: Apache MADlib
>          Issue Type: Improvement
>          Components: Module: Parallel Latent Dirichlet Allocation
>            Reporter: Frank McQuillan
>            Assignee: Himanshu Pandey
>            Priority: Minor
>             Fix For: v1.17
>
>
> In LDA 
> http://madlib.apache.org/docs/latest/group__grp__lda.html
> make stopping criteria on perplexity rather than just number of iterations.
> Suggested approach is to do what scikit-learn does
> https://scikit-learn.org/stable/modules/generated/sklearn.decomposition.LatentDirichletAllocation.html
> evaluate_every : int, optional (default=0)
> How often to evaluate perplexity. Set it to 0 or negative number to not evaluate perplexity in training at all. Evaluating perplexity can help you check convergence in training process, but it will also increase total training time. Evaluating perplexity in every iteration might increase training time up to two-fold.
> perp_tol : float, optional (default=1e-1)
> Perplexity tolerance in batch learning. Only used when evaluate_every is greater than 0.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)