Vol. 4, No.

8 Aug 2013 ISSN 2079-8407

Journal of Emerging Trends in Computing and Information Sciences
©2009-2013 CIS Journal. All rights reserved.

http://www.cisjournal.org

A Neural-CBR System for Real Property Valuation

1
Adebola G. Musa, 2 Olawande Daramola, 3 Alfred Owoloko, 4 Oludayo Olugbara
1
Dept. of Software Engineering, Tshwane University of Technology, South Africa
2
Dept of Computer Science, Covenant University, Nigeria
3
Dept of Mathematics, Covenant University, Nigeria
4
Dept of Information Technology, Durban Univ. of Tech, South Africa

ABSTRACT
In recent times, the application of artificial intelligence (AI) techniques for real property valuation has been on the
increase. Some expert systems that leveraged on machine intelligence concepts include rule-based reasoning, case-based
reasoning and artificial neural networks. These approaches have proved reliable thus far and in certain cases outperformed
the use of statistical predictive models such as hedonic regression, logistic regression, and discriminant analysis. However,
individual artificial intelligence approaches have their inherent limitations. These limitations hamper the quality of
decision support they proffer when used alone for real property valuation. In this paper, we present a Neural-CBR system
for real property valuation, which is based on a hybrid architecture that combines Artificial Neural Networks and Case-
Based Reasoning techniques. An evaluation of the system was conducted and the experimental results revealed that the
system has higher satisfactory level of performance when compared with individual Artificial Neural Network and Case-
Based Reasoning systems.

Keywords: Case-based reasoning, artificial neural networks, real property valuation, intelligent system

1. INTRODUCTION together in parallel for the purpose of performing a

In recent years, the application of artificial common task. It is a model of computation that emulates
intelligence (AI) approaches for real property valuation the operational principles of the biological nervous
has been on the increase. Expert systems that leverage systems by providing a mathematical equivalent of the
machine intelligence concepts such as rule-based combination of neurons connected in a network. The
reasoning [39, 29], case-based reasoning (CBR) [30], and neurons of an ANN are linked with each other through
artificial neural networks (ANN) [10, 15, 27, 43, 45] have connections. Each connection is assigned a weight that
been used. These AI approaches have been found to controls the flow of information among the neurons.
outperform traditional statistical approaches such as Whenever data is fed into a neuron through the
hedonic regression, logistic regression, and discriminant connections, it is summed up first and then gets
analysis, and very capable to complement the decision transformed by an activation function. The outputs of this
making process [37]. activation function are then sent to other neurons (for feed
forward networks) or back to itself as input (for recurrent
However, these AI approaches have their networks) [35]. ANN has very strong adaptive learning
individual strengths and weaknesses, which inherently ability from which it derives its strong interpolative
affect the quality of performance when used alone for real capability. This makes it very suitable for prediction,
estate valuation. For example, in order to engage a rule- especially in instances of noisy data or incomplete data,
based expert system approach, optimal weights must be which many other alternative prediction models are not
assigned to individual property attributes that are to be able to handle [44]. However, ANN has very weak
used for composing rules of the rule-base, by using explanation mechanism, which makes it difficult to
standardized regression coefficients. However, a major understand the reasoning behind its conclusions [2]. This
challenge of a rule-based system is that these optimal is a major limitation particularly in the real estate domain
weights derived from regression are not generalized, but where it is essential to have a strong rationale for making
rather are location dependent, therefore, the rules and investment decisions.
weights must be updated regularly in order to sustain the
relevance of the system, which is usually a very CBR is an approach that entails the use of the
demanding task [12]. Data mining offers an alternative experience gained in previous problem episodes to arrive
approach to developing intelligent systems for real at a solution for a new problem [1, 21]. It is a machine
property valuation but their viability is only guaranteed learning paradigm that closely models the human
when there is a large pool of transaction data to work with, reasoning process. Solving a problem using CBR involves
which may not exist or may be unreliable in some a number of processes: (1) case matching and retrieval of a
locations [26]. relevant case using defined similarity metrics; (2) case
adaptation for reuse; (3) case revision for appropriateness
The use of Artificial Neural Networks (ANN) for and; (4) case retention in the case base [2, 1]. The nature
the appraisal of real estate property is particularly of CBR, which relates every new episode to similar past
prevalent [44, 2]. ANN could be defined as a group of episodes, makes it very suitable for building intelligent
simple interconnected units, called neurons that function systems with effective explanation mechanism. It has

611
 Vol. 4, No. 8 Aug 2013 ISSN 2079-8407
Journal of Emerging Trends in Computing and Information Sciences
©2009-2013 CIS Journal. All rights reserved.

