A NEW STATISTICAL MODEL
To Designing a Decision Support System
Morteza Zahedi, Ali Pouyan and Esmat Hejazi
Computer and Information Technology Department, Shahrood University of Technology, Shahrood, Iran
Keywords: Decision Support System, Technical Support Group, Statistical Pattern Recognition, Hidden Markov Model,
Human Judgment, Expert Systems.
Abstract: In this paper we propose a new statistical approach to simulate a technical support center as a help desk for
a web site which makes use of scientific documents and university protocols for the students and lecturers.
In contrary to the existing statistical approaches which are modelled by general statistical graphs named
Bayesian network or decision graph, we propose a statistical approach which can be used consistently in
different domains and problem spaces without any need for a new designing regarding the new domain.
Furthermore, the proposed statistical model which is trained by a set of training data collected from the
experts in a special field is applicable to high-dimensional, large-sized, non-geometric-based data for
decision making support.
1 INTRODUCTION
Corporations and companies often provide a help
desk support to their customers in the sense of
installation and usage of the products and
troubleshooting the problems. Also, some schools
offer classes in which they perform similar tasks as a
help desk to help the students. In addition to the
typical help desks, there are also many technical
support forums freely available on the Internet,
wherein expert and experienced users volunteer to
help novices particularly in the field of computer
programming and coding. As inside the companies,
institutions and schools, employees and teachers
need some information and technical support guides,
there are also in-house help desks providing the
same kind of help for employees, lecturers, or other
internal associates only. It is very important for these
services to be accessible 24-hours a day.
The various kinds of help desks introduced here
are implemented via a toll-free telephone number or
various online media such as website and/or e-mail.
Not only a help desk is very useful to support the
costumers and offered services of a company to the
costumers, but a help desk software can also be an
extremely beneficial tool when it records the
receiving questions and problems from the costumer
side. The recorded information can be used to find,
analyze, and eliminate common problems in an
organization's computing environment. As a help
desk communicates daily with numerous customers
or employees, this gives the help desk the ability to
monitor the user environment for issues from
technical problems to user preferences and
satisfactions. Such information gathered at the help
desk is very valuable in planning and preparation to
other units in the departments such as sales and
product development (Middleton, 1996).
A very important problem for many help desks is
to be strictly rostered. Time is an important
parameter in such help desks to perform some tasks
such as following up problems, returning phone
calls, and answering questions via e-mail. The
incoming phone calls and receiving queries by email
or recorded messages via web portals are random in
nature, so a rostered help desk agent schedules
should ensure that all analysts get time to follow up
on calls, and also ensure that analysts are always
available to take incoming phone calls.
Due to the time constraints to response the receiving
questions and high variaty of problems that occur in
the services, causing different queries from the user
side, the help desk is often referred to as the “hell
desk” by the desk staff who work there. Thus, a
decision support system (DSS) which is able to give
some advices to the desk staff or answering the
queries on the desk automatically is very useful. It
can be emplemented along many new upcoming
524
Zahedi M., Pouyan A. and Hejazi E. (2008).
A NEW STATISTICAL MODEL - To Designing a Decision Support System.
In Proceedings of the Tenth International Conference on Enterprise Information Systems - AIDSS, pages 524-530
DOI: 10.5220/0001704905240530
Copyright
c
 SciTePress

technical support organizations which offer
comprehensive computer repair services or guiding
the users of a web portal.
Figure 1: A two-level help-desk.
A typical help desk has two levels to handle
different types of questions. It provides the users a
central point to receive help on various problem
issues. The user notifies the help desk of his or her
issue, and the help desk issues a ticket including
details of the problem. The first-level help desk is
prepared to answer the general queries, the most
commonly asked questions, or provide resolutions
that often belong to the list of frequently asked
questions (FAQ). The second level help desk
consists of specialized teams ready to solve more
complicated problems which are not solved in the
first level. If the first level is able to solve the issue,
the ticket is closed and updated with documentation
of the solution to allow other help desk technicians
to use it as a reference. If not, it will be dispatched to
a second level where more experienced and expert
people in a particular field are waiting for the
queries. The queue manager will assign a ticket to
one of the specialized teams based on the type of the
issue. These comprehensive help desks need a team
of experts in the first and second level to answer
general questions and queries, and some specialists
in the particular fields of the services, to work on-
line 24 hours a day.
