journal of computing::a neural-cbr system for real...

Vol. 4, No. 8 Aug 2013 ISSN 2079-8407Journal of Emerging Trends in Computing and Information Sciences

©2009-2013 CIS Journal. All rights reserved.

http://www.cisjournal.org

611

A Neural-CBR System for Real Property Valuation1 Adebola G. Musa, 2 Olawande Daramola, 3 Alfred Owoloko, 4 Oludayo Olugbara

1 Dept. of Software Engineering, Tshwane University of Technology, South Africa2 Dept of Computer Science, Covenant University, Nigeria

3 Dept of Mathematics, Covenant University, Nigeria4 Dept of Information Technology, Durban Univ. of Tech, South Africa

ABSTRACTIn recent times, the application of artificial intelligence (AI) techniques for real property valuation has been on theincrease. Some expert systems that leveraged on machine intelligence concepts include rule-based reasoning, case-basedreasoning and artificial neural networks. These approaches have proved reliable thus far and in certain cases outperformedthe 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 ofdecision support they proffer when used alone for real property valuation. In this paper, we present a Neural-CBR systemfor 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 thesystem 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. INTRODUCTIONIn recent years, the application of artificial

intelligence (AI) approaches for real property valuationhas been on the increase. Expert systems that leveragemachine intelligence concepts such as rule-basedreasoning [39, 29], case-based reasoning (CBR) [30], andartificial neural networks (ANN) [10, 15, 27, 43, 45] havebeen used. These AI approaches have been found tooutperform traditional statistical approaches such ashedonic regression, logistic regression, and discriminantanalysis, and very capable to complement the decisionmaking process [37].

However, these AI approaches have theirindividual strengths and weaknesses, which inherentlyaffect the quality of performance when used alone for realestate valuation. For example, in order to engage a rule-based expert system approach, optimal weights must beassigned to individual property attributes that are to beused for composing rules of the rule-base, by usingstandardized regression coefficients. However, a majorchallenge of a rule-based system is that these optimalweights derived from regression are not generalized, butrather are location dependent, therefore, the rules andweights must be updated regularly in order to sustain therelevance of the system, which is usually a verydemanding task [12]. Data mining offers an alternativeapproach to developing intelligent systems for realproperty valuation but their viability is only guaranteedwhen there is a large pool of transaction data to work with,which may not exist or may be unreliable in somelocations [26].

The use of Artificial Neural Networks (ANN) forthe appraisal of real estate property is particularlyprevalent [44, 2]. ANN could be defined as a group ofsimple interconnected units, called neurons that function

together in parallel for the purpose of performing acommon task. It is a model of computation that emulatesthe operational principles of the biological nervoussystems by providing a mathematical equivalent of thecombination of neurons connected in a network. Theneurons of an ANN are linked with each other throughconnections. Each connection is assigned a weight thatcontrols the flow of information among the neurons.Whenever data is fed into a neuron through theconnections, it is summed up first and then getstransformed by an activation function. The outputs of thisactivation function are then sent to other neurons (for feedforward networks) or back to itself as input (for recurrentnetworks) [35]. ANN has very strong adaptive learningability from which it derives its strong interpolativecapability. This makes it very suitable for prediction,especially in instances of noisy data or incomplete data,which many other alternative prediction models are notable to handle [44]. However, ANN has very weakexplanation mechanism, which makes it difficult tounderstand the reasoning behind its conclusions [2]. Thisis a major limitation particularly in the real estate domainwhere it is essential to have a strong rationale for makinginvestment decisions.

CBR is an approach that entails the use of theexperience gained in previous problem episodes to arriveat a solution for a new problem [1, 21]. It is a machinelearning paradigm that closely models the humanreasoning process. Solving a problem using CBR involvesa number of processes: (1) case matching and retrieval of arelevant case using defined similarity metrics; (2) caseadaptation for reuse; (3) case revision for appropriatenessand; (4) case retention in the case base [2, 1]. The natureof CBR, which relates every new episode to similar pastepisodes, makes it very suitable for building intelligentsystems with effective explanation mechanism. It has




612

proved useful for real property valuation in some locationswhere transaction data is not readily available or isunreliable. The past experiences of experts have been usedas a basis to implement a CBR system for decision supportpurposes [32]. CBR systems have very strong explanationmechanism because of the existence of sufficiently similarprevious cases that provides good rationale for newsolutions obtained. However, the disadvantage of CBR isthat the quality of its solutions depends solely on theexistence of good cases that are relevant to solving thenew case at hand. This brings the tendency to overly relyon previous experiences without validating them in a newsituation.