http://www.cisjournal.org
proved useful for real property valuation in some locations estate values. In [22], a multi-resolution approach was
where transaction data is not readily available or is used to determine real estate price in the Chinese real
unreliable. The past experiences of experts have been used estate market applying three theories: (1) Unascertained
as a basis to implement a CBR system for decision support theory, (2) Principal Component Analysis - PCA and (3)
purposes [32]. CBR systems have very strong explanation Ant Colony Optimization ACO-based ANN. The result
mechanism because of the existence of sufficiently similar forecasted is in good agreement with the actual values, and
previous cases that provides good rationale for new have been very accurate and meet the actual needs. Lin
solutions obtained. However, the disadvantage of CBR is and Chen [23] applied Back Propagation Neural Networks
that the quality of its solutions depends solely on the (BPN) and Support Vector Regression (SVR) to property
existence of good cases that are relevant to solving the valuation in Taiwan. The results of the BPN & SVR were
new case at hand. This brings the tendency to overly rely compared. It was found that SVR with trial-and-error
on previous experiences without validating them in a new method performed the best with MAPE = 4.466% and R2
situation. = 0.8540. That is, stepwise regression is efficient but not
the best variable selection method with both BPN and
Our approach in this work, innovates the SVR. Also ANN was used as a valuation technique in [10,
combination of ANN and CBR in a single system 15, 27, 43, 45, 44, 8, 9, 33].
framework leveraging the strengths of the two instance-
based learning techniques. The experienced-based Most of the existing CBR systems that have been
problem solving capability of CBR systems and its viable reported in literature find application in the fields of
explanation mechanism is combined with the strong medicine, law, planning and design. These include CHEF
interpolative capability of ANN in producing a Hybrid [18], PESUADER [41], CABOT [11], GINA [14]), and
Intelligent System (HIS) for qualitative decision support CYRUS [20]. Relatively few CBR systems have been
for real property valuation. This, to the best of our reported to have application in the real estate domain.
knowledge represents a first attempt at hybridizing these However, PROFIT [6], is a Fuzzy CBR (FCBR) system
two approaches in a practical scenario for improved for residential property valuation. It is an advanced
decision support in the real estate domain. To achieve this, prototype system developed to estimate residential
data from selected input variables of new cases are property values for real estate transactions that was based
transformed via a pre-processing procedure into numerical on the use of CBR techniques with Fuzzy predicates.
data that are suitable for ANN computation, and the result PROFIT has been successfully tested on thousands of real
of the ANN is passed to the CBR component. Thereafter, estate transactions. Also, Pacharavanich et al. [32]
the CBR component seeks for existing past cases that are reported the application of a CBR tool for the valuation of
sufficiently similar to the input case whose solution and residential property in Bangkok, also an evaluation of the
explanation can be adapted to the new context. Hence, CBR tool was conducted. Juan et al. [19] developed a
this work introduces the novel hybridization of ANN and “pre-sale housing”-based decision support system for the
CBR decision support in real property valuation for Taiwan real estate market using a hybrid of CBR and
improved performance relative to the application of a Genetic Algorithm (GA). Based on the customer’s needs,
solitary ANN or CBR approach. CBR is used to retrieve relevant housing layout. Out of the
retrieved cases, nearest neighbour method was used to
The remaining part of the paper is described as calculate similarity of cases. GA was then applied to
follows: In section 2 we give an overview of related work, optimize cost and housing conditions. Hybrid CBR
while Section 3 discusses the hybrid architecture of the systems are those that combine other forms of knowledge
Neural-CBR system. Section 4 is a case study report of the and reasoning methods with CBR. Examples include
application of the Neural-CBR system to business data of Fuzzy-CBR [7], rule-based and case-based [36, 16],
properties sales obtained from a Nigerian company (Dan combining case-based and model-based [34], case-based
Odiete and Co. Ltd. based in Benin City, Nigeria). The and inductive learning [13, 5, 25, 3]. Thus far, relatively
paper is concluded in section 5 with a brief note. few instances of Neural-CBR hybrids have been reported
in the literature with no report of its application to the real
2. RELATED WORK estate domain. Although, Taffasse [42] discussed the
A number of machine learning methods and prospects of Neural-CBR approach to real property
techniques that are applicable to property appraisal and appraisal, the paper did not report any implementation
valuation have been reported in literature. Wilson et al. experience to practically validate the propositions made.
[44] reported the implementation of an intelligent system Specifically, a combination of CBR and a Radial basis
for valuation of residential property. The intelligent ANN in the implementation of a Sales-Advisory system
system was built using a hybrid of Multi-Layer Perceptron was reported in [28].
(MLP) ANN and rule-based expert system. The study by
Guan et al [17] describes the design and implementation of In [24] an ongoing work on the development of
an Adaptive Neuro-Fuzzy Inference System-based hybrid Neural-CBR classifiers for building on-line
(ANFIS) approach to estimate prices for residential communities was reported. The objective of the work is to
properties. The paper represents a first attempt to evaluate identify communities of use in the context of an organized
the feasibility and effectiveness of ANFIS in assessing real group of people. The process involves mining users