In this work we introduce a comprehensive help
desk which uses the knowledge of the experts in a
particular case. Not only can it help the desk staff as
a decision support system by providing some
advices, but it can also work alone instead of an
expert team answering the questions and giveing
advices to the visitors and users of a web portal
automatically.
2 STATE-OF-THE-ART AND
RESEARCH OBJECTIVES
In contrary to the expert systems implemented by
knowledge base construction rules in the artificial
intelligence (AI) discipline like genetic algorithms
(Leu S. S., 2002, Turban, 2004), and knowledge
representation methods (Michalewicz, Z.,2005,
Bing
Nan Li, 2008, and Ren-Jye Dzeng, 2007), the statistical
methods aims at providing a rational decision in the
context of probability theory and decision theory.
The decision support systems which make rational
decisions use collection of data gathered from the
experts in a special field to construct statistical
models.
It has been rather convincingly presented in
numerous empirical researches that human
judgment and decisions made by the experts are
based on intuitive strategies. The intuitive strategies
oppose to rational decision making methods which
theoretically use reasoning rules. Empirical
evidence and also several studies of expert
performance in realistic settings show that experts
and experienced people are more accurate than
novices within their area of expertise, even though
they are also liable to the same judgmental biases as
novices and apparent errors and inconsistencies
occur in their judgment. An informal review of the
available evidence and literature review can be
found in the book by (Robyn M. Dawes, 1988).
Although the decisions made by heuristic methods
are not based on optimal decision making rules and
violates probability axioms by judgment biases, the
intuitive strategies or judgmental heuristics help the
experts and expert systems in the context of
decision making by reducing the cognitive load. In
an anthology edited by Kahneman, Slovic, and
Tversky (Kahneman, 1982), there is a formal
discussion of the most important research results
along with experimental data.
In the general statistical approaches proposed for
the decision support systems, very detailed analysis
of domain tasks and information analysis theory is
used (Bertin, 1983) to construct statistical models.
Typically, the statistical approaches construct a
Bayesian network or other kinds of decision graphs
which is strictly dependent on a problem space
(
Dorner S., 2007). These methods focus on the special
domain and problem space, analyze each node in the
Experts
team 1
Experts
team 2
Experts
team 3
First level experts
Costumers
Answers
Questions
A NEW STATISTICAL MODEL - To Designing a Decision Support System
525

problem space, identify each data item and
information component (variable) for the nodes, the
characteristics of the information components, and
the relationships among the nodes. Finally, a
conceptual database is used to store and retrieve the
data collection gathered from the experts. The
relationships among variables and rules that apply to
problem-solving activities are described as a set of
knowledge and form a conceptual knowledge base.
Also, some researchers have tried to represent
huge data sets graphically. Computer hard disk
usage is represented by development of TreeMaps,
where huge data sets are involved and compressed
(Shneiderman, 1992). Also, a punctuation graph is
used to represent a technical document to allow the
writer to detect potentially overly complex
sentences, as well as to recognize familiar patterns
(Perlman, 1983). The approaches using Bayesian
networks or statistical graphs suffer from the
limitation that exists in many business domains; the
relationships among data are very complicated and
cannot be presented with geometric structures such
as hierarchies, linear, or networks. In other words,
due to the nature of non-geometric data or non-
spatial data, there is no obvious physical model that
can be used to represent the data that humans can
understand objectively.
Furthermore, in most business and management
domains, problem-solving is overwhelming because
of the large amount of complicated data, multiple
complex relationships among data, and the
negotiability of the constraints. Thus, in such
systems including the data with complicated
structure, it is difficult to construct a decision graph.