Our approach in this work, innovates thecombination of ANN and CBR in a single systemframework leveraging the strengths of the two instance-based learning techniques. The experienced-basedproblem solving capability of CBR systems and its viableexplanation mechanism is combined with the stronginterpolative capability of ANN in producing a HybridIntelligent System (HIS) for qualitative decision supportfor real property valuation. This, to the best of ourknowledge represents a first attempt at hybridizing thesetwo approaches in a practical scenario for improveddecision support in the real estate domain. To achieve this,data from selected input variables of new cases aretransformed via a pre-processing procedure into numericaldata that are suitable for ANN computation, and the resultof the ANN is passed to the CBR component. Thereafter,the CBR component seeks for existing past cases that aresufficiently similar to the input case whose solution andexplanation can be adapted to the new context. Hence,this work introduces the novel hybridization of ANN andCBR decision support in real property valuation forimproved performance relative to the application of asolitary ANN or CBR approach.

The remaining part of the paper is described asfollows: In section 2 we give an overview of related work,while Section 3 discusses the hybrid architecture of theNeural-CBR system. Section 4 is a case study report of theapplication of the Neural-CBR system to business data ofproperties sales obtained from a Nigerian company (DanOdiete and Co. Ltd. based in Benin City, Nigeria). Thepaper is concluded in section 5 with a brief note.

2. RELATED WORKA number of machine learning methods and

techniques that are applicable to property appraisal andvaluation have been reported in literature. Wilson et al.[44] reported the implementation of an intelligent systemfor valuation of residential property. The intelligentsystem was built using a hybrid of Multi-Layer Perceptron(MLP) ANN and rule-based expert system. The study byGuan et al [17] describes the design and implementation ofan Adaptive Neuro-Fuzzy Inference System-based(ANFIS) approach to estimate prices for residentialproperties. The paper represents a first attempt to evaluatethe feasibility and effectiveness of ANFIS in assessing real

estate values. In [22], a multi-resolution approach wasused to determine real estate price in the Chinese realestate market applying three theories: (1) Unascertainedtheory, (2) Principal Component Analysis - PCA and (3)Ant Colony Optimization ACO-based ANN. The resultforecasted is in good agreement with the actual values, andhave been very accurate and meet the actual needs. Linand Chen [23] applied Back Propagation Neural Networks(BPN) and Support Vector Regression (SVR) to propertyvaluation in Taiwan. The results of the BPN & SVR werecompared. It was found that SVR with trial-and-errormethod performed the best with MAPE = 4.466% and R2= 0.8540. That is, stepwise regression is efficient but notthe best variable selection method with both BPN andSVR. Also ANN was used as a valuation technique in [10,15, 27, 43, 45, 44, 8, 9, 33].

Most of the existing CBR systems that have beenreported in literature find application in the fields ofmedicine, law, planning and design. These include CHEF[18], PESUADER [41], CABOT [11], GINA [14]), andCYRUS [20]. Relatively few CBR systems have beenreported to have application in the real estate domain.However, PROFIT [6], is a Fuzzy CBR (FCBR) systemfor residential property valuation. It is an advancedprototype system developed to estimate residentialproperty values for real estate transactions that was basedon the use of CBR techniques with Fuzzy predicates.PROFIT has been successfully tested on thousands of realestate transactions. Also, Pacharavanich et al. [32]reported the application of a CBR tool for the valuation ofresidential property in Bangkok, also an evaluation of theCBR tool was conducted. Juan et al. [19] developed a“pre-sale housing”-based decision support system for theTaiwan real estate market using a hybrid of CBR andGenetic Algorithm (GA). Based on the customer’s needs,CBR is used to retrieve relevant housing layout. Out of theretrieved cases, nearest neighbour method was used tocalculate similarity of cases. GA was then applied tooptimize cost and housing conditions. Hybrid CBRsystems are those that combine other forms of knowledgeand reasoning methods with CBR. Examples includeFuzzy-CBR [7], rule-based and case-based [36, 16],combining case-based and model-based [34], case-basedand inductive learning [13, 5, 25, 3]. Thus far, relativelyfew instances of Neural-CBR hybrids have been reportedin the literature with no report of its application to the realestate domain. Although, Taffasse [42] discussed theprospects of Neural-CBR approach to real propertyappraisal, the paper did not report any implementationexperience to practically validate the propositions made.Specifically, a combination of CBR and a Radial basisANN in the implementation of a Sales-Advisory systemwas reported in [28].

