BUILDING A WEB EFFORT ESTIMATION MODEL THROUGH

KNOWLEDGE ELICITATION

Emilia Mendes

Computer Science Department, The University of Auckland, Auckland, New Zealand

Keywords: Web engineering, Web effort estimation, Expert-based effort models, Knowledge elicitation, Case studies.

Abstract: OBJECTIVE – The objective of this paper is to describe a case study where Bayesian Networks (BNs) were

used to construct an expert-based Web effort model. METHOD – We built a single-company BN model

solely elicited from expert knowledge, where the domain experts were two experienced Web project

managers from a medium-size Web company in Auckland, New Zealand. This model was validated using

data from eleven past finished Web projects. RESULTS – The BN model has to date been successfully used

to estimate effort for numerous Web projects. CONCLUSIONS – Our results suggest that, at least for the

Web Company that participated in this case study, the use of a model that allows the representation of

uncertainty, inherent in effort estimation, can outperform expert-based estimates. Another nine companies

have also benefited from using Bayesian Networks, with very promising results.

1 INTRODUCTION

A cornerstone of Web project management is effort

estimation, the process by which effort is forecasted

and used as basis to predict costs and allocate

resources effectively, so enabling projects to be

delivered on time and within budget. Effort

estimation is a very complex domain where the

relationship between factors is non-deterministic and

has an inherently uncertain nature. E.g. assuming

there is a relationship between development effort

and an application’s size (e.g. number of Web pages,

functionality), it is not necessarily true that increased

effort will lead to larger size. However, as effort

increases so does the probability of larger size.

Effort estimation is a complex domain where

corresponding decisions and predictions require

reasoning with uncertainty.

Within the context of Web effort estimation,

numerous studies investigated the use of effort

prediction techniques. However, to date, only

Mendes (2007a, 2007b, 2007c, 2008), Mendes and

Mosley (2008), and Mendes et al. (2009)

investigated the explicit inclusion, and use, of

uncertainty, inherent to effort estimation, into

models for Web effort estimation. Mendes (2007a,

2007b, 2007c) built a Hybrid Bayesian Network

(BN) model (structure expert-driven and

probabilities data-driven), which presented

significantly superior predictions than the mean- and

median-based effort (Mendes 2007b), multivariate

regression (Mendes 2007a; 2007b; 2007c), case-

based reasoning and classification and regression

trees (Mendes 2007c). Mendes (2008), and Mendes

and Mosley (2008) extended their previous work by

building respectively four and eight BN models

(combinations of Hybrid and data-driven). These

models were not optimised, as previously done in

Mendes (2007a, 2007b, 2007c), which might have

been the reason why they presented significantly

worse accuracy than regression-based models.

Finally, Mendes et al. (2009) details a case study

where a small expert-based Web effort estimation

BN model was successfully used to estimate effort

for projects developed by a small Web company in

Auckland, New Zealand.

A BN is a model that supports reasoning with

uncertainty due to the way in which it incorporates

existing complex domain knowledge (Jensen, 1996).

Herein, knowledge is represented using two parts.

The first, the qualitative part, represents the structure

of a BN as depicted by a directed acyclic graph

(digraph) (see Fig. 1). The digraph’s nodes represent

the relevant variables (factors) in the domain being

modelled, which can be of different types (e.g.

observable or latent, categorical). The digraph’s arcs

represent the causal relationships between variables,

128

Mendes E..

BUILDING A WEB EFFORT ESTIMATION MODEL THROUGH KNOWLEDGE ELICITATION.

DOI: 10.5220/0003562701280135

In Proceedings of the 13th International Conference on Enterprise Information Systems (ICEIS-2011), pages 128-135

ISBN: 978-989-8425-55-3

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

where relationships are quantified probabilistically.