Furthermore, general purpose representations are not
easy to apply to a specific domain due to the
complexity of data in different domains and
sophisticated underlying functionality.
Another limitation of general statistical graphs,
is that the data to be graphed must have controllable
size or dimensions. In other words, the statistical
approaches based on decision graphs cannot
represent high-dimensional, large-sized, non-
geometric-based data for decision-making support.
This research paper focuses on developing a
research strategy for building a statistical model able
to be used for non-geometric data that are massive in
both size and dimensionality to help decision makers
eventually to improve problem-solving performance
or work alone instead of a group of experts. We will
then apply the proposed statistical model to concrete
a realistic domain to verify the effectiveness of the
model. However, the proposed statistical model
itself is domain-independent. It indicates the
procedure of a decision support system as an
automatic machine translation system which first
maps receiving questions from the user-side into an
answer from the desk-side. The final output of the
proposed model is a set of most probable answers
offering the desk staff supporting the entire human
problem-solving process in a specific business
domain. By employing a nearest neighbor (NN)
classifier the best answer obtained from the
statistical model can be chosen as an exact answer
from the desk-side.
3 STATISTICAL MODELING OF
THE DECISIONS
A simple view of decision making is that it is a
problem of a choice among several alternatives. A
somewhat more sophisticated view includes the
process of constructing the alternatives, i.e. given a
problem statement, developing a list of options. A
complete picture includes a search for opportunities
of decisions, i.e. discovering that there is a decision
to be made. For instance, a manager of a company
may face a choice in which the options are clear, e.g.
the choice of a supplier from among all existing
suppliers. There are a lot of anecdotal and some
empirical evidence that structuring decision
problems and identifying creative decision
alternatives determine the ultimate quality of
decisions. Decision support systems aim mainly at
this broadest type of decision making, and in
addition to supporting choice, they aid in modeling
and analyzing systems, such as complex
organizations, identifying decision opportunities,
and structuring decision problems.
In other words a decision support system can be
simplified to a machine translation (MT) system
translating a source sentence to a target sentence.
Machine translation is a sub-field of
computational linguistics, investigating the use of
computer software to translate text or speech from
one natural language to another. Simple machine
translation methods perform simple substitution of
words in one natural language for words in another.
More complex translations may be attempted, by
using corpus techniques in order to allow better
handling of differences in linguistic typology, phrase
recognition, and translation of idioms, as well as the
isolation of anomalies. In order to improve accuracy
of the MT methods, some research groups allow for
customization by domain or profession (such as
ICEIS 2008 - International Conference on Enterprise Information Systems
526

weather reports) by limiting the scope of allowable
substitutions.
Statistical machine translation is a kind of
translation method trying to generate translations
using statistical methods based on bilingual text
corpora, such as the English-French record of the
Canadian parliament and EUROPARL, the record of
the European Parliament. By using such corpora,
impressive results are obtained translating texts of a
similar kind, but the scarceness of such corpora is
still a critical problem for machine translation.
Although the CANDIDE is the first statistical
machine translation software from IBM, currently
Google employs a statistical translation method
improving their translation capabilities by inputting
approximately 200 billion words from United
Nations materials to train their system (
Hutchins, W.
John, 1992).
Although a typical MT system translates a
source text into a target text in another language, it
looks very similar to a decision support system
which maps the user-side questions into the domain
of desk-side answers. Machine translation is not
always applied for a complete and accurate
translation of texts. Sometimes a machine translation
system is employed to perform a rough translation of
a foreign language text, like a web page or news,
which gives an idea of its contents. An MT machine
is also applied for translation aid systems to help
human translators. An inaccurate machine
translation system can be used similarly as a
decision support system.
In all these contexts, it is important to know
when the system possibly made an error, and when
one can be sure of obtaining a good translation.
Since often a translation of a sentence as a whole is
incorrect, but contains correct parts, the output of the
MT system can provide the desk staff with a list of
most probable answers. Also it is possible to employ
a nearest neighbor classifier to find the most similar
answer to the target sentence among the list of the
answers collected from the experts in the particular
field.
