2017 1st International Conference on Informatics and Computational Sciences (ICICoS)

Predicting House Sale Price Using Fuzzy Logic,

Artificial Neural Network and K-Nearest Neighbor

Muhammad Fahmi Mukhlishin, Ragil Saputra, Adi Wibowo

Department of Informatics
Faculty of Science and Mathematics, Diponegoro University
Semarang, Indonesia
fahmi.mukhlish@gmail.com, ragil.saputra@undip.ac.id, bowo.adi@undip.ac.id

Abstract—Determining the value of land and home are market in Hong Kong [6]. Several machine learning methods
regularly determined at the earliest by the seller, however are compared to get the best prediction of house pricing [7].
determining the right price in the sales process will affect the
buyer's desire to elect and bid. Special characteristics in In advance, spatial analysis is used to determine house and
Indonesia, tax object value (NJOP) and location parameters are land prices [8]. Jeffrey et. al discussed the spatial effects on the
high influence to the price. In this paper we proposed the dynamics of house prices in the USA that have a significant
prediction of land and house value using several methods. Fuzzy influence in urban house price growth rates [9]. In addition,
logic, Artificial Neural Network and K-Nearest Neighbor are real estate pricing in Vienna (Austria) was investigated that
compared in this paper to discover the most appropriate method requires recognition of spatial heterogeneity in housing prices
that can be used as a reference for determining the price by the [10]. However, only few studies that apply the prediction
sellers. Google Maps is used to represent the spatial data for model using spatial data as analysis factor for the prediction.
prediction parameter. The variables that used in the methods are
NJOP of land, the locations, the age, NJOP of house, and the In this study, several prediction methods were compared to
valuable location of the land. The experimental methods are find out the best predicted results for determining the selling
tested by comparing between the real price transaction and the price of a house compared to the real price. The methods
prediction using MAPE formula. (Fuzzy, Neural Network, k-nearest neighbor) were tested to
predict house prices in Indonesia. In this study, spatial data and
Keywords—House Price, Fuzzy, Artificial Neural Networks, K- attribute data are employed as determinant parameters in the
Nearest Neighbor. determination of house and land prices. Spatial data consists of
the selling value of the tax object (NJOP) of the land and the
I. INTRODUCTION strategic location of the land, while the attribute data includes
House is one of the important needs for people which has the selling value of tax object (NJOP) of the building, the
function as a place for rest and gather with family [1]. In the condition of the house and the age of the house. This study
housing market, the initial prices are an important factor in the uses spatial data in the form of a strategic location of the land,
process of buying and selling houses and land. Determining the called accessibility level that influence by infrastructure
initial selling price of a house or land usually depends on the facilities such as public health services, education, downtown
seller, however determining the right price in the sales process and economy. In spatial data the Geographic Information
will affect the buyer's desire to bid and make selections. The System (GIS) has three elements, namely vertices, lines (arcs),
initial price for each house and land are varied according to and broad (polygons) in the form of vectors or raster
residential facilities and home geographical conditions. In representing geometric topology, size, shape, position and
Indonesia, one of the parameters to determine the selling price direction.
of the house that can be quantity calculated are the value of the
tax object (NJOP) of the land and the value of the tax object of II. METHOD
the building [2]. Both parameters are influenced by several A. Fuzzy Logic
factors such as strategic location and also age of building.
Fuzzy logic was first introduced by Lotfi Zadeh in 1965.
The initial house prices prediction is challenging and Fuzzy logic is a mathematical framework used to represent
requires the best method to get the best prediction accuracy. In uncertainty. The presentation of set A with the zero-one
the predicting the sale price of a house that has an uncertainty membership function, otherwise known as the crisp set can be
parameter, fuzzy logic becomes one of the solutions that can be represented by the following equation [12].
used in solving the problem [3-4]. Moreover, artificial neural
network methods are used to predict house selling prices [5]. In μ_(A ) (x)={ _(0 if x ∉ A)^(1 if x ∈ A) (1)
addition to using fuzzy logic and artificial neural networks, The above equation shows that the Fuzzy set is a
predictions can also use the K-Nearest Neighbors algorithm, generalization of the crisp set by allowing the membership
for estimating residential prices for the residential property function to retrieve the values in the interval [0, 1]. In other
words, the membership function of the crisp set can only take a