612
 Vol. 4, No. 8 Aug 2013 ISSN 2079-8407
Journal of Emerging Trends in Computing and Information Sciences
©2009-2013 CIS Journal. All rights reserved.

http://www.cisjournal.org
bookmark files in order to identify communities that share Where w0 is the bias weight, wi and xi are the
the same information interests. An intelligent agent is used weight and input vectors respectively. The activation at
to observe user behavior in order to learn the user each neuron is given by the sigmoid function:
bookmark classification strategy before hybrid neural
case-based reasoning component is used as incremental 1
classifier. Also, Bajo et al. [4] reported the implementation f(a) = (2)
of a Case-Based Planner for Monitoring Patients (1  e  sum )
(CBPMP) system. It is an autonomous deliberative case-
based planner designed to plan the nurses' working time The sigmoid function is continuous and
dynamically, to maintain the standard working reports differentiable in the interval [0, 1], and is one of mostly
about the nurses' activities, and to guarantee that the used activation function for the MLP. The MLP is trained
patients assigned to the nurses are given the right care. The using the back propagation algorithm [35], which is a form
CBPMP was integrated with a Routing Problems with of supervised learning, by presenting sample input-output
Time Windows (RPTW) neural network component in pairs to the network. The error difference between the
order to realize an intelligent environment for monitoring network’s output and the expected target output are fed
patients' health care in execution time in hospital back into the network for updating the weights connecting
environments. Hence, the contribution of this work stems the hidden-output layers and the input-hidden layers.
from the novel hybridization of MLP-ANN and CBR in
the implementation of a Neural-CBR decision support In our Neural-CBR system, symbolic data
system in real estate property valuation. obtained from case instances are transformed into
numerical data through pre-processing and fed as input
into the MLP. The data pre-processing is therefore, a
3. THE HYBRID NEURAL-CBR SYSTEM
necessary precursor to the operations of the MLP-ANN
The Neural-CBR system is a hybrid modular
component of the Neural-CBR system.
architecture of two components with five user interfaces.
The two components are the ANN component and the
3.2 The CBR Component
CBR component (see Figure 1).
The set of input variable values and the predicted
output obtained from the MLP component is passed to the
CBR component. The CBR component has a case base
indexed on unique case identity field (case_id) and the
computed similarity score of each case. The typical
structure of a case consists of the following:
 Case_id – which is an auto-generated primary
key of the table;
 Case_simscore – which is the computed
similarity score of a case in the case base relative
to a particular case instance;
 Case_attributesset – input values of individual
attributes stored as a string separated by
delimiters;
Fig 1: A Schematic View of the hybrid Neural-CBR  Case_sellingprice – the predicted valuation of a
Architecture case instance;
 Case_weightSet – the set of weights associated
3.1 The Multi-Layer Perceptron ANN Component with each attribute variable such that wi
The multi-layer perceptron (MLP) ANN is a represents the weight of the ith attribute.
powerful neural network model that can be used for
solving approximation, estimation, classification and 3.2.1 Similarity Analysis
prediction problems. Generally, it consists of an input Similarity analysis was done using the nearest
layer, an output layer and one hidden layer. The hidden neighbour algorithm. The similarity measure used was the
layer(s) and the output layer are the processing layers in inverse of weighted normalized Euclidian distance. A
the network where activation takes place. The knowledge similarity score is computed by:
of the network is encoded in the weights connecting the
neurons. Each neuron in an inner layer acts as a linear SIM(X, Y) = 1- DIST(X,Y) (3)
combiner whose summation function is given as:
DIST(X, Y) = (4)
n
Sum = w 
0 w
j 1
i xi (1)
Where X and Y are the new and stored case
respectively with n number of attributes while and
are the normalized values for the ith attribute. A

613
 Vol. 4, No. 8 Aug 2013 ISSN 2079-8407
Journal of Emerging Trends in Computing and Information Sciences
©2009-2013 CIS Journal. All rights reserved.

http://www.cisjournal.org
normalized weight wi is assigned to each attribute based categories, g1, g2, g3..., gk, with their
on contextual experts’ knowledge of the location corresponding Pcbr as P1, P2,.., Pk;
concerned. This calculation is repeated for every stored iii. Take count of the number of cases retrieved in
case in the case base. The cases with the highest similarity each category gi (i = 1...k) and store them as t1,
score (that is up to or above the similarity benchmark t2...tk);
value) are picked as candidates for adaptation in providing iv. Choose category gi with the highest frequency;
explanation for the new case scenario. The algorithm of v. If there exist only one category gi with highest
the Neural-CBR system is given in Figure 2. frequency then take Pcbr of a case in gi as Pvalue
(i.e. final output) else, if there is more than one
category gi with highest frequency, such as gi1,
gi2... gim then compute average value Pave of all
Pim in gi as Pvalue (final output) and as Pcbr for Cnew
.
vi. Next, use descriptions of case attributes
(Case_attributesset) in category gi as explanation
for Cnew (line 9-12).