The statistical approach to machine translation
has received growing interest over the last years
since its introduction by the IBM research group in
the early nineties. In various comparative
evaluations, it has been proven to be competitive or
superior to other traditional approaches. The
translation quality achieved in restricted domains is
relatively high. Examples include the domains of
appointment scheduling, which was the scope of the
project Verbmobil (
W. Wahlster, 2000), or tourism
which is used in the IWSLT evaluations (Y. Akiba.,
2004). In recent years, more challenging tasks have
been tackled in SMT research. The TC-STAR
project (TCS, 2005), for example, deals with speech
translation of the plenary sessions of the European
Parliament. The domain and the vocabulary of these
speeches are open.
The goal of machine translation is the automatic
translation of a source language string
f
J
1
=
f
1
...
f
j
…
f
J
of words
f
j
into a target language
string
e
I
1
=
e
1
…
e
i
…
e
I
. In statistical
machine translation (SMT), the translation is
modeled as a decision process:
Given a source string
f
J
1
, the target string
e
I
1
with maximal posterior probability is determined:
⎭
⎬
⎫
⎩
⎨
⎧
⎪
⎪
⎭
⎪
⎪
⎬
⎫
⎪
⎪
⎩
⎪
⎪
⎨
⎧
⎭
⎬
⎫
⎩
⎨
⎧
=
=
=
)
1
Pr().
11
Pr(
maxarg
1
,
)
1
Pr(
)
1
Pr().
11
Pr(
maxarg
1
,
)
11
Pr(
maxarg
1
,
ˆ
ˆ
1
I
e
I
e
J
f
I
eI
J
f
I
e
I
e
J
f
I
eI
J
f
I
e
I
eI
e
I
(1)
Through this decomposition of the posterior
probability
)Pr(
11
JI
fe , two knowledge sources are
obtained: the translation model
)Pr(
11
IJ
ef
and the
language model
)Pr(
1
I
e . Both of them can be
modeled independently of each other. The
translation model is responsible of linking the source
string
J
f
1
and the target string
I
e
1
, i.e. for capturing
the semantics of the sentence. The target language
model
)Pr(
1
I
e assigns probabilities to target word
sequences. It models the well-formedness or the
syntax in the target language.
The probability of the source sentence,
)Pr(
1
J
f ,
is usually omitted in the maximization because it
does not affect the choice of the target word
sequence. Nevertheless, it will be shown later that
this probability is important for the methods
suggested in this thesis.
A NEW STATISTICAL MODEL - To Designing a Decision Support System
527

The overall architecture of the statistical
translation approach is depicted in figure 1.1.
The correspondence between the words in the
source and the target string is described by
alignments which can be viewed as mappings
{}
Iiaja
j
,...,,...,1: ∈→ assigning a target
position
j
a to each source position
j
(Brown et al.
1993]. An artificial target position zero is introduced
for mapping source words that do not have any
equivalence in the target string. The alignment is
introduced into the model as a hidden variable:
∑
=
J
a
IJJIJ
eafef
1
),Pr()Pr(
11111
(2)
Finally, using a nearest neighbour classifier, we
find the most similar sentence among the target
sentences to the translated sentence as the final
choice of the decision support system.
Figure 2: Architecture of the statistical approach for a
technical support center.
4 PORSAJ TSG: A CASE STUDY
The PORSAJ is an on-line portal, sharing and selling
electronic books, scientific documents, lecture notes,
university examinations, and technical reports in
Persian (http://www.porsaj.com). In this section we
study the proposed statistical method to simulate the
PORSAJ technical support group as a case study. In
this section we introduce the technical support center
of PORSAJ, and then we explain how the data set
for the training, development and evaluation of the
proposed statistical system is created.
4.1 The Technical Support Center
The Internet users consist of amateur visitors in the
sense of smart, surfing through the pages to find
needed information. When using an on-line portal,
the visitors may face some problems which can be
solved by asking a question from the technical staff
of the web site. As it is explained before, a technical
support center can be implemented by a web
interface with the architecture of two-level help desk
which at least a member from the technical support
group is online and ready to answer the questions.