978-1-5386-0903-3/17/$31.00 © 2017 IEEE 171

value of zero and one, while the membership function of the
fuzzy set is a continuous function with the range [0, 1] [11].
B. Fuzzy Tsukamoto
The Fuzzy Tsukamoto model is one of the models in the
fuzzy inference system shown in Figure 2.7. [12]. The
Inference System is a process of reasoning about a given state,
using all available knowledge to produce the best estimates of
output. Fuzzy Inference System (FIS) or often called fuzzy
expert system is a popular computing framework based on
fuzzy set theory, IF-THEN fuzzy rules, and fuzzy reasoning. In Fig. 1. Artificial Neural Networks
the fuzzy system, the inference engine is used to adjust the
fuzzy set pattern of the inputs with the antecedents of all fuzzy The working principle of K-Nearest Neighbor is to find the
rules and combine all responses to produce the fuzzy set of closest distance between the data to be evaluated with the
output [12]. In the Fuzzy Tsukamoto model, any consequent nearest neighbor in the training data. Training data is projected
IF-THEN-shaped rules should be represented by a fuzzy set into many-dimensional spaces, of which each dimension
with a monotonous membership function. Here are the steps of explains the features of the data. This space is divided into
Fuzzy Tsukamoto's model [12]. sections based on training data classification. A point in this
1) Calculate the value of α-predicate (fire strength) of space is denoted by class c, if class c is the most commonly
each rule during the process of evaluation of rules in encountered classification of the nearest neighbor of the point
the inference engine by using the MIN implication [14].
function.
2) Counting the inference results expressly (crisp) of E. Geographic Information System
each rule (z_1, z_2, ..., z_n). The calculation is based According to S. Murai, Geographic Information System
on the fire strength value of each rule (w_1, w_2, ..., (GIS) is an information system used to enter, store, recall,
w_n). process, analyze and produce geo-referenced data or geospatial
3) The defuzzification process uses the weighted average data, to support decision making in land use planning and
method . management, natural resources, environment, transportation,
C. Artificial Neural Network City facilities, and other public services [15].
Artificial Neural Networks are one of the artificial Geospatial data sources are digital maps, aerial
representations of the human brain that always simulate the photographs, satellite images, statistics tables, and other related
learning process in the human brain. The term artificial here is documents. Geospatial data can be divided into graphical data
used because this neural network is implemented by using a called geometric data and attribute data (thematic data).
computer program capable of completing a number of Graphic data has three elements, namely the point (node), the
line (arc), and the area (polygon) in the form of vectors or
calculation processes during the learning process. [13]
raster representing the geometry of topology, size, shape,
The back-propagation algorithm for the training process position, and direction [15].
using in feed forward neural network algorithm are employed
F. Google Maps API
in this study. The back propagation neural network algorithm
apply to do looping at two stages of propagation and repeated, Google Maps uses HTML and Javascript programming
until achieved acceptable results in training. Moreover, in the languages to make it possible to place Google Maps apps on
another web. The Google Maps API is a feature issued by
feed forward neural network, the information moves forward,
google to facilitate the user integrating Google Maps into the
one direction from the input to the output (via a hidden node)
user's website. In order for Google Maps to appear on the web
without the loop. as desired, it needs API-key. API-key is a special code
D. K-Nearest Neighbor(K-NN) generated by google for a particular website, so the Google
Maps server can recognize [16].
The K-NN algorithm is one of the methods used for
classification analysis, but the last few decades the KNN G. Mean Absolute Percentage Error
method has also been used for prediction [14]. K-Nearest Mean Absolute Percentage Error (MAPE) Is the average of
Neighbor algorithm is a method to classify objects based on the overall percentage error (difference) between the actual
learning data closest to the object. Nearest Neighbor is an data with the data forecasting results. The accuracy measure is
approach to finding cases by calculating the proximity matched with time series data, and is shown in percentages
between the new case and the old case that is based on with the MAPE formula as follows:
matching the weights of a number of existing features [14]. | |
∑ × 100%
=

MAPE = Average absolute percentage error

172
 = Value of Real Transaction
= Prediction Value
n = Number of data
III. PROPOSED MODEL
The real house prices in Pedurungan Sub-district of
Semarang City are used in this study. The training data consists
of 7 predictor’s attributes and 1 label attribute. In order to
improve the data quality and eliminating duplication,
preprocessing was applied. Fig. 2. Membership fuction of NJOP-L

