[jira] [Created] (SPARK-4500) Improve exact stratified sampling implementation

["Joseph K. Bradley (JIRA)" <ji...@apache.org>]

Joseph K. Bradley created SPARK-4500:
----------------------------------------

             Summary: Improve exact stratified sampling implementation
                 Key: SPARK-4500
                 URL: https://issues.apache.org/jira/browse/SPARK-4500
             Project: Spark
          Issue Type: Improvement
          Components: MLlib
    Affects Versions: 1.2.0
            Reporter: Joseph K. Bradley


The current implementation for exact stratified sampling (sampleByKeyExact) could be more efficient.  Proposed algorithm sketch:

* Sampling is done separately for each stratum.  Here, all counts are w.r.t. a fixed stratum.
* Let:
 * N = total number of elements in all partitions
 * N_partition = number of elements in a given partition
* This method uses 2-tiered sampling:
 * Tier 1 (on driver): Select the sample sizes {n_partition} for each partition.
  * Without replacement, this uses a multivariate hypergeometric distribution.
  * With replacement, this should use a multinomial distribution.
 * Tier 2 (in parallel): Select a sample of size n_partition.

If anyone is interested in implementing this, I have a rough draft which works without replacement, but it needs to be cleaned up and augmented to do sampling with replacement too.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org