Currently, due to the high value of expenses the
PORSAJ TSG can not work on-line for 24 hours a
day. On the other hand, the PORSAJ visitors are
transitive visitors who are linked from a search
engine when searching for a document, they like to
face an on-line comprehensive help desk. Thus a
help desk equipped with a queuing system is not a
good choice when the users have to wait for the
answer of their questions some hours.
4.2 Data Set
The data set consists of 10,000 questions and
sentences about the technical problems and some
queries to get more information about the services of
the website. These questions are collected from the
help desk of the PORSAJ web site and also by
filling the questionnaire forms by the web site
visitors. The questions and sentences in the data set
can be divided into three categories:
 Communicative messages like “Hello”, “Hi”,
“Good morning”, etc.
 Questions about technical services and
regarding problems.
 Some statements irrelevant to the subject of the
services
These statements being received from the visitors
should be answered by the on-line technical groups
from the web site staff at the first level of the help
desk. Any question or message received from the
visitors is answered by a message from the technical
group. The answers from the technical group are
)
Pr
(
1
I
e
I
e
1
Transformation
Transformation
N
earest
Neighbor
Classifie
r
Language
Model
Lexicon
and
Alignment
Model
Global search
Maximize
)Pr().Pr(
111
IIJ
eef
Over
I
e
1
)Pr(
11
IJ
ef
J
f
1
Use
r
-side question
Des
k
-side answe
r
ICEIS 2008 - International Conference on Enterprise Information Systems
528

reduced to 800 individual answers which are enough
to answer the collected messages from the visitors.
The collected data set is divided into two parts; a
training and an evaluation set. As the proposed
statistical approach needs some parameters of the
model to be trained we split the training set into two
parts, a smaller training set and a development set.
First, the statistical model is trained by the smaller
training set and the parameters of the model are
optimized based on the results obtained by using the
model on the development set. Then, the optimized
parameters are used to train the model by using the
whole samples of the training and development set.
Finally, the resulting model is tested on the
evaluation set. The number of samples in the
training, development, and evaluation set is listed in
the Table 1.
Table 1: The statistics of the data set.
User-side
(questions)
Desk-
side
(answers
)
Train # Sentences 8000 800
# Run.
Words
90423 11096
Vocab. Size 576 186
# Singletons 142 98
Developme
nt
# Sentences 1000 442
# Run.
Words
8937 3820
Vocab. Size 248 82
Evaluation
# Sentences 1000 800
# Run.
Words
9022 11096
Vocab. Size 308 186
5 EXPERIMENTAL RESULTS
In order to evaluate the proposed method, we do the
experiments on the data set which is introduced in
the former section. First, we train the hidden Markov
models for the lexicon and alignment models and
also the language model by using the training part of
the database. Then, the parameters of the HMMs
like time distortion penalty, and language model
scales are optimized on the development set. Finally,
we test the resulting model of the DSS on the
evaluation part of the data set. The results of the
classifier on the development and evaluation set are
listed in Table 2.
Table 2: The error rate of the DSS on the development and
evaluation set.
Sentence error rate
Development 8.2%
Evaluation 12.8%
The results show that the help desk can work
alone with an accepted rate of correct answers.
However, the results rely on training the data
strongly and for some applications which the
vocabulary size of the training set is very big, we
expect to have more sentences in the training part in
order to help the construction of the statistical
model. Nevertheless, if we can not obtain an
acceptable rate of right answers for any other
applications, the model can be used as a DSS to help
the technical support staff instead of working alone.
6 CONCLUSIONS
In this paper, a decision support system is suggested
to be simplified as a machine translation system. In
contrary to the general statistical graphs which suffer
from some constraints like controllable size or
dimensions of the data, The proposed approach
focuses on developing a research strategy for
building a domain-independent statistical model
capable to be used for non-geometric data that are
massive in both size and dimensionality to help
decision makers eventually to improve problem-
solving performance or work alone instead of a
group of experts. We will then apply the proposed
statistical model to concrete a realistic domain to
verify the effectiveness of the model. The
experimental results show that the statistical model
can work with an acceptable rate of correct answers
from the desk side.
