Thank you so very much for the info, Ted.
-----Original Message-----
From: Ted Dunning
Sent: Friday, June 07, 2013 7:58 AM
To: Mahout Dev List
Subject: Re: Clarification ---- Re: Can Mahout do pattern recognition?
On Wed, Jun 5, 2013 at 5:04 PM, Mimi Tam <mimi....@ieee.org> wrote:
I have one input stream of many Radio network related Signaling records in
XML from various sources. I need to group records that belong to a Base
Station/Mobile Operator/mobile device and monitor its usage pattern e.g.
high traffic at noon time and low traffic at midnight, location of device
path tracking,... etc. The detected usage pattern will be used for various
purposes. These are not personal device but M2M devices.
Where can I find info on an implementation/integration of this kind of
pattern recognition algorithm in Mahout?
Mahout does not have anything that specifically does these things.
But ...
Mahout does have good k-means clustering (thanks to Dan making my prototype
good enough to commit). With k-means clustering you can do volume
quantization on positions which turns paths into sequences of symbols. You
can then to cooccurrence analysis on lagged versions of these sequences to
get interesting path predictions.
You can also use similar techniques to find traffic anomalies. The basic
idea is that can use pure spatial or spatio-temporal clustering to build
models that predict traffic. Comparing non-temporal models can tell you
where time is important and comparing actual versus predicted traffic can
help you find anomalies. These anomalies might be network disruptions or
might be disturbances in your own systems.
