DM: Introductions

[Subrata Chatterjee]

Hello Folks:

Here is my introduction.
I am doctorate student in Industrial Engineering
at Purdue University, West Lafayette, Indiana, USA.
For my doctorate thesis, I have developed a connectionist 
model for classifying motor vehicle accidents. These
accidents are *actual* reports from *real* people
involved in automobile accidents who are insured
with a leading insurance company. 

The objective of my thesis is to help in understanding 
the patterns that exist in automobile accidents 
(such as whether a rear-ending accident is always preceded by
a stopping or braking activity). While rule-based systems can help
in capturing relationships among objects in many settings,
connectionist models (such as spreading activation and models 
based on competitive learning) have been shown to more successful in
applications that involve natural language with no obvious or poor
grammar (which is typical for most real-life datasets). 
The network in my model provides an estimate
of the posterior probability of an accident category given the 
description of an accident. Techniques from  NLP(Porter's stemming
algorithm and a very short stop-word list), 
signal processing (feature extraction and sphering of the database)
and large-scale optimization were used to 
develop the model. Feature extraction was done using
LSI. However, I am exploring other techniques such as
Fishers discriminant analysis. I have also used projection
pursuit to view the feature-extracted dataset which allowed me 
to use skip layer connections in my network. 

Scaled conjugate gradient
was used to train the network. Stopped training with 10-fold 
cross-validation was used for determining the generalization
performance of the network. The network was tested on an
independent set of accident reports. Three hypothesis were
tested for validating the model and they have shown that the
network can perform reasonably well on unseen data sets. However,
enhancements are possible, using improved feature detection,
and time-delay networks. 

I am graduating in the Fall semester and will look for a
data mining job in the industry.

Thanks
Subrata
==================================================================
Subrata Chatterjee
218 Grissom Hall
School of Industrial Engineering
Purdue University
West Lafayette, IN 47907
Office: (765) 494 6147
Home:   (765) 743 5546 
=================================================================