If cases with similarity score of at least 0.5, but

less than SimBenchmark exists then the Neural-CBR
system uses the neural computed Pneural (ANN predicted
output) in case adaptation thus:

1) Use retrieved cases to retrain ANN-MLP in order

to determine Pneural;
2) Assign Pneural as solution for Pcbr and also as Pvalue
in this case;
3) Use descriptions of attributes of retrieved cases
that are closest to corresponding Cnew variables
for explanation. Where no sufficiently similar
case is found, then restate description attributes
of Cnew for explanation (see line 15-18) i.e. for all
Fig 2: Neural-CBR System’s Operational Procedure retrieved cases Cr = {c1, c2 …cn) where fi
(i=1...15) is a specific attribute feature of Ck  Cr,
3.3 Algorithm of Neural-CBR System fj a corresponding attribute feature of Cnew, wi, the
In this section, we present the algorithm of the weight associated with attribute feature fi,
Neural-CBR systems that details how it reaches its compute SIM(fi, fj) = │wifi-wifj│ and rank cases
conclusions (see Figure 2). Given an input Cnew and the where SIM(fi, fj) ≤ 0.5. Pick as explanation for
existence of the case base CB. A variable Sim Benchmark Cnew highest ranked fi in Cr, for each instance
is set as the minimum acceptable value for sufficient where SIM(fi, fj) ≤ 0.5 is not found use the
similarity between cases. The algorithm selects the description of fj in Cnew as explanation for Cnew.
SimBenchmark value to be in the interval [0.75, 1.0], such end for
that the value 0.75 implies a degree of similarity in the
upper quartile while the value 1.0 denotes perfect If sufficiently similar cases do not exist then the
similarity. Likewise, the value 0.5 connotes an average system takes Pneural as Pvalue. However in this case it
similarity score. Step 4, 5, 6 in the algorithm represents the restates the features of Cnew as explanation for output (line
pre-processing, ANN training and ANN prediction phases 20-21).
of the system’s operation. In Step 7, a one pass scan of the
case base (CB) was carried out to compute similarity 4. THE CASE STUDY
between Cnew and the existing cases in the case base using A prototype system was developed using data
the Weighted Euclidean distance (see equations 1, 2). If collected from a real estate firm (Dan Odiete and Co. Ltd)
cases with similarity score up to or above the in Nigeria [31]. The instance data used in the case study is
SimBenchmark exist (i.e. similarity ≥ SimBenchmark) for the periods of 2002 to 2005, a sample is shown in
then case adaptation is conducted as follows: Table 1. Data associated with fifteen core attribute features
used in the appraisal of residential properties were
extracted and used to train the neural network component
i. Rank all cases found by their similarity score; to yield an estimate of the price of the property. The
ii. Group the retrieved cases based on their solution features, description and range of values obtained from the
values Pcbr (case selling price) into different training set used are given in Table 2.

614
 Vol. 4, No. 8 Aug 2013 ISSN 2079-8407
Journal of Emerging Trends in Computing and Information Sciences
©2009-2013 CIS Journal. All rights reserved.

http://www.cisjournal.org

Table 1: Bedroom Bungalow BQ optional

Table 2: Description of features for real property valuation

4.1 Data-Preprocessing represented as A (available) or B (not available), which

The raw input data (see Table 1) were normalized are assigned vales 1 for A and 0 for B. In the case of
based on min-max normalization [40] to values between 0 neighbourhood_group, A (high), B (medium), or C (low)
and 1 using the data-preprocessing interface of the Neural- are assigned values 1, 0.5 and 0 respectively.
CBR system. The rescaled values of the attribute features
extracted from the data are as shown in Table 3. All the
values are numerical except the following: borehole, BQ,
fenced_round, neighbourhood_group and accessibility.
These five were category inputs, therefore were

615
 Vol. 4, No. 8 Aug 2013 ISSN 2079-8407
Journal of Emerging Trends in Computing and Information Sciences
©2009-2013 CIS Journal. All rights reserved.

http://www.cisjournal.org
Table 3: Showing normalized feature values from Table 1

4.2 Training the MLP-ANN

The configuration of the MLP-ANN model used
for our training instance is a 15-16-1 MLP in which the Table 5: Variation of learning rate
number of core attribute variables (15 of them)
corresponds to the number of input neurons with one
hidden layer containing 16 neurons and 1 neuron in the
output layer, which returns as output the predicted sales
price estimate. The Neural-CBR system environment
allows for the specification of a desired configuration for
the MLP network, which is then dynamically created. The
MLP was trained using the back propagation algorithm
with three sets of data, the training set, the validation set
(to verify correctness during training) and the test set. The
summary of the ANN experiments that was conducted
with the Neural-CBR system framework in order to arrive
at the 15-16-1 MLP configuration and other optimal
training parameters such as number of neurons in the
hidden layer, learning rate and threshold are presented in Table 6: Variation of threshold
Tables 4, 5 and 6. In each occasion recordings were taken
and the average computed to determine the optimum value
in each case.

