[
https://issues.apache.org/jira/browse/MAHOUT-1177?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13647841#comment-13647841
]
Yu Lee commented on MAHOUT-1177:
--------------------------------
Hello Robin Anil, Jeff Eastman, Dan Filimon, and Ted Dunning,
Yexi and I (Yu Lee) are new to this Mahout community. We want to contribute to
the improvement of Mahout by reforming and simplifying the clustering APIs per
the following link:
https://issues.apache.org/jira/browse/MAHOUT-1177?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13644120#comment-13644120
We have gone through the code of Mahout clustering. Now we have some ideas
about improving it:
=========================================================================================
Addressing the problems in the current interface:
Testing cases are missing. For example, in spectral kmeans clustering, the run
methods of SpectralKmeansDriver and EigencutsDriver are not tested
Documentations are missing for some methods. For example: in the run method of
DirichletDriver, the description of parameter 'numModels' is missing; in the
run method of SpectralKmeansDriver, the description of some arguments are
missing
Some testing methods do not contain the specific description of some arguments.
For example: in the run method of FuzzyKmeansDriver, the description of an
argument of "m" (fuzzification factor) is missing. Although a wiki link
regarding "Clustering Analysis" is given, it is not clear enough.
-----------------------------------------------------------------------------------------
Implementing some new clustering algorithms
Agglomerative hierarchical clustering, which will cluster the data points into
a dendragram, so that user could indicate whatever number of clusters as they
want. (http://en.wikipedia.org/wiki/Hierarchical_clustering)
Dbscan, which is a density based clustering method being able to identify
clusters with arbitrary shapes, and is useful in spatial clustering.
(http://en.wikipedia.org/wiki/DBSCAN)
-----------------------------------------------------------------------------------------
Providing a new unified interface
Currently, each clustering algorithm has its own implemented class with
different interfaces (i.e., run methods in different Drivers have different
argument list). However, it is better to have a unified interface to execute
all available clustering methods, and an example interface is as follows:
Clustering-run(input, output, methodClass,clusteringConfig)
Here, the "methodClass" indicates a specific clustering method, while
"clusteringConfig" indicates the configuration for this specific clustering
method.
=========================================================================================
Could you please let us know what you think about our ideas?
> GSOC 2013: Reform and simplify the clustering APIs
> --------------------------------------------------
>
> Key: MAHOUT-1177
> URL: https://issues.apache.org/jira/browse/MAHOUT-1177
> Project: Mahout
> Issue Type: Improvement
> Reporter: Dan Filimon
> Labels: gsoc2013, mentor
>
> Clustering is one of the most used features in Mahout and has many
> applications [http://en.wikipedia.org/wiki/Cluster_analysis#Applications].
> We have of lots clustering algorithms. There's:
> - basic k-means
> - canopy clustering
> - Dirichlet clustering
> - Fuzzy k-means
> - Spectral k-means
> - Streaming k-means [coming soon]
> We want to make them easier to use by updating the APIs and make sure they
> all work in the same way have consistent inputs, outputs, diagnostics and
> documentation.
> This is a great way to gain an in-depth understanding of clustering
> algorithms, familiarize yourself with Hadoop, Mahout clustering and good
> software engineering principles.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