ACKNOWLEDGEMENTS
We are very thankful for the Information and
Communication Treasure Company (ICT Co.) and
Shahrood University of Technology (SUT) funding
A NEW STATISTICAL MODEL - To Designing a Decision Support System
529

this research. Also we appreciate the efforts of the
students of the Computer and Information
Technology Department at the SUT.
REFERENCES
Bing Nan Li, Ming Chui Dong and Sam Chao, 2008, On
decision making support in blood bank information
systems, Expert Systems with Applications: An
International Journal, ELSEVIER, ScienceDirect, Vol.
34 , No. 2, pp. 1522-1532.
Ren-Jye Dzeng and Hsin-Yun Lee, 2007, Activity and
value orientated decision support for the development
planning of a theme park, Expert Systems with
Applications: An International Journal, ELSEVIER,
ScienceDirect, Vol. 33, pp. 923-935.
Dorner S., Shi J, Swayne D., 2007, Multi-Objective
Modelling and Decision Support Using a Bayesian
Network Approximation to a Non-point Source
Pollution Model, Environmental Modelling &
Software, Vol. 22, pp. 211-222.
TCS, 2005, TC-STAR - Technology and Corpora for
Speech to Speech Translation, Integrated project TC-
STAR (IST-2002-FP6-506738) funded by the
European
Commission. http://www.tc-star.org/.
Michalewicz, Z.; Schmidt, M.; Michalewicz, M.; Chiriac,
C., 2005, Case study: an intelligent decision support
system, Intelligent Systems, IEEE Vol. 20, Issue 4, pp.
44-49.
Turban, 2004, Decision Support System and Intelligent
Systems, 6
th
Edition, Millan, 2004.
Y. Akiba, M. Federico, N. Kando, H. Nakaiwa, M. Paul,
J. Tsujii, 2004, Overview of the IWSLT04 Evaluation
Campaign. In Proc. of the Int. Workshop on Spoken
Language Translation (IWSLT), pp. 1–12, Kyoto,
Japan.
Leu S. S., and Hwang, S. T., 2002, GA-based resources-
constrained flowshop scheduling model for mixed
precast production. Automation in Cnstruction, Vol.
11, No. 4, pp. 439-452.
Leu S. S., and Hwang, S. T., 2002, GA-based resources-
constrained flowshop scheduling model for mixed
precast production. Automation in Cnstruction, Vol.
11, No. 4, pp. 439-452.
W. Wahlster, 2000, editor: Verbmobil: Foundations of
speech-to-speech translations. Springer Verlag,
Berlin, Germany.
Middleton, I., 1996. Key Factors in Help Desk Success
(An analysis of areas critical to help desk development
and functionality.) In British Library R&D Report
6247, The British Library.
Hutchins, W. John, and Harold L. Somers, 1992. An
Introduction to Machine Translation, London:
Academic Press, Press. ISBN 0-12-362830-X.
Shneiderman, Ben, 1992. Tree Visualization with
TreeMaps: A 2-D Space-filling Approach, ACM
Transaction on Graphics, Vol. 11, No. 1, pp. 92-99.
Robyn M. Dawes, 1988. Rational Choice in an Uncertain
Choice, Hartcourt Brace Jovanovich, Publishers.
Bertin, Jacques, 1983. Semilogy of Graphics. Translated
by William J. Berg, At The University of Wisconsin.
Perlman, G. and T.D. Erickson, 1983. Graphical
Abstractions of Technical Documents, In Visible
Language, Vol. 23, No. 4.
Daniel Kahneman, Paul Slovic, and Amos Tversky, 1982.
Judgment under uncertainty: Heuristics and biases,
Cambridge University Press.
ICEIS 2008 - International Conference on Enterprise Information Systems
530