Table 4: Variation of neurons in the hidden layer of MLP

After the training experiment, the MLP

configuration was 15-16-1, with 1.0 learning rate and
0.005 threshold value. This was used for predicting the
sale price of a property and the predicted result passed to
the CBR component.

4.3 Implementation of the Neural-CBR System

The Borland C++ Builder version 6.0 software
was used as the programming platform to realize the
Neural-CBR system. The case base was implemented as
SQL Server database table of records (cases) indexed on
case_id and the computed weighted Case_simscore fields.
The CBR component does similarity analysis and uses
parameterized SQL statements to determine the best-case

616
 Vol. 4, No. 8 Aug 2013 ISSN 2079-8407
Journal of Emerging Trends in Computing and Information Sciences
©2009-2013 CIS Journal. All rights reserved.

http://www.cisjournal.org
matches. The procedure employed by the Neural-CBR are determined. In our specific case, the output of the
system to reach its final conclusion is based on the Neural-CBR system was validated by using a direct
algorithm shown in figure 2. Figures 3-5 show method [38]. In the direct method of expert system
validation a human expert does a quantitative assessment
of the expert system software by engaging it to perform a
simple benchmark problem over a specified period. The
expert then responds to a set of questions about the
software based on past experience. The questions are
quantitative and are based on a 0 (very false) to 5 (very
true) numerical scales. Thereafter, a single numerical
factor results called the Satisfaction Level that ranges from
0 (least satisfied user) to 5 (most satisfied user) is
computed, which is used to rank the expert system
software in terms of its likelihood to satisfy a prospective
end user.

Fig 3: Feature Extraction Interface of the Neural-CBR 4.4.1 Description of the Evaluation Experiment
System The objective of the validation experiment was to
determine the level of users’ satisfaction with the hybrid
Neural-CBR system relative to CBR alone and ANN alone
systems. In order to do this, a real property valuation
system was created that allows the user to alternate
between three system modalities, Neural-CBR, CBR and
ANN, such that, if one of the modality is activated the
other two modalities are automatically disabled. A
comparative assessment of the three system modalities
was then undertaken using some selected human experts
from the real estate industry for the evaluation. The
participants in the experiment were persons with
significant professional experience in the real property
valuation. Fifteen (15) real estate experts drawn from two
firms, Dan Odiete & Co. and Ajeb Associates (both
Fig 4: Network Training Interface of the Neural-CBR located in Benin City, Nigeria) with varying professional
System experience ranging from 2 (lowest) to 15 (highest) years
were selected to participate in the user-based system
usability experiment. Each received a copy of a
questionnaire instrument and had the system to be
evaluated installed for them. The participants who were
people with tangible knowledge of the use of software
were given one-week training on how to make use the
system prior to the commencement of the experiment.
Details of how to switch modalities and to engage the
system in specific modality for operations such as data
preprocessing, training, and prediction were clearly
outlined. In addition, the participants had the license to
engage the system in as many trial sessions as deemed
convenient commencing with the actual assessment.
Moreover, one IT support staff was temporarily assigned
to each company for the trial period of three weeks. The
Fig 5: Prediction and Explanation Interface of the Neural- human experts were asked to give a rating of 0 to 5 of their
CBR System assessment of the performance of the three (Neural-CBR,
CBR, ANN) system modalities covering seven aspects and
4.4 Performance Evaluation of Neural-CBR System a total of sixteen questions. Each question has specific
Evaluation is the process of determining the weight assigned to it, which had the mutual consent of
appropriateness of a specific system relative to its human experts that participated in the experimental
functional requirements and objectives. Validation of an process. Table 7 shows a sample of a typical response to
expert system is conducted by determining whether the the questionnaire test instrument and how the satisfaction
system's outcome is consistent with the conclusions of the score for an evaluator is computed.
human experts. Validation focuses on evaluating the
outcomes rather than the process by which the outcomes

617
 Vol. 4, No. 8 Aug 2013 ISSN 2079-8407
Journal of Emerging Trends in Computing and Information Sciences
©2009-2013 CIS Journal. All rights reserved.

http://www.cisjournal.org

Table 7: A Typical System Evaluation Template

4.5 Results and Discussion Table 8: Result of Evaluation Experiment

At the end of the evaluation experiment, the mean
satisfaction level as computed from the assessments of the
fifteen real estate expert evaluators for the three systems
are as shown in Table 8. The Neural-CBR system had a
mean score of 3.83/5.0 (viz. 3.83 out of the possible
maximum score of 5.0); CBR system had 3.64/5.0; and
ANN system 3.75/5.0; all which are indicative of good
performance.