In [24] an ongoing work on the development ofhybrid Neural-CBR classifiers for building on-linecommunities was reported. The objective of the work is toidentify communities of use in the context of an organizedgroup of people. The process involves mining users




613

bookmark files in order to identify communities that sharethe same information interests. An intelligent agent is usedto observe user behavior in order to learn the userbookmark classification strategy before hybrid neuralcase-based reasoning component is used as incrementalclassifier. Also, Bajo et al. [4] reported the implementationof a Case-Based Planner for Monitoring Patients(CBPMP) system. It is an autonomous deliberative case-based planner designed to plan the nurses' working timedynamically, to maintain the standard working reportsabout the nurses' activities, and to guarantee that thepatients assigned to the nurses are given the right care. TheCBPMP was integrated with a Routing Problems withTime Windows (RPTW) neural network component inorder to realize an intelligent environment for monitoringpatients' health care in execution time in hospitalenvironments. Hence, the contribution of this work stemsfrom the novel hybridization of MLP-ANN and CBR inthe implementation of a Neural-CBR decision supportsystem in real estate property valuation.

3. THE HYBRID NEURAL-CBR SYSTEMThe Neural-CBR system is a hybrid modular

architecture of two components with five user interfaces.The two components are the ANN component and theCBR component (see Figure 1).

Fig 1: A Schematic View of the hybrid Neural-CBRArchitecture

3.1 The Multi-Layer Perceptron ANN ComponentThe multi-layer perceptron (MLP) ANN is a

powerful neural network model that can be used forsolving approximation, estimation, classification andprediction problems. Generally, it consists of an inputlayer, an output layer and one hidden layer. The hiddenlayer(s) and the output layer are the processing layers inthe network where activation takes place. The knowledgeof the network is encoded in the weights connecting theneurons. Each neuron in an inner layer acts as a linearcombiner whose summation function is given as:

Sum =

n

jii xww

10

(1)

Where w0 is the bias weight, wi and xi are theweight and input vectors respectively. The activation ateach neuron is given by the sigmoid function:

f(a) =)1(

1sume

(2)

The sigmoid function is continuous anddifferentiable in the interval [0, 1], and is one of mostlyused activation function for the MLP. The MLP is trainedusing the back propagation algorithm [35], which is a formof supervised learning, by presenting sample input-outputpairs to the network. The error difference between thenetwork’s output and the expected target output are fedback into the network for updating the weights connectingthe hidden-output layers and the input-hidden layers.

In our Neural-CBR system, symbolic dataobtained from case instances are transformed intonumerical data through pre-processing and fed as inputinto the MLP. The data pre-processing is therefore, anecessary precursor to the operations of the MLP-ANNcomponent of the Neural-CBR system.

3.2 The CBR ComponentThe set of input variable values and the predicted

output obtained from the MLP component is passed to theCBR component. The CBR component has a case baseindexed on unique case identity field (case_id) and thecomputed similarity score of each case. The typicalstructure 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 relativeto a particular case instance;

Case_attributesset – input values of individualattributes stored as a string separated bydelimiters;

Case_sellingprice – the predicted valuation of acase instance;

Case_weightSet – the set of weights associatedwith each attribute variable such that wi

represents the weight of the ith attribute.

3.2.1 Similarity AnalysisSimilarity analysis was done using the nearest

neighbour algorithm. The similarity measure used was theinverse of weighted normalized Euclidian distance. Asimilarity score is computed by:

SIM(X, Y) = 1- DIST(X,Y) (3)

DIST(X, Y) = (4)

Where X and Y are the new and stored caserespectively with n number of attributes while andare the normalized values for the ith attribute. A




614

normalized weight wi is assigned to each attribute basedon contextual experts’ knowledge of the locationconcerned. This calculation is repeated for every storedcase in the case base. The cases with the highest similarityscore (that is up to or above the similarity benchmarkvalue) are picked as candidates for adaptation in providingexplanation for the new case scenario. The algorithm ofthe Neural-CBR system is given in Figure 2.

