David Leifker created SPARK-19979:
-------------------------------------
Summary: [MLLIB] Multiple Estimators/Pipelines In CrossValidator
Key: SPARK-19979
URL: https://issues.apache.org/jira/browse/SPARK-19979
Project: Spark
Issue Type: Improvement
Components: MLlib
Affects Versions: 2.1.0
Reporter: David Leifker
Update CrossValidator and TrainValidationSplit to be able to accept multiple
pipelines and grid parameters for testing different algorithms and/or being
able to better control tuning combinations. Maintains backwards compatible API
and reads legacy serialized objects.
The same could be done using an external iterative approach. Build different
pipelines, throwing each into a CrossValidator, and then taking the best model
from each of those CrossValidators. Then finally picking the best from those.
This is the initial approach I explored. It resulted in a lot of boiler plate
code that felt like it shouldn't need to exist if the api simply allowed for
arrays of estimators and their parameters.
A couple advantages to this implementation to consider come from keeping the
functional interface to the CrossValidator.
1. The caching of the folds is better utilized. An external iterative approach
creates a new set of k folds for each CrossValidator fit and the folds are
discarded after each CrossValidator run. In this implementation a single set of
k folds is created and cached for all of the pipelines.
2. A potential advantage of using this implementation is for future
parallelization of the pipelines within the CrossValdiator. It is of course
possible to handle the parallelization outside of the CrossValidator here too,
however I believe there is already work in progress to parallelize the grid
parameters and that could be extended to multiple pipelines.
Both of those behind-the-scene optimizations are possible because of providing
the CrossValidator with the data and the complete set of pipelines/estimators
to evaluate up front allowing one to abstract away the implementation.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