However, in order to determine the best of the

three systems; we compared the computed mean
satisfaction level of the three systems to determine
whether the differences in the mean values are statistically
significant. To achieve this we formulated the following
hypothesis:
Hypothesis one

618
 Vol. 4, No. 8 Aug 2013 ISSN 2079-8407
Journal of Emerging Trends in Computing and Information Sciences
©2009-2013 CIS Journal. All rights reserved.

http://www.cisjournal.org
H0: There is no significant difference in the mean
satisfaction level of the CBR, Neural-CBR, and ANN
systems, and hence the three systems are at par
performance wise.

H1: There is significant difference in the mean

satisfaction level of the CBR, Neural-CBR, and ANN
system, and the Neural-CBR systems is better than the
other two systems.
Testing the hypothesis
Generally the observations from the case study
In order to test the hypothesis that was reveal potential benefits of the novel hybridization of
formulated, the Analysis of Variance (ANOVA) statistic ANN and CBR. First, the output of the system shows the
was employed to compare the three sets of data obtained interpolative power of ANN especially in instances where
from the evaluation experiment. The coefficient of other predictive models may be deficient. Second, the
variation (CV) (see Table 10) of the sets of data was results from the system reveal how well the ANN
computed in order to determine the system with the best component could make up for the limitation of the CBR
rating distribution. This is given as: component in instances when there is lack of sufficiently
similar old cases in the case base for a new case. At the
same time the Neural-CBR system leverages the existence
of a case base, in providing justifiable explanation for
results instead of being a black box like a typical ANN.
Where σ = standard deviation of the data distribution, µ = Additionally, the provisioning of rich GUI interfaces for
mean of the data distribution preprocessing and training of ANN enables real-time
acquisition of expert knowledge in the process of solving a
At 1% significance level (i.e. p < 0.01), it was problem such as being able to assign weights to specific
found that the mean satisfaction levels of the three systems real property attribute variables, which makes the system
(ANN, Neural-CBR and CBR) systems are significantly very adaptive and suitable as a decision support tool.
different because the p-value of 0.008305 from the
ANOVA test is less than 0.01 (see Table 9). In addition, 5. CONCLUSION
the coefficient of variation (CV) for the hybrid Neural- In this work, a novel hybridization of ANN and
CBR system is the lowest when compared to the other two CBR techniques for real estate property valuation has been
systems, which is indicative of a relatively better user demonstrated. A prototype Neural-CBR system was
rating. Therefore, H0 is rejected and H1 accepted, which developed and evaluated in a case study to confirm the
states that there is significant difference in the mean viability of the concept. The result obtained revealed that
satisfaction level of ANN, Neural-CBR and CBR systems the Neural-CBR combination offers more acceptable level
and the Neural-CBR system is better than the other two of usability and performance satisfaction relative to
systems. solitary ANN systems and CBR systems. Also, the system
showed significant strength in key areas of weaknesses
Table 9: ANOVA Table for Mean Satisfaction Level usually associated with solitary ANN and CBR intelligent
Comparison systems, which gives merit to the hybridization. In future
work, we intend to investigate the applicability of hybrid
Neural-CBR systems to other business application
domains such as education, health, and finance where the
potential of Neural-CBR is yet to be fully explored.

REFERENCES
[1] Aamodt A and Plaza “Case-based reasoning:
foundational issues, methodological variations, and
system approaches”. AI Communications. 7(1):39-
59. (1994)

[2] Almond N, Lewis O, Jenkins D, Gronow S and

Ware A “Intelligent systems for the valuation of
residential property”. RICS Cutting Edge,
Conference, Dublin 5-6 September. (1997)

[3] Armengol E. and Plaza E. “Integrating induction in

Table 10: Coefficient of Variation for three Systems a case-based reasoned”. Proceedings Second

619
 Vol. 4, No. 8 Aug 2013 ISSN 2079-8407
Journal of Emerging Trends in Computing and Information Sciences
©2009-2013 CIS Journal. All rights reserved.