Fig 2: Neural-CBR System’s Operational Procedure

3.3 Algorithm of Neural-CBR SystemIn this section, we present the algorithm of the

Neural-CBR systems that details how it reaches itsconclusions (see Figure 2). Given an input Cnew and theexistence of the case base CB. A variable Sim Benchmarkis set as the minimum acceptable value for sufficientsimilarity between cases. The algorithm selects theSimBenchmark value to be in the interval [0.75, 1.0], suchthat the value 0.75 implies a degree of similarity in theupper quartile while the value 1.0 denotes perfectsimilarity. Likewise, the value 0.5 connotes an averagesimilarity score. Step 4, 5, 6 in the algorithm represents thepre-processing, ANN training and ANN prediction phasesof the system’s operation. In Step 7, a one pass scan of thecase base (CB) was carried out to compute similaritybetween Cnew and the existing cases in the case base usingthe Weighted Euclidean distance (see equations 1, 2). Ifcases with similarity score up to or above theSimBenchmark exist (i.e. similarity ≥ SimBenchmark)then case adaptation is conducted as follows:

i. Rank all cases found by their similarity score;ii. Group the retrieved cases based on their solution

values Pcbr (case selling price) into different

categories, g1, g2, g3..., gk, with theircorresponding Pcbr as P1, P2,.., Pk;

iii. Take count of the number of cases retrieved ineach category gi (i = 1...k) and store them as t1,t2...tk);

iv. Choose category gi with the highest frequency;v. If there exist only one category gi with highest

frequency then take Pcbr of a case in gi as Pvalue

(i.e. final output) else, if there is more than onecategory gi with highest frequency, such as gi1,gi2... gim then compute average value Pave of allPim in gi as Pvalue (final output) and as Pcbr for Cnew

.vi. Next, use descriptions of case attributes

(Case_attributesset) in category gi as explanationfor Cnew (line 9-12).

If cases with similarity score of at least 0.5, butless than SimBenchmark exists then the Neural-CBRsystem uses the neural computed Pneural (ANN predictedoutput) in case adaptation thus:

1) Use retrieved cases to retrain ANN-MLP in orderto 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 variablesfor explanation. Where no sufficiently similarcase is found, then restate description attributesof Cnew for explanation (see line 15-18) i.e. for allretrieved cases Cr = {c1, c2 …cn) where fi

(i=1...15) is a specific attribute feature of Ck Cr,fj a corresponding attribute feature of Cnew, wi, theweight associated with attribute feature fi,compute SIM(fi, fj) = │wifi-wifj│ and rank caseswhere SIM(fi, fj) ≤ 0.5. Pick as explanation forCnew highest ranked fi in Cr, for each instancewhere SIM(fi, fj) ≤ 0.5 is not found use thedescription of fj in Cnew as explanation for Cnew.end for

If sufficiently similar cases do not exist then thesystem takes Pneural as Pvalue. However in this case itrestates the features of Cnew as explanation for output (line20-21).

4. THE CASE STUDYA prototype system was developed using data

collected from a real estate firm (Dan Odiete and Co. Ltd)in Nigeria [31]. The instance data used in the case study isfor the periods of 2002 to 2005, a sample is shown inTable 1. Data associated with fifteen core attribute featuresused in the appraisal of residential properties wereextracted and used to train the neural network componentto yield an estimate of the price of the property. Thefeatures, description and range of values obtained from thetraining set used are given in Table 2.




615

Table 1: Bedroom Bungalow BQ optional

Table 2: Description of features for real property valuation

4.1 Data-PreprocessingThe raw input data (see Table 1) were normalized

based on min-max normalization [40] to values between 0and 1 using the data-preprocessing interface of the Neural-CBR system. The rescaled values of the attribute featuresextracted from the data are as shown in Table 3. All thevalues are numerical except the following: borehole, BQ,fenced_round, neighbourhood_group and accessibility.These five were category inputs, therefore were

represented as A (available) or B (not available), whichare assigned vales 1 for A and 0 for B. In the case ofneighbourhood_group, A (high), B (medium), or C (low)are assigned values 1, 0.5 and 0 respectively.




616

Table 3: Showing normalized feature values from Table 1

4.2 Training the MLP-ANNThe configuration of the MLP-ANN model used

for our training instance is a 15-16-1 MLP in which thenumber of core attribute variables (15 of them)corresponds to the number of input neurons with onehidden layer containing 16 neurons and 1 neuron in theoutput layer, which returns as output the predicted salesprice estimate. The Neural-CBR system environmentallows for the specification of a desired configuration forthe MLP network, which is then dynamically created. TheMLP was trained using the back propagation algorithmwith three sets of data, the training set, the validation set(to verify correctness during training) and the test set. Thesummary of the ANN experiments that was conductedwith the Neural-CBR system framework in order to arriveat the 15-16-1 MLP configuration and other optimaltraining parameters such as number of neurons in thehidden layer, learning rate and threshold are presented inTables 4, 5 and 6. In each occasion recordings were takenand the average computed to determine the optimum valuein each case.

