White box non-linear prediction models

bannerclubAI and Robotics

Oct 20, 2013 (5 years and 29 days ago)

223 views

Special Issue of
IEEE Transactions on Neural Networks


White box non
-
linear prediction models


Guest Editors:



Baesens Bart,
K.U.Leuven, Belgium
;
University of Southampton, United Kingdom

Martens David,
University College Ghent, Association Ghent Universi
ty, Belgium;
K.U.Leuven, Belgium

Setiono Rudy,

National University of Singapore, Singapore


Zurada Jacek
, University of Louisville, Louisville, KY, USA

Dr. B
art Baesens serves as
the
Corresponding Guest Ed
itor;
all related inquiries should be
directed
to
Bart.Baesens@econ.kuleuven.be


Scope of the Special Issue


Predictive modeling aims at predicting the future using patterns learnt on past data.
Both
classification and regression are key predictive mod
eling activities.
For both tasks, a
myriad of techniques have been introduced, going from simple linear regression, to
advanced non
-
linear
prediction methods such as neural network
s
,
support vector
machines
and kernel methods in general.

Although
non
-
lin
ear techniques

typically

provide the most accurate predictive models, they are
often
not suitable
to be used
in
many
practical application
domains because of their lack of transparency and
comprehensibility.
I
n domains where validation of the underlying m
odel
is required, e.g.
credit risk analysis

and medical diagnosis, a clear insight into the reasoning

made by the
non
-
linear prediction

model is necessary

and desired
.
It is the purpose of this special
issue to solicit papers discussing various ways of ma
king non
-
linear prediction models
more interpretable and transparent, illustrated in domains where model understandability
is a key requirement.


Potential top
ics include (but not limited to
):



Non
-
linear modeling techniques

o

Neural networks, Support Vecto
r Machines,
kernel methods
, …



Integrating domain knowledge into non
-
linear models

o

Formulating domain knowledge

o

Taxonomy of constraints

o

Monotonicity constraints

o

Dealing with hard or soft domain knowledge constraints

o

Identification and resolution of knowledg
e conflicts

o

Knowledge representation

o

Knowledge fusion



Rule extraction from non
-
linear models

o

Types of rules (propositional, M
-
of
-
N, oblique, fuzzy, …)

o

Rule representation

o

Decompositional methods

o

Pedagogical methods

o

Eclectic methods

o

Decision tree extraction



Sensitivity analysis and input selection methods



Semi
-
supervised learning

o

Transductive approaches

o

Co
-
training



Active learning

o

Cost
-
based learning

o

Intelligent sampling


o

Data generation



Graph based methods and representations



Two stage models

o

Generalised ad
ditive models



Model correctness

o

Model performance

o

Model interpretability

o

Model justifiability



Real
-
life applications

o

Financial engineering (credit scoring, bankruptcy prediction, stock
prediction, fraud det
ection, anti
-
money laundering
)

o

Marketing engineeri
ng (churn

prediction, response modeling
, customer
lifetime value, …
)

o

Public sector
applications

o

Terrorism prevention

o

Medical diagnosis

o

Bio
-
informatics

o

Text mining

o

Web analytics

Submission Instructions


Both Regular (full) and Brief papers can be submitted.

All submitted papers for the
special issue should be submitted
by
May 31
st
, 2010

to the TNN submission webpage for
review using Manuscript Central. Please log on to
http://
mc.manuscriptcentral.com/tnn

and follow the directions for submiss
ion of your paper. On the first
page of the submitted
manuscript as well
as on the Author's Cover Letter
(during the subm
i
ssion procedure),
please indicate cle
arly that the submitted manuscript is submitted to the specific TNN
Special Issue. Before submission, please read carefully the Information for Authors on
the TNN webpage:
http://ieee
-
cis.org/pubs/tnn