http://www.cisjournal.org
European Workshop on Case-Based Reasoning: [15] Evans A, James H and Collins A “Artificial Neural
243-252. (1994) Networks: an Application to Residential Valuation
in the UK”. Journal of Property Valuation &
[4] Bajo J, de Paz JF , de Paz Y , Corchado JM Investment 11:195-204. (1993)
“Integrating case-based planning and RPTW neural
networks to construct an intelligent environment for [16] Gayer, G, Gilboa, I. and Lieberman O “Rule-Based
health care”, Expert Systems with Applications and Case-Based Reasoning in Housing Prices”. The
36(3): 5844-5858. (2009) B.E. Journal of Theoretical Economics, Vol. 7,
Issue 1 (2007)
[5] Bamberger, S and Goos K “Integration of case-
based reasoning and inductive learning methods”, [17] Guan, J, Zurada, J and Levitan, A S “An Adaptive
Proceedings First European Workshop on Case- Neuro-Fuzzy Inference System Based Approach to
Based Reasoning, ed. Richter, Wess, Althoff, and Real Estate Property Assessment”. Journal of Real
Maurer, 2:296-300. (1993) Estate Research”. Vol. 30, No. 4. (2008)

[6] Bonissone P and Cheetham W “Fuzzy Case-Based [18] Hammond K “Learning to anticipate and avoid
Reasoning for Residential Property Valuation”. problems through the explanation of failures”.
Handbook on Fuzzy Computing (G 15.1), Oxford Proceedings of Fifth International Conference on
University Press. (1998) Artificial Intelligence:556-560, Philadelphia, PA.,
[7] Bonissone P, Mantaras R “Fuzzy Case-Based Morgan Kaufmann. (1986)
Reasoning Systems”. Handbook of Fuzzy
Computing Section F4.3, Ruspini, Bonissone, [19] Juan, Y, Shih, Shen and Perng, Y “Decision support
Pedrycz (Eds.), Institute of Physics Publishers. for housing customization: A hybrid approach using
(1998) case-based reasoning and genetic algorithm”.
Expert Systems with Applications, Elsevier, Vol.
[8] Borst RA “Artificial Neural Networks: The Next 31, Issue 1, pp 83–93 (2006)
Modeling/Calibration Technology for the
Assessment community?” Property Tax Journal, [20] Koldoner J “Maintaining organization in a dynamic
10(1):69-94. (1991) long-term memory”. Cognitive Science, Vol. 7(4):
281-328. (1983)
[9] Borst RA “A Method for the Valuation of
Residential Properties using Artificial Neural [21] Koldoner J “An introduction to case-based
Networks in Conjunction with Geographical reasoning”. Artificial Intelligence Review 6(1):3-
Information Systems”. IAAO Conference, Dublin. 34. (1992)
(1994)
[22] Li W and Shi H “Applying Unascertained Theory,
[10] Borst R “Artificial neural networks in mass Principal Component Analysis and ACO-based
appraisal”. Journal of Property Tax Assessment & Artificial Neural Networks for Real Estate Price
Administration 1(2):5-15. (1995) Determination”. Journal of Software, Vol. 6, No. 9,
pp. 1672-1679. (2011)
[11] Callan JP, Fawcett TE, Rissland EL “CABOT: An
adaptive approach to case-based Search”, [23] Lin, H and Chen, K “Predicting Price of Taiwan
Proceedings IJCA. (1991) Real Estates By Neural Networks and Support
Vector Regression”. In Proc of the 15th WSEAS
[12] Churbuck DC “Learning by example”. Forbes, international Conference. (2010)
6/8/92, 149(2):130-132. (1992)
[24] Malek M and Kanawati R “A Cooperating Hybrid
[13] Connolly D and Christey S “Learning Neural-CBR Classifiers for Building On-line
representation by integrating case-based and Communities”.
inductive learning”, Proceedings AAAI Case-Based www.aic.nrl.navy.mil/papers/2001/AIC-01-
Reasoning Workshop: 157, Washington, DC. 003/ws5/ws5toc8.PDF. (2009)
(1993)
[25] Manago M, Althoff K., Auriol E, Traphoner R,
[14] De Jong, KA and Schultz AC “Using experience- Wess S, Conruyt N. and Maurer F. “Induction and
based learning in game playing”. Proceedings of reasoning from cases”. Proceedings First European
Fifth International Conference on Machine Workshop on Case-Based Reasoning, ed. Richter,
Learning: 284-290, Ann Arbor, MI, Morgan Wess, Althoff and Maurer 2:313-318. (1993)
Kaufmann. (1988)
[26] McCluskey W and Anand S “The application of
intelligent hybrid techniques for the mass appraisal

620
 Vol. 4, No. 8 Aug 2013 ISSN 2079-8407
Journal of Emerging Trends in Computing and Information Sciences
©2009-2013 CIS Journal. All rights reserved.