In our proposed method, the determination of the initial

selling price of the house and land are considered by the value
of sales value of taxable object land (NJOP-L), sales value of
taxable object building / house (NJOP-B), house age, house
condition and land strategic location. The prediction system is
equipped with Google Maps to facilitate in determining the
strategic location of the land, get the sales value of taxable
object price of land and the strategic location value of the land.
A. Fuzzy Logic
In this experiment, intuition is used to determine the Fig. 3. Membership fuction of NJOP-B
membership function. Intuition is based on human intelligence
and deep understanding to develop membership functions.
Determination of the shape or type of curve, and the overlap
between curves used adjusted based on research that has been
done that is represented in Fig 2-8. The fuzzy set of input
variables and output variables of the house and land sale
pricing system are as follow:
1) Input Variables
a) Sales Value of Taxable Object Building (NJOP-B)
Variables : input variables obtained from sales value of Fig. 4. Membership Function of House Age
taxable object data building set by the Government of
Semarang City
b) Sales Value of Taxable Object Land (NJOP-L)
Variables: input variables obtained from sales value of taxable
object data of land determined by Semarang City Government
represented in map form which can be seen on Fig.9.
c) House Age Variables : input variables obtained from
the calculation of the age of the house refer to the building
establishment permit (IMB) as shown in Fig. 4.
d) House Condition Variables: input variables obtained Fig. 5. Membership Function of House Condition
from the state and structure of the house as a whole according
to assumptions of homeowners and assumptions of home
buyers that assessed in the percentage as shown in Fig. 5.
e) Strategic Value of Land Location Variables: input
variables that can be represented as the level of accessibility
locations as shown in Fig. 6.
2) Ouput Variables
a) Predicted House Prices Variables: the output variable
that became the initial output of the first fuzzy calculation on Fig. 6. Membership Function of Land Location Strategic Value
the house price calculation as shown in Fig. 7.
b) Predicted Land Prices Variables: the output variable
that becomes the output of the second fuzzy calculation on the
house price calculation as shown in Fig. 8.

173
 TABLE I. FIRST RULEBASE FUZZY
Sales Value
of Taxable House Output Predicted
No House Age
Object Condition House Prices
Building
1 LOW LOW LOW LOW
2 LOW MID LOW LOW
3 LOW HIGH LOW LOW
4 LOW LOW MID LOW
5 LOW MID MID LOW
Fig. 7. Membership Function of Predicted House Prices Value 6 LOW HIGH MID LOW
7 LOW LOW HIGH HIGH
8 LOW MID HIGH LOW
9 LOW HIGH HIGH LOW
10 MID LOW LOW HIGH
11 MID MID LOW LOW
12 MID HIGH LOW LOW
13 MID LOW MID HIGH
14 MID MID MID LOW
15 MID HIGH MID LOW
16 MID LOW HIGH HIGH
17 MID MID HIGH HIGH
Fig. 8. Membership Function of Predicted Land Prices Value 18 MID HIGH HIGH LOW
19 HIGH LOW LOW HIGH
20 HIGH MID LOW HIGH
21 HIGH HIGH LOW LOW
22 HIGH LOW MID HIGH
23 HIGH MID MID HIGH
24 HIGH HIGH MID LOW
25 HIGH LOW HIGH HIGH
26 HIGH MID HIGH HIGH
27 HIGH HIGH HIGH HIGH

TABLE II. SECOND RULEBASEFUZZY

Sales Value of Strategic Location Output Predicted
No
Taxable Object Land Land Land Prices
1 LOW LOW LOW
2 LOW MID LOW
3 LOW HIGH HIGH
4 MID LOW LOW
5 MID MID HIGH
Fig. 9. Map of Land Classification Based on sales value of taxable object 6 MID HIGH HIGH
Price 7 HIGH LOW LOW
8 HIGH MID HIGH
The fuzzy rules are operated using a series of IF-THEN 9 HIGH HIGH HIGH
statements. In the first fuzzy rules there are 3 input variables;
sales value of taxable objects building (NJOP-B), house age
and house condition. The output variable for this fuzzy rule is Training process used back-propagation neural network with
the house price prediction. Moreover, two input variables are parameter: MSE 0.00001, 0.3 learning rate, 500 epochs, and
employed in the second fuzzy rules; sales value of taxable 0.3 momentum. The spatial data is converted to value in this
object land (NJOP-L) and the strategic value of land location. method for perdition using neural networks.
The output of the second fuzzy rule is the land price prediction. C. K – Nearest Neigbor
The number of statements on each rule base is the
multiplication of the number of input sets. The first fuzzy rule In this experiment, the similar data training and testing
has 27 statements formed from 3 input variables with each with neural networks are employed. The k-Nearest Neighbor
variable having 3 sets LOW, MID and HIGH as shown in algorithm is based the comparing a given test example with
Table I. The second fuzzy rule has 9 statements formed from 2 training examples. The k-nearest neighbor is applied using
input variables with each variable having 3 sets LOW, MID, k=1.
and HIGH as shown in Table II.
IV. RESULTS
B. Artificial Neural Network The comparing between the real price and the prediction
In this study, the artificial neural network architecture built price are employ in the testing experiments using MAPE. The
consists of 3 layers, namely: the input layer, a hidden layer, and real price is collected from the real transactions from the
output layer. Each layer is associated with weights. property agent.

174
 .

