spark-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Sergio Ramírez-Gallego (JIRA) <>
Subject [jira] [Commented] (SPARK-6531) An Information Theoretic Feature Selection Framework
Date Wed, 25 Mar 2015 10:45:52 GMT


Sergio Ramírez-Gallego commented on SPARK-6531:

Yes, but the reviewer from Spark in GitHub suggested to me to create a different JIRA issue
for this proposal. The issue that you reference is for general discussing according to this

> An Information Theoretic Feature Selection Framework
> ----------------------------------------------------
>                 Key: SPARK-6531
>                 URL:
>             Project: Spark
>          Issue Type: New Feature
>          Components: MLlib
>    Affects Versions: 1.3.0
>            Reporter: Sergio Ramírez-Gallego
> **Information Theoretic Feature Selection Framework**
> The present framework implements Feature Selection (FS) on Spark for its application
on Big Data problems. This package contains a generic implementation of greedy Information
Theoretic Feature Selection methods. The implementation is based on the common theoretic framework
presented in [1]. Implementations of mRMR, InfoGain, JMI and other commonly used FS filters
are provided. In addition, the framework can be extended with other criteria provided by the
user as long as the process complies with the framework proposed in [1].
> -- Main features:
> * Support for sparse data (in progress).
> * Pool optimization for high-dimensional.
> * Improved performance from previous version.
> This work has associated two submitted contributions to international journals which
will be attached to this request as soon as they are accepted This software has been proved
with two large real-world datasets such as:
> - A dataset selected for the GECCO-2014 in Vancouver, July 13th, 2014 competition, which
comes from the Protein Structure Prediction field ( The
dataset has 32 million instances, 631 attributes, 2 classes, 98% of negative examples and
occupies, when uncompressed, about 56GB of disk space.
> - Epsilon dataset:
400K instances and 2K attributes.
> -- Brief benchmark results:
> * 150 seconds by selected feature for a 65M dataset with 631 attributes. 
> *  For epsilon dataset, we have outperformed the results without FS for three classifers
(from MLLIB) using only 2.5% of original features.
> References
> [1] Brown, G., Pocock, A., Zhao, M. J., & Luján, M. (2012). 
> "Conditional likelihood maximisation: a unifying framework for information theoretic
feature selection." 
> The Journal of Machine Learning Research, 13(1), 27-66.

This message was sent by Atlassian JIRA

To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message