http://www.cisjournal.org
of residential properties”. Journal of Property [38] Salim, MD, Villavicencio A and Timmerman MA
Investment and Finance 17(3):218-238. (1999) “A method for evaluating expert system shells for
classroom instruction”. Journal of Industrial
[27] McCluskey W, Dyson K, McFall D & Anand S Technology, Vol. 19, No. 1, pp. 1-11. (2002)
“Mass Appraisal for Property Taxation: An
Artificial Intelligence Approach”. Land Economic [39] Scott I and Gronow S “Valuation expertise: its
Review 2(1):25-32. (1996) nature and application”. Journal of Valuation
8(4):362-375. (1989)
[28] Murray-Smith R and Thakar S “Combining case-
based reasoning with neural networks”. AAAI [40] Shalabi LA, Shaaban Z and Kasasbeh B “Data
Workshop on AI in Service and support, Mining: A Preprocessing Engine”. Journal of
Washington. (1993) Computer Science. 2(9): 735-739. (2006)

[29] Nawawi AH, Jenkins D and Gronow S “Expert [41] Sycara KP “Using case-based reasoning for plan
system development for the mass appraisal of adaptation and repair”. Proceedings of DARPA
commercial property in Malaysia”. Journal of the workshop on case-based reasoning. San Mateo,
Society of Surveying Technicians 18(8): 66-72. Calif. Morgan Kaufmann. (1988)
(1997)
[42] Taffese WZ “Case-based reasoning and neural
[30] O’Roarty B, Patterson D, McGreal WS, Adair AS. networks for real estate valuation”. Proceedings of
“A case based reasoning approach to the selection the 25th IASTED International Multi-Conference:
of comparable evidence for retail rent artificial intelligence and applications, Innsbruck,
determination”. Expert System with Application Austria: 98-104. (2007)
12(4): 417-428. (1997)
[43] Tay D and Ho D “Intelligent Mass Appraisal”.
[31] Owoloko EA “Neural Networks: A data mining Journal of Property Tax Assessment &
technique for property appraisal”. M.Sc. Thesis Administration 1(1): 5-25. (1994)
(Unpublished), University of Benin, Nigeria. (2005)
[44] Wilson ID, Paris SD, Ware JA and Jenkins DH
[32] Pacharavanich P and Wongpinunwatana N “The “Residential property price time series forecasting
Development of a Case-Based Reasoning System as with neural networks”. Knowledge Based Systems
a Tool for Residential Valuation in Bangkok”. Proc. 15(5–6): 335–341. (2002)
of the 6th Annual Pacific-Rim Real Estate Society
Conference:1-14, Sydney. (2000) [45] Worzala E, Lenk, M., & Silva A “An Exploration
of neural networks and its application to real estate
[33] Peterson, S and Flanagan, A B “Neural Network valuation”. Journal of Real Estate Research
Hedonic pricing models in mass real estate 10(2):185-201.
appraisal”. Journal of Real Estate Research, Vol.
31, No. 2, pp 147-164. (2009) AUTHOR PROFILES
[34] Portinale L, Torasso P, Ortalda C and Giardino A Adebola G. Musa is a doctoral
“Using case-based reasoning to focus model-based candidate at Tshwane
diagnostic problem solving”. Proceedings First University of Technology,
European Workshop on Case-Based Reasoning, ed. South Africa. He earns MS and
Richter,Wess, Althoff and Maurer 2:335-340. BS Computer Science at
(1993) Covenant University and
University of Ilorin, Nigeria
[35] Rich E and Knight K “Artificial Intelligence 2nd respectively. His research
ed”. McGraw-Hill, Inc. (1991) interests include the Semantic
Web, Artificial Intelligence and
[36] Rissland E and Skalak D “CABARET: Rule Neural Networks, e-Government. He has to his credit
interpretation in a hybrid architecture”, Int. J. of several research articles in conference proceedings and
Man-Machine Studies, 34(6): 839-887. (1991) refereed journals. He is a member of IITPSA.

[37] Rossini PA, “Accuracy Issues for Automated and

Artificial Intelligent Residential Valuation
Systems”, International Real Estate Society
Conference, Kuala Lumpur, January 26-30. (1999)

621
 Vol. 4, No. 8 Aug 2013 ISSN 2079-8407
Journal of Emerging Trends in Computing and Information Sciences
©2009-2013 CIS Journal. All rights reserved.

http://www.cisjournal.org
Dr. Olawande Daramola is a
Senior Faculty at Department Olugbara O.O. is a
of Computer and Information Professor of Computer
Sciences, Covenant University Science and Information
Nigeria. His research interests Technology at Durban
include the Semantic Web, University of Technology,
Intelligent Systems, and Durban, South Africa.
Automated Software
Engineering.
He is widely published in top range journals and
conferences. He is a member of ACM, IEEE, and the
Nigerian Computer Society.

Owoloko E. Alfred is a
doctoral candidate at
Covenant University, Ota,
Nigeria. He earned his MS
and BS from the University
of Benin, Nigeria. His
research areas include
stochastic differential
equation to valuation,
application of Neural
Networks to solving real life problems, mathematical
statistics and operation research. He has to his credit
several research articles in refereed journals.

622