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

Table 5: Variation of learning rate

Table 6: Variation of threshold

After the training experiment, the MLPconfiguration was 15-16-1, with 1.0 learning rate and0.005 threshold value. This was used for predicting thesale price of a property and the predicted result passed tothe CBR component.

4.3 Implementation of the Neural-CBR SystemThe Borland C++ Builder version 6.0 software

was used as the programming platform to realize theNeural-CBR system. The case base was implemented asSQL Server database table of records (cases) indexed oncase_id and the computed weighted Case_simscore fields.The CBR component does similarity analysis and usesparameterized SQL statements to determine the best-case




617

matches. The procedure employed by the Neural-CBRsystem to reach its final conclusion is based on thealgorithm shown in figure 2. Figures 3-5 show

Fig 3: Feature Extraction Interface of the Neural-CBRSystem

Fig 4: Network Training Interface of the Neural-CBRSystem

Fig 5: Prediction and Explanation Interface of the Neural-CBR System

4.4 Performance Evaluation of Neural-CBR SystemEvaluation is the process of determining the

appropriateness of a specific system relative to itsfunctional requirements and objectives. Validation of anexpert system is conducted by determining whether thesystem's outcome is consistent with the conclusions of thehuman experts. Validation focuses on evaluating theoutcomes rather than the process by which the outcomes

are determined. In our specific case, the output of theNeural-CBR system was validated by using a directmethod [38]. In the direct method of expert systemvalidation a human expert does a quantitative assessmentof the expert system software by engaging it to perform asimple benchmark problem over a specified period. Theexpert then responds to a set of questions about thesoftware based on past experience. The questions arequantitative and are based on a 0 (very false) to 5 (verytrue) numerical scales. Thereafter, a single numericalfactor results called the Satisfaction Level that ranges from0 (least satisfied user) to 5 (most satisfied user) iscomputed, which is used to rank the expert systemsoftware in terms of its likelihood to satisfy a prospectiveend user.

4.4.1 Description of the Evaluation ExperimentThe objective of the validation experiment was to

determine the level of users’ satisfaction with the hybridNeural-CBR system relative to CBR alone and ANN alonesystems. In order to do this, a real property valuationsystem was created that allows the user to alternatebetween three system modalities, Neural-CBR, CBR andANN, such that, if one of the modality is activated theother two modalities are automatically disabled. Acomparative assessment of the three system modalitieswas then undertaken using some selected human expertsfrom the real estate industry for the evaluation. Theparticipants in the experiment were persons withsignificant professional experience in the real propertyvaluation. Fifteen (15) real estate experts drawn from twofirms, Dan Odiete & Co. and Ajeb Associates (bothlocated in Benin City, Nigeria) with varying professionalexperience ranging from 2 (lowest) to 15 (highest) yearswere selected to participate in the user-based systemusability experiment. Each received a copy of aquestionnaire instrument and had the system to beevaluated installed for them. The participants who werepeople with tangible knowledge of the use of softwarewere given one-week training on how to make use thesystem prior to the commencement of the experiment.Details of how to switch modalities and to engage thesystem in specific modality for operations such as datapreprocessing, training, and prediction were clearlyoutlined. In addition, the participants had the license toengage the system in as many trial sessions as deemedconvenient commencing with the actual assessment.Moreover, one IT support staff was temporarily assignedto each company for the trial period of three weeks. Thehuman experts were asked to give a rating of 0 to 5 of theirassessment of the performance of the three (Neural-CBR,CBR, ANN) system modalities covering seven aspects anda total of sixteen questions. Each question has specificweight assigned to it, which had the mutual consent ofhuman experts that participated in the experimentalprocess. Table 7 shows a sample of a typical response tothe questionnaire test instrument and how the satisfactionscore for an evaluator is computed.




618

Table 7: A Typical System Evaluation Template

4.5 Results and DiscussionAt the end of the evaluation experiment, the mean

satisfaction level as computed from the assessments of thefifteen real estate expert evaluators for the three systemsare as shown in Table 8. The Neural-CBR system had amean score of 3.83/5.0 (viz. 3.83 out of the possiblemaximum score of 5.0); CBR system had 3.64/5.0; andANN system 3.75/5.0; all which are indicative of goodperformance.