The second, the quantitative part, associates a node

conditional probability table (CPT) to each node, its

probability distribution. A parent node’s CPT

describes the relative probability of each state

(value); a child node’s CPT describes the relative

probability of each state conditional on every

combination of states of its parents (e.g. in Fig. 1,

the relative probability of Total effort (TE) being

‘Low’ conditional on Size (new Web pages)

(SNWP) being ‘Low’ is 0.8). Each column in a CPT

represents a conditional probability distribution and

therefore its values sum up to 1 (Jensen, 1996).



Size (new

Web pages)

Total Effort

Size (total Web

pages)

Child node

Parent node

CPT for node Size (new Web pages)

Low 0.2

Medium 0.3

High 0.5

CPT for node Total Effort (TE)

Size (new Web pages) Low Medium High

Low 0.8 0.2 0.1

Medium 0.1 0.6 0.2

High 0.1 0.2 0.7

Figure 1: A small BN model and two CPTs.

Once a BN is specified, evidence (e.g. values)

can be entered into any node, and probabilities for

the remaining nodes automatically calculated using

Bayes’ rule (Pearl, 1988). Therefore BNs can be

used for different types of reasoning, such as

predictive and “what-if” analyses to investigate the

impact that changes on some nodes have on others

(Fenton et al. 2004).

Within the context of Web effort estimation there

are issues with building data-driven or hybrid

Bayesian models, as follows:

1. Any dataset used to build a BN model should be

large enough to provide sufficient data capturing all

(or most) relevant combinations of states amongst

variables such that probabilities can be learnt from

data, rather than elicited manually. Under such

circumstance, it is very unlikely that the dataset

would contain project data volunteered by only a

single company (single-company dataset). As far as

we know, the largest dataset of Web projects

available is the Tukutuku dataset (195 projects)

(Mendes et al., 2005). This dataset has been used to

build data-driven and hybrid BN models; however

results have not been encouraging overall, and we

believe one of the reasons is due to the small size of

this dataset.

2. Even when a large dataset is available, the next

issue relates to the set of variables part of the

dataset. It is unlikely that the variables identified,

represent all the factors within a given domain (e.g.

Web effort estimation) that are important for

companies that are to use the data-driven or hybrid

model created using this dataset. This was the case

with the Tukutuku dataset, even though the selection

of which variables to use had been informed by two

surveys (Mendes et al., 2005). However, one could

argue that if the model being created is hybrid, then

new variables (factors) can be added to, and existing

variables can be removed from the model. The

problem is that every new variable added to the

model represents a set of probabilities that need to

be elicited from scratch, which may be a hugely time

consuming task.

3. Different structure and probability learning

algorithms can lead to different prediction accuracy

(Mendes and Mosley, 2008); therefore one may need

to use different models and compare their accuracy,

which may also be a very time consuming task.

4. When using a hybrid model, the BN’s structure

should ideally be jointly elicited by more than one

domain expert, preferably from more than one

company, otherwise the model built may not be

general enough to cater for a wide range of

companies (Mendes and Mosley, 2008). There are

situations, however, where it is not feasible to have

several experts from different companies

cooperatively working on a single BN structure. One

such situation is when the companies involved are

all consulting companies potentially sharing the

same market. This was the case within the context of

this research.

5. Ideally the probabilities used by the data-driven

or hybrid models should be revisited by at least one

domain expert, once they have been automatically

learnt using the learning algorithms available in BN

tools. However, depending on the complexity of the

BN model, this may represent having to check

thousands of probabilities, which may not be

feasible. One way to alleviate this problem is to add

additional factors to the BN model in order to reduce

the number of causal relationships reaching child

nodes; however, all probabilities for the additional

factors would still need to be elicited from domain

experts.

BUILDING A WEB EFFORT ESTIMATION MODEL THROUGH KNOWLEDGE ELICITATION

129

6. The choice of variable discretisation, structure

learning algorithms, parameter estimation

algorithms, and the number of categories used in the

discretisation all affect the accuracy of the results

and there are no clear-cut guidelines on what would

be the best choice to employ. It may simply be

dependent on the dataset being used, the amount of

data available, and trial and error to find the best

solution (Mendes and Mosley, 2008).

