[
https://issues.apache.org/jira/browse/SPARK-6509?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15431378#comment-15431378
]
Sean Owen commented on SPARK-6509:
----------------------------------
The outcome of many "add X to MLlib" proposals, where it's not clear obvious
interest in adding it straight away, is to implement it outside Spark and
perhaps let it demonstrate from there that it's used widely. This is how things
like CSV parsing came in. MLlib implementations are so separable that we don't
really need or even want everything to be part of Spark itself. Some things are
useful just niche.
> MDLP discretizer
> ----------------
>
> Key: SPARK-6509
> URL: https://issues.apache.org/jira/browse/SPARK-6509
> Project: Spark
> Issue Type: New Feature
> Components: MLlib
> Reporter: Sergio Ramírez
>
> Minimum Description Lenght Discretizer
> This method implements Fayyad's discretizer [1] based on Minimum Description
> Length Principle (MDLP) in order to treat non discrete datasets from a
> distributed perspective. We have developed a distributed version from the
> original one performing some important changes.
> -- Improvements on discretizer:
> Support for sparse data.
> Multi-attribute processing. The whole process is carried out in a single
> step when the number of boundary points per attribute fits well in one
> partition (<= 100K boundary points per attribute).
> Support for attributes with a huge number of boundary points (> 100K
> boundary points per attribute). Rare situation.
> 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
> (http://cruncher.ncl.ac.uk/bdcomp/). 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:
> http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/binary.html#epsilon.
> 400K instances and 2K attributes
> We have demonstrated that our method performs 300 times faster than the
> sequential version for the first dataset, and also improves the accuracy for
> Naive Bayes.
> Publication: S. Ramírez-Gallego, S. García, H. Mouriño-Talin, D.
> Martínez-Rego, V. Bolón, A. Alonso-Betanzos, J.M. Benitez, F. Herrera. "Data
> Discretization: Taxonomy and Big Data Challenge", WIRES Data Mining and
> Knowledge Discovery. In press, 2015.
> Design doc:
> https://docs.google.com/document/d/1HOaPL_HJzTbL2tVdzbTjhr5wxVvPe9e-23S7rc2VcsY/edit?usp=sharing
> References
> [1] Fayyad, U., & Irani, K. (1993).
> "Multi-interval discretization of continuous-valued attributes for
> classification learning."
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