Table 8: Result of Evaluation Experiment

However, in order to determine the best of thethree systems; we compared the computed meansatisfaction level of the three systems to determinewhether the differences in the mean values are statisticallysignificant. To achieve this we formulated the followinghypothesis:Hypothesis one




619

H0: There is no significant difference in the meansatisfaction level of the CBR, Neural-CBR, and ANNsystems, and hence the three systems are at parperformance wise.

H1: There is significant difference in the meansatisfaction level of the CBR, Neural-CBR, and ANNsystem, and the Neural-CBR systems is better than theother two systems.Testing the hypothesis

In order to test the hypothesis that wasformulated, the Analysis of Variance (ANOVA) statisticwas employed to compare the three sets of data obtainedfrom the evaluation experiment. The coefficient ofvariation (CV) (see Table 10) of the sets of data wascomputed in order to determine the system with the bestrating distribution. This is given as:

Where σ = standard deviation of the data distribution, µ =mean of the data distribution

At 1% significance level (i.e. p < 0.01), it wasfound that the mean satisfaction levels of the three systems(ANN, Neural-CBR and CBR) systems are significantlydifferent because the p-value of 0.008305 from theANOVA test is less than 0.01 (see Table 9). In addition,the coefficient of variation (CV) for the hybrid Neural-CBR system is the lowest when compared to the other twosystems, which is indicative of a relatively better userrating. Therefore, H0 is rejected and H1 accepted, whichstates that there is significant difference in the meansatisfaction level of ANN, Neural-CBR and CBR systemsand the Neural-CBR system is better than the other twosystems.

Table 9: ANOVA Table for Mean Satisfaction LevelComparison

Table 10: Coefficient of Variation for three Systems

Generally the observations from the case studyreveal potential benefits of the novel hybridization ofANN and CBR. First, the output of the system shows theinterpolative power of ANN especially in instances whereother predictive models may be deficient. Second, theresults from the system reveal how well the ANNcomponent could make up for the limitation of the CBRcomponent in instances when there is lack of sufficientlysimilar old cases in the case base for a new case. At thesame time the Neural-CBR system leverages the existenceof a case base, in providing justifiable explanation forresults instead of being a black box like a typical ANN.Additionally, the provisioning of rich GUI interfaces forpreprocessing and training of ANN enables real-timeacquisition of expert knowledge in the process of solving aproblem such as being able to assign weights to specificreal property attribute variables, which makes the systemvery adaptive and suitable as a decision support tool.

5. CONCLUSIONIn this work, a novel hybridization of ANN and

CBR techniques for real estate property valuation has beendemonstrated. A prototype Neural-CBR system wasdeveloped and evaluated in a case study to confirm theviability of the concept. The result obtained revealed thatthe Neural-CBR combination offers more acceptable levelof usability and performance satisfaction relative tosolitary ANN systems and CBR systems. Also, the systemshowed significant strength in key areas of weaknessesusually associated with solitary ANN and CBR intelligentsystems, which gives merit to the hybridization. In futurework, we intend to investigate the applicability of hybridNeural-CBR systems to other business applicationdomains such as education, health, and finance where thepotential of Neural-CBR is yet to be fully explored.

REFERENCES[1] Aamodt A and Plaza “Case-based reasoning:

foundational issues, methodological variations, andsystem approaches”. AI Communications. 7(1):39-59. (1994)

[2] Almond N, Lewis O, Jenkins D, Gronow S andWare A “Intelligent systems for the valuation ofresidential property”. RICS Cutting Edge,Conference, Dublin 5-6 September. (1997)

[3] Armengol E. and Plaza E. “Integrating induction ina case-based reasoned”. Proceedings Second




620

European Workshop on Case-Based Reasoning:243-252. (1994)

[4] Bajo J, de Paz JF , de Paz Y , Corchado JM“Integrating case-based planning and RPTW neuralnetworks to construct an intelligent environment forhealth care”, Expert Systems with Applications36(3): 5844-5858. (2009)

[5] Bamberger, S and Goos K “Integration of case-based reasoning and inductive learning methods”,Proceedings First European Workshop on Case-Based Reasoning, ed. Richter, Wess, Althoff, andMaurer, 2:296-300. (1993)

[6] Bonissone P and Cheetham W “Fuzzy Case-BasedReasoning for Residential Property Valuation”.Handbook on Fuzzy Computing (G 15.1), OxfordUniversity Press. (1998)

[7] Bonissone P, Mantaras R “Fuzzy Case-BasedReasoning Systems”. Handbook of FuzzyComputing Section F4.3, Ruspini, Bonissone,Pedrycz (Eds.), Institute of Physics Publishers.(1998)

[8] Borst RA “Artificial Neural Networks: The NextModeling/Calibration Technology for theAssessment community?” Property Tax Journal,10(1):69-94. (1991)

[9] Borst RA “A Method for the Valuation ofResidential Properties using Artificial NeuralNetworks in Conjunction with GeographicalInformation Systems”. IAAO Conference, Dublin.(1994)

[10] Borst R “Artificial neural networks in massappraisal”. Journal of Property Tax Assessment &Administration 1(2):5-15. (1995)

[11] Callan JP, Fawcett TE, Rissland EL “CABOT: Anadaptive approach to case-based Search”,Proceedings IJCA. (1991)

[12] Churbuck DC “Learning by example”. Forbes,6/8/92, 149(2):130-132. (1992)

[13] Connolly D and Christey S “Learningrepresentation by integrating case-based andinductive learning”, Proceedings AAAI Case-BasedReasoning Workshop: 157, Washington, DC.(1993)

[14] De Jong, KA and Schultz AC “Using experience-based learning in game playing”. Proceedings ofFifth International Conference on MachineLearning: 284-290, Ann Arbor, MI, MorganKaufmann. (1988)

[15] Evans A, James H and Collins A “Artificial NeuralNetworks: an Application to Residential Valuationin the UK”. Journal of Property Valuation &Investment 11:195-204. (1993)

[16] Gayer, G, Gilboa, I. and Lieberman O “Rule-Basedand Case-Based Reasoning in Housing Prices”. TheB.E. Journal of Theoretical Economics, Vol. 7,Issue 1 (2007)

[17] Guan, J, Zurada, J and Levitan, A S “An AdaptiveNeuro-Fuzzy Inference System Based Approach toReal Estate Property Assessment”. Journal of RealEstate Research”. Vol. 30, No. 4. (2008)

[18] Hammond K “Learning to anticipate and avoidproblems through the explanation of failures”.Proceedings of Fifth International Conference onArtificial Intelligence:556-560, Philadelphia, PA.,Morgan Kaufmann. (1986)

[19] Juan, Y, Shih, Shen and Perng, Y “Decision supportfor housing customization: A hybrid approach usingcase-based reasoning and genetic algorithm”.Expert Systems with Applications, Elsevier, Vol.31, Issue 1, pp 83–93 (2006)

[20] Koldoner J “Maintaining organization in a dynamiclong-term memory”. Cognitive Science, Vol. 7(4):281-328. (1983)

[21] Koldoner J “An introduction to case-basedreasoning”. Artificial Intelligence Review 6(1):3-34. (1992)

[22] Li W and Shi H “Applying Unascertained Theory,Principal Component Analysis and ACO-basedArtificial Neural Networks for Real Estate PriceDetermination”. Journal of Software, Vol. 6, No. 9,pp. 1672-1679. (2011)

[23] Lin, H and Chen, K “Predicting Price of TaiwanReal Estates By Neural Networks and SupportVector Regression”. In Proc of the 15th WSEASinternational Conference. (2010)

[24] Malek M and Kanawati R “A Cooperating HybridNeural-CBR Classifiers for Building On-lineCommunities”.www.aic.nrl.navy.mil/papers/2001/AIC-01-003/ws5/ws5toc8.PDF. (2009)

[25] Manago M, Althoff K., Auriol E, Traphoner R,Wess S, Conruyt N. and Maurer F. “Induction andreasoning from cases”. Proceedings First EuropeanWorkshop on Case-Based Reasoning, ed. Richter,Wess, Althoff and Maurer 2:313-318. (1993)

[26] McCluskey W and Anand S “The application ofintelligent hybrid techniques for the mass appraisal




621

of residential properties”. Journal of PropertyInvestment and Finance 17(3):218-238. (1999)

[27] McCluskey W, Dyson K, McFall D & Anand S“Mass Appraisal for Property Taxation: AnArtificial Intelligence Approach”. Land EconomicReview 2(1):25-32. (1996)

[28] Murray-Smith R and Thakar S “Combining case-based reasoning with neural networks”. AAAIWorkshop on AI in Service and support,Washington. (1993)

[29] Nawawi AH, Jenkins D and Gronow S “Expertsystem development for the mass appraisal ofcommercial property in Malaysia”. Journal of theSociety of Surveying Technicians 18(8): 66-72.(1997)

[30] O’Roarty B, Patterson D, McGreal WS, Adair AS.“A case based reasoning approach to the selectionof comparable evidence for retail rentdetermination”. Expert System with Application12(4): 417-428. (1997)

[31] Owoloko EA “Neural Networks: A data miningtechnique for property appraisal”. M.Sc. Thesis(Unpublished), University of Benin, Nigeria. (2005)

[32] Pacharavanich P and Wongpinunwatana N “TheDevelopment of a Case-Based Reasoning System asa Tool for Residential Valuation in Bangkok”. Proc.of the 6th Annual Pacific-Rim Real Estate SocietyConference:1-14, Sydney. (2000)

[33] Peterson, S and Flanagan, A B “Neural NetworkHedonic pricing models in mass real estateappraisal”. Journal of Real Estate Research, Vol.31, No. 2, pp 147-164. (2009)

[34] Portinale L, Torasso P, Ortalda C and Giardino A“Using case-based reasoning to focus model-baseddiagnostic problem solving”. Proceedings FirstEuropean Workshop on Case-Based Reasoning, ed.Richter,Wess, Althoff and Maurer 2:335-340.(1993)

[35] Rich E and Knight K “Artificial Intelligence 2nd

ed”. McGraw-Hill, Inc. (1991)

[36] Rissland E and Skalak D “CABARET: Ruleinterpretation in a hybrid architecture”, Int. J. ofMan-Machine Studies, 34(6): 839-887. (1991)

[37] Rossini PA, “Accuracy Issues for Automated andArtificial Intelligent Residential ValuationSystems”, International Real Estate SocietyConference, Kuala Lumpur, January 26-30. (1999)

[38] Salim, MD, Villavicencio A and Timmerman MA“A method for evaluating expert system shells forclassroom instruction”. Journal of IndustrialTechnology, Vol. 19, No. 1, pp. 1-11. (2002)

[39] Scott I and Gronow S “Valuation expertise: itsnature and application”. Journal of Valuation8(4):362-375. (1989)

[40] Shalabi LA, Shaaban Z and Kasasbeh B “DataMining: A Preprocessing Engine”. Journal ofComputer Science. 2(9): 735-739. (2006)

[41] Sycara KP “Using case-based reasoning for planadaptation and repair”. Proceedings of DARPAworkshop on case-based reasoning. San Mateo,Calif. Morgan Kaufmann. (1988)

[42] Taffese WZ “Case-based reasoning and neuralnetworks for real estate valuation”. Proceedings ofthe 25th IASTED International Multi-Conference:artificial intelligence and applications, Innsbruck,Austria: 98-104. (2007)

[43] Tay D and Ho D “Intelligent Mass Appraisal”.Journal of Property Tax Assessment &Administration 1(1): 5-25. (1994)

[44] Wilson ID, Paris SD, Ware JA and Jenkins DH“Residential property price time series forecastingwith neural networks”. Knowledge Based Systems15(5–6): 335–341. (2002)

[45] Worzala E, Lenk, M., & Silva A “An Explorationof neural networks and its application to real estatevaluation”. Journal of Real Estate Research10(2):185-201.

AUTHOR PROFILESAdebola G. Musa is a doctoralcandidate at TshwaneUniversity of Technology,South Africa. He earns MS andBS Computer Science atCovenant University andUniversity of Ilorin, Nigeriarespectively. His researchinterests include the SemanticWeb, Artificial Intelligence and

Neural Networks, e-Government. He has to his creditseveral research articles in conference proceedings andrefereed journals. He is a member of IITPSA.




622

Dr. Olawande Daramola is aSenior Faculty at Departmentof Computer and InformationSciences, Covenant UniversityNigeria. His research interestsinclude the Semantic Web,Intelligent Systems, andAutomated SoftwareEngineering.

He is widely published in top range journals andconferences. He is a member of ACM, IEEE, and theNigerian Computer Society.

Owoloko E. Alfred is adoctoral candidate atCovenant University, Ota,Nigeria. He earned his MSand BS from the Universityof Benin, Nigeria. Hisresearch areas includestochastic differentialequation to valuation,application of Neural

Networks to solving real life problems, mathematicalstatistics and operation research. He has to his creditseveral research articles in refereed journals.

Olugbara O.O. is aProfessor of ComputerScience and InformationTechnology at DurbanUniversity of Technology,Durban, South Africa.

journal of computing::a neural-cbr system for real...

Documents