Therefore, given the abovementioned constraints, as

part of a NZ-government-funded project on using

Bayesian Networks to Web effort estimation, we

decided to develop several expert-based company-

specific Web effort BN models, with the

participation of numerous local Web companies in

the Auckland region, New Zealand. The

development and successful deployment of one of

these models is the subject and contribution of this

paper. The model detailed herein, as will be

described later on, is a large model containing 37

factors and over 40 causal relationships. This model

is much more complex than the one presented in

(Mendes et al., 2009), where an expert-based Web

effort estimation model is described, comprising 15

factors and 14 causal relationships. This is the first

time that a study in either Web or Software

Engineering describes the creation and use of a large

expert-based BN model. In addition, we also believe

that our contribution goes beyond the area of Web

engineering given that the process presented herein

can also be used to build BN models for non-Web

companies.

Note that we are not suggesting that data-driven

and hybrid BN models should not be used. On the

contrary, they have been successfully employed in

numerous domains (Woodberry et al., 2004);

however the specific domain context of this paper –

that of Web effort estimation, provides other

challenges (described above) that lead to the

development of solely expert-driven BN models.

We would also like to point out that in our view

Web and software development differ in a number

of areas, such as: Application Characteristics,

Primary Technologies Used, Approach to Quality

Delivered, Development Process Drivers,

Availability of the Application, Customers

(Stakeholders), Update Rate (Maintenance Cycles),

People Involved in Development, Architecture and

Network, Disciplines Involved, Legal, Social, and

Ethical Issues, and Information Structuring and

Design. A detailed discussion on this issue is

provided in (Mendes et al. 2005).

The remainder of the paper is organised as

follows: Section 2 provides a description of the

overall process used to build and validate BNs;

Section 3 details this process, focusing on the

expert-based Web Effort BN focus of this paper.

Finally, conclusions and comments on future work

are given in Section 4.

2 GENERAL PROCESS USED TO

BUILD BNS

The BN presented in this paper was built and

validated using an adaptation of the Knowledge

Engineering of Bayesian Networks (KEBN) process

proposed in (Woodberry et al., 2004). Within the

context of this paper the author was the KE, and two

Web project managers from a well-established Web

company in Auckland were the DEs.

The three main steps within the adapted KEBN

process are the Structural Development, Parameter

Estimation, and Model Validation. This process

iterates over these steps until a complete BN is built

and validated. Each of these three steps is detailed

below:

Structural Development

: This step represents the

qualitative component of a BN, which results in a

graphical structure comprised of, in our case, the

factors (nodes, variables) and causal relationships

identified as fundamental for effort estimation of

Web projects. In addition to identifying variables,

their types (e.g. query variable, evidence variable)

and causal relationships, this step also comprises the

identification of the states (values) that each variable

should take, and if they are discrete or continuous. In

practice, currently available BN tools require that

continuous variables be discretised by converting

them into multinomial variables, also the case with

the BN software used in this study. The BN’s

structure is refined through an iterative process. This

structure construction process has been validated in

previous studies (Druzdzel and van der Gaag, 2000,

Fenton et al., 2004, Mahoney and Laskey, 1996;