TABLE III. PREDICTION PRICING TEST

Transaction Fuzzy Percent ANN. Percent Percent

No Suitability Suitability KNN Pred. Suitability
Price Pred. Error Pred Error Error
1 425 531 25% 75% 504 19% 81% 600 41% 59%
2 355 433 22% 78% 461 30% 70% 550 55% 45%
3 800 679 15% 85% 686 14% 86% 1500 88% 13%
4 1400 1017 27% 73% 1191 15% 85% 1300 7% 93%
5 1300 1311 1% 99% 734 44% 56% 1300 0% 100%
6 650 646 1% 99% 658 1% 99% 1500 131% -31%
7 1000 1017 2% 98% 1244 24% 76% 1100 10% 90%
8 600 646 8% 92% 492 18% 82% 550 8% 92%
9 550 526 4% 96% 771 40% 60% 1500 173% -73%
Average 12 % 88% 23 % 77% 57 % 43%

TABLE IV. PREDICTION PRICING TESTWITHOUT VARIABLE STRATEGIC LOCATION LAND

Transaction Percent Percent KNN Percent

No Fuzzy Pred. Suitability ANN. Pred Suitability Suitability
Price Error Error Pred. Error
1 425 428 1% 99% 730 72% 28% 600 41% 59%
2 355 306 14% 86% 565 59% 41% 550 55% 45%
3 800 554 31% 69% 833 4% 96% 1500 88% 13%
4 1400 1131 19% 81% 1314 6% 94% 1300 7% 93%
5 1300 1354 4% 96% 1858 43% 57% 1300 0% 100%
6 650 548 16% 84% 833 28% 72% 1500 131% -31%
7 1000 900 10% 90% 1220 22% 78% 1100 10% 90%
8 600 520 13% 87% 1747 191% -91% 550 8% 92%
9 550 520 5% 95% 1074 95% 5% 1500 173% -73%
Average 13 % 87% 58 % 42% 57 % 43%

Fig. 10. Prediction Pricing Test Chart

Fig. 11. Prediction Pricing TestWithout Variable Strategic Location Land

In table III, the prediction experiment using all input .

variables are compared with the real prices. In this experiment,
the fuzzy method produces the best accuracy with 88%, a V. CONCLUSIONS
significant difference compared to the artificial neural network In this study, fuzzy, artificial neural network and k-nearest
method with 77% and the k-nearest neighbor method with neighbor are used to predict the selling price of a house. We
43%. The second experiment, the strategic location of the land tested the performance of this method by measuring how
is removed as input parameter. Based on the experiments accurate the results of the predictions were generated compared
results as show in Table IV, the fuzzy method still has the best to the real house selling price using MAPE. The spatial data
performance compare to ANN and k-NN. The comparison and attribute data are used in the prediction system. Based on
graph of prediction price shows in Fig 11 and 12. Therefore, the experiment results show that the fuzzy method is superior
fuzzy method has high accuracy due to un-employing training to neural networks as well as k-nearest neighbor for the house
method in the method modeling. This is extremely overturned price prediction in limited data training. In future, the
with ANN and k-NN that are heavily affected by training optimization of fuzzy rule and increasing data training are
process. In this study we only use 18 training data that required for improving the prediction accuracy.
represented several conditions.

175
 REFERENCES [8] Basu, Sabyasachi, and Thomas G. Thibodeau. "Analysis of spatial
autocorrelation in house prices." The Journal of Real Estate Finance and
[1] Raharjo, N. P. “Dinamika pemenuhan Kebutuhan Perumahan Economics 17, no. 1, pp. 61-85. 1998
Masyarakat Berpenghasilan rendah”. Semarang: Magister Teknik
[9] Cohen, Jeffrey P., Yannis M. Ioannides, and Win Wirathip
Pembangunan wilayah dan Kota Universitas Diponegoro, p. 30. 2010.
Thanapisitikul. "Spatial effects and house price dynamics in the USA."
[2] Fahirah, F., Basong, A. & Tagala, H. H. “Identifikasi Faktor yang Journal of Housing Economics 31, pp. 1-13. 2016.
Mempengaruhi Nilai Jual Lahan dan Bangunan pada Perumahan Tipe
Sederhana”. Jurnal Smartek, Volume 4, pp. 251 - 269 . 2010. [10] Helbich, Marco, and Daniel A. Griffith. "Spatially varying coefficient
models in real estate: Eigenvector spatial filtering and alternative
[3] Kuşan, Hakan, Osman Aytekin, and İlker Özdemir. "The use of fuzzy approaches." Computers, Environment and Urban Systems 57, pp. 1-11.
logic in predicting house selling price." Expert systems with 2016
Applications 37, no. 3, pp. 1808-1813. 2010.
[11] Klir, George, and Bo Yuan. Fuzzy sets and fuzzