/beymani

Collection of Hadoop based outlier analysis implementations for fraud detection

Primary LanguageJava

Introduction
============

Beymani consists of set of Hadoop based tools for outlier and anamoly 
detection, which can be used for fraud detection.

Blogs
=====
The following blogs of mine are good source of details of beymani

http://pkghosh.wordpress.com/2012/01/02/fraudsters-outliers-and-big-data-2/
http://pkghosh.wordpress.com/2012/02/18/fraudsters-are-not-model-citizens/
http://pkghosh.wordpress.com/2012/06/18/its-a-lonely-life-for-outliers/
http://pkghosh.wordpress.com/2012/10/18/relative-density-and-outliers/


Distribution Method
===================
Use the MR  class MultiVarHistogram from the project chombo. As the name 
suggests it calculates multivariate distribution and detects outliers. 
Here is my blog post on this

http://pkghosh.wordpress.com/2012/02/18/fraudsters-are-not-model-citizens/

Average Distance
================
Use SameTypeSimilarity MR from the sifarish, to find pair wise distance  
for all data points. The outout of this MR is used as input to 
AverageDistance MR in this  project. Here is the relevanr blog post on this

http://pkghosh.wordpress.com/2012/06/18/its-a-lonely-life-for-outliers/

Relative Density
================
This approach is appropriate when the feature space is not homogegeneous 
and density varie. First, use SameTypeSimilrity to find pairwise distance.
Then use AverageDistance MR to find density. Use AverageDistance MR again
to find neighborhood groups. Use the results of the last two steps and and 
run NeighborDensity to find group wise density of each data point. Finally
run the MR RelativeDensity to find relative density of each point

More Coming.......