Neil et al., 2000, Woodberry et al., 2004) and uses

the principles of problem solving employed in data

modelling and software development (Studer et al.,

1998). As will be detailed later, existing literature in

Web effort estimation, and knowledge from the

domain expert were employed to elicit the Web

effort BN’s structure. Throughout this step the

knowledge engineer(s) also evaluate(s) the structure

of the BN, done in two stages. The first entails

checking whether: variables and their values have a

clear meaning; all relevant variables have been

included; variables are named conveniently; all

ICEIS 2011 - 13th International Conference on Enterprise Information Systems

130

states are appropriate (exhaustive and exclusive); a

check for any states that can be combined. The

second stage entails reviewing the BN’s graph

structure (causal structure) to ensure that any

identified d-separation dependencies comply with

the types of variables used and causality

assumptions. D-separation dependencies are used to

identify variables influenced by evidence coming

from other variables in the BN (Jensen, 1996; Pearl,

1988). Once the BN structure is assumed to be close

to final knowledge engineers may still need to

optimise this structure to reduce the number of

probabilities that need to be elicited or learnt for the

network. If optimisation is needed, techniques that

change the causal structure (e.g. divorcing (Jensen,

1996)) are employed.

Parameter Estimation

: This step represents the

quantitative component of a BN, where conditional

probabilities corresponding to the quantification of

the relationships between variables (Jensen, 1996;

Pearl, 1988) are obtained. Such probabilities can be

attained via Expert Elicitation, automatically from

data, from existing literature, or using a combination

of these. When probabilities are elicited from

scratch, or even if they only need to be revisited, this

step can be very time consuming. In order to

minimise the number of probabilities to be elicited

some techniques have been proposed in the literature

(Das, 2004; Druzdzel and van der Gaag, 2000; Tang

and McCabe, 2007); however, as far as we are

aware, there is no empirical evidence to date

comparing their effectiveness for prediction,

compared to probabilities elicited from scratch,

using large and realistic BNs. This is one of the

topics of our future work.

Model Validation

: This step validates the BN that

results from the two previous steps, and determines

whether it is necessary to re-visit any of those steps.

Two different validation methods are generally used

- Model Walkthrough and Predictive Accuracy.

Model walkthrough represents the use of real case

scenarios that are prepared and used by domain

experts to assess if the predictions provided by a BN

correspond to the predictions experts would have

chosen based on their own expertise. Success is

measured as the frequency with which the BN’s

predicted value for a target variable (e.g. quality,

effort) that has the highest probability corresponds to

the experts’ own assessment.

Predictive Accuracy uses past data (e.g. past

project data), rather than scenarios, to obtain

predictions. Data (evidence) is entered on the BN

model, and success is measured as the frequency

with which the BN’s predicted value for a target

variable (e.g. quality, effort) that has the highest

probability corresponds to the actual past data.

However, previous literature also documents a

different measure of success, proposed by

Pendharkar et al. (2005), and later used by Mendes

(2007a, 2007c), and Mendes and Mosley (2009).

This was the measure employed herein.

3 PROCESS USED TO BUILD

THE EXPERT-BASED BN

This Section revisits the adapted KEBN process,

detailing the tasks carried out for each of the three

main steps that form part of that process. Before

starting the elicitation of the Web effort BN model,

the Domain Experts (DEs) participating were

presented with an overview of Bayesian Network

models, and examples of “what-if” scenarios using a

made-up BN. This, we believe, facilitated the entire

process as the use of an example, and the brief

explanation of each of the steps in the KEBN

process, provided a concrete understanding of what

to expect. We also made it clear that the knowledge

Engineers were facilitators of the process, and that

the Web company’s commitment was paramount for

the success of the process. The entire process took

54 person hours to be completed, corresponding to

nine 3-hour slots.

The DEs who took part in this case study were

project managers of a well-established Web

company in Auckland (New Zealand). The company

had ~20 employees, and branches overseas. The

project managers had each worked in Web

development for more than 10 years. In addition,

this company developed a wide range of Web

applications, from static & multimedia-like to very

large e-commerce solutions. They also used a wide

range of Web technologies, thus enabling the

development of Web 2.0 applications. Previous to

using the BN model created, the effort estimates

provided to clients would deviate from actual effort

within the range of 20% to 60%.

Detailed Structural Development and Parameter

Estimation: In order to identify the fundamental

factors that the DEs took into account when

preparing a project quote we used the set of

variables from the Tukutuku dataset (Mendes et al.,

2005) as a starting point. We first sketched them out

on a white board, each one inside an oval shape, and

then explained what each one meant within the

context of the Tukutuku project. Our previous

BUILDING A WEB EFFORT ESTIMATION MODEL THROUGH KNOWLEDGE ELICITATION

131

experience eliciting BNs in other domains (e.g.

ecology) suggested that it was best to start with a

few factors (even if they were not to be reused by

the DE), rather than to use a “blank canvas” as a

starting point. Once the Tukutuku variables had been

sketched out and explained, the next step was to

remove all variables that were not relevant for the

DEs, followed by adding to the white board any

additional variables (factors) suggested by them. We

also documented descriptions for each of the factors

suggested. Next, we identified the states that each

factor would take. All states were discrete.

Whenever a factor represented a measure of effort

(e.g. Total effort), we also documented the effort

range corresponding to each state, to avoid any

future ambiguity. For example, ‘very low’ Total

effort corresponded to 4+ to 10 person hours, etc.

Once all states were identified and documented, it

was time to elicit the cause and effect relationships.

As a starting point to this task we used a simple

medical example from (Jensen, 1996) (see Figure 2).

This example clearly introduces one of the most

important points to consider when identifying cause

and effect relationships – timeline of events. If

smoking is to be a cause of lung cancer, it is

important that the cause precedes the effect. This

may sound obvious with regard to the example used;

however, it is our view that the use of this simple

example significantly helped the DEs understand the

notion of cause and effect, and how this related to

Web effort estimation and the BN being elicited.

Figure 2: A small example of a cause & effect

relationship.

Once the cause and effect relationships were

identified, the original BN structure needed to be

simplified in order to reduce the number of

probabilities to be elicited. New nodes were

suggested by the KE (names ending in ‘_O’), and

validated by the DEs. The DEs also made a few

more changes to some of the relationships. At this

point the DEs seemed happy with the BN’s causal

structure and the work on eliciting the probabilities

was initiated. All probabilities were created from

scratch, and the probabilities elicitation took ~24

hours. While entering the probabilities, the DEs

decided to re-visit the BN’s causal structure after

revisiting their effort estimation process; therefore a

new iteration of the Structural Development step

took place. The final BN causal structure is shown in

Figure 3. Here we present the BN using belief bars

rather than labelled factors, so readers can see the

probabilities that were elicited. Note that this BN

corresponds to the current model being used by the

Web company (also validated, to be detailed next).

Detailed Model Validation

: Both Model

walkthrough and Predictive accuracy were used to

validate the Web Effort BN model, where the former

was the first type of validation to be employed. The

DEs used four different scenarios to check whether

the node Total_effort would provide the highest

probability to the effort state that corresponded to

the DEs’ own suggestions. All scenarios were run

successfully; however it was also necessary to use

data from past projects, for which total effort was

known, in order to check the model’s calibration. A

validation set containing data on 11 projects was

used. The DEs selected a range of projects

presenting different sizes and levels of complexity,

where all 11 projects were representative of the

types of projects developed by the Web company:

five were small projects; two were medium, two

large, and one very large.

For each project, evidence was entered in the BN

model, and the effort range corresponding to the

highest probability provided for ‘Total Effort’ was

compared to that project’s actual effort (see an

example in Figure 4). The company had also defined

the range of effort values associated with each of the

categories used to measure ‘Total Effort’. In the case

of the company described herein, Medium effort

corresponds to 25 to 40 person hours. Whenever

actual effort did not fall within the effort range

associated with the category with the highest

probability, there was a mismatch; this meant that

some probabilities needed to be adjusted. In order to

know which nodes to target first we used a

Sensitivity Analysis report, which provided the

effect of each parent node upon a given query node.

Within our context, the query node was ‘Total

Effort’.

Whenever probabilities were adjusted, we re-

entered the evidence for each of the projects in the

validation set that had already been used in the

validation step to ensure that the calibration already

carried out had not affected. This was done to ensure

that each calibration would always be an improved

upon the previous one. Within the scope of the

model presented herein, of the 11 projects used for

validation, only one required the model to be re-

calibrated. This means that for all the 10 projects

remaining, the BN model presented the highest

ICEIS 2011 - 13th International Conference on Enterprise Information Systems

132

probability to the effort range that contained the

actual effort for the project being used for validation.

Once all 11 projects were used to validate the model

the DEs assumed that the Validation step was

complete.

The BN model was completed in September

2009, and has been successfully used to estimate

effort for new projects developed by the company.

In addition, the two DEs changed their approach to

estimating effort as follows: prior to using the BN

model, these DEs had to elicit requirements using

very short meetings with clients, given that these

clients assumed that short meetings were enough in

order to understand what the applications needed to

provide once delivered. The DEs were also not fully

aware of the factors that they subjectively took into

account when preparing an effort estimate; therefore

many times they ended up providing unrealistic

estimates to clients.

Once the BN model was validated, the DEs

started to use the model not only for obtaining better

estimates than the ones previously prepared by

subjective means, but also as means to guide their

requirements elicitation meetings with prospective

clients. They focused their questions targeting at

obtaining evidence to be entered in the model as the

requirements meetings took place; by doing so they

basically had effort estimates that were practically

ready to use for costing the projects, even when

meeting with clients had short durations. Such

change in approach provided extremely beneficial to

the company given that all estimates provided using

the model turned out to be more accurate on average

than the ones previously obtained by subjective

means.

Clients were not presented the model due to its

complexity; however by entering evidence while a

requirements elicitation meeting took place enabled

the DEs to optimize their elicitation process by

being focused and factor-driven.

We believe that the successful development of

this Web effort BN model was greatly influenced by

the commitment of the company, and also by the

DEs’ experience estimating effort.

4 CONCLUSIONS

This paper has presented a case study where a

Bayesian Model for Web effort estimation was built

using solely knowledge of two Domain Experts from

a well-established Web company in Auckland, New

Zealand.

Figure 3: Final expert-based Web effort BN model.

BUILDING A WEB EFFORT ESTIMATION MODEL THROUGH KNOWLEDGE ELICITATION

133

Figure 4: Example of evidence being entered in the Web effort BN model.

This model was developed using an adaptation of

the knowledge engineering for Bayesian Networks

process. Its causal structure went through three

versions, because as the work progressed the

experts’ views on which factors were fundamental

when they estimated effort also matured. Each

session with the DEs lasted for no longer than 3

hours. The final BN model was calibrated using data

on eleven past projects. These projects represented

typical projects developed by the company, and

believed by the experts to provide enough data for

model calibration.

Since the model’s adoption, it has been

successfully used to provide effort quotes for the

new Web projects managed by the company.

The entire process used to build and validate the

BN model took 54 person hours, where the largest

amount of time was spent eliciting the probabilities.

This is an issue to those building BN models from

domain expertise only, and is currently the focus of

our future work.

The elicitation process enables experts to think

deeply about their effort estimation process and the

factors taken into account during that process, which

in itself is already advantageous to a company. This

has been pointed out to us not only by the domain

experts whose model is presented herein, but also by

other companies with which we worked on model

elicitations.

To date we have completed the elicitation of six

expert-driven Bayesian Models for Web effort

estimation and have merged their causal structures in

order to identify common Web effort predictors, and

causal relationships (Baker and Mendes, 2010).

ACKNOWLEDGEMENTS

We thank the Web company who participated in this

case study, and also all the participating companies in

this research. This work was sponsored by the Royal

Society of New Zealand (Marsden research grant 06-

UOA-201).

REFERENCES

Baker, S., and Mendes, E.: Aggregating Expert-driven

causal maps for Web Effort Estimation, Proceedings

of the International Conference on Advanced Software

Engineering & Its Applications (accepted for

publication). (2010).

Das, B.: Generating Conditional Probabilities for Bayesian

Networks: Easing the Knowledge Acquisition

Problem, arxiv.org/pdf/cs/0411034v1 (accessed in

2008). (2004)

Druzdzel, M. J., and van der Gaag, L. C.: Building

Probabilistic Networks: Where Do the Numbers Come

From? IEEE Trans. on Knowledge and Data

Engineering, 12(4), 481-486. (2000)

Fenton, N., Marsh, W., Neil, M., Cates, P., Forey, S. and

Tailor, M.: Making Resource Decisions for Software

Projects, Proc. ICSE’04, pp. 397-406. (2004)

Jensen, F. V.: An introduction to Bayesian networks. UCL

Press, London. (1996)

ICEIS 2011 - 13th International Conference on Enterprise Information Systems

134

Mahoney, S. M., and Laskey, K. B.: Network Engineering

for Complex Belief Networks, Proc. Twelfth Annual

Conference on Uncertainty in Artificial Intelligence,

pp. 389-396. (1996).

Mendes, E.: Predicting Web Development Effort Using a

Bayesian Network, Proceedings of EASE'07, pp. 83-

93. (2007).

Mendes, E.: The Use of a Bayesian Network for Web

Effort Estimation, Proceedings of International

Conference on Web Engineering, pp. 90-104, LNCS

4607. (2007).

Mendes, E.: A Comparison of Techniques for Web Effort

Estimation, Proceedings of the ACM/IEEE

International Symposium on Empirical Software

Engineering, pp. 334-343. (2007).

Mendes, E.: The Use of Bayesian Networks for Web

Effort Estimation: Further Investigation, Proceedings

of ICWE’08, pp. 2-3-216. (2008).

Mendes, E., and Mosley, N.: Bayesian Network Models

for Web Effort Prediction: a Comparative Study,

Transactions on Software Engineering, Vol. 34, Issue:

6, Nov/Dec 2008, pp. 723-737. (2008).

Mendes, E., Mosley, N. and Counsell, S.: The Need for

Web Engineering: An Introduction, Web Engineering,

Springer-Verlag, Mendes, E. and Mosley, N. (Eds.)

ISBN: 3-540-28196-7, pp. 1-28. (2005).

Mendes, E., Mosley, N., and Counsell, S.: Investigating

Web Size Metrics for Early Web Cost Estimation,

Journal of Systems and Software, 77(2), 157-172.

(2005).

Mendes, E., Pollino, C., and Mosley, N.: Building an

Expert-based Web Effort Estimation Model using

Bayesian Networks, Proceedings of the EASE

Conference, pp. 1-10. (2009).

Neil, M., Fenton, N., and Nielsen, L.: Building Large-

scale bayesian networks, The knowledge Engineering

Review, KER, Vol. 15, No. 3, 257-284. (2000).

Pearl J.: Probabilistic Reasoning in Intelligent Systems,

Morgan Kaufmann, San Mateo, CA. (1988).

Pendharkar, P. C., Subramanian, G. H., and Rodger, J.A.:

A Probabilistic Model for Predicting Software

Development Effort, IEEE Trans. Software Eng. Vol.

31, No. 7, 615-624. (2005)

Studer, R., Benjamins, V. R. and Fensel, D.: Knowledge

engineering: principles and methods. Data &

Knowledge Engineering, vol. 25, 161-197. (1998).

Tang, Z., and B. McCabe: Developing Complete

Conditional Probability Tables from Fractional Data

for Bayesian Belief Networks, Journal of Computing

in Civil Engineering, 21(4), pp. 265-276. (2007).

Woodberry, O., A. Nicholson, K. Korb, and C. Pollino:

Parameterising Bayesian Networks, Proc. Australian

Conference on Artificial Intelligence, pp. 1101-1107.

(2004).

BUILDING A WEB EFFORT ESTIMATION MODEL THROUGH KNOWLEDGE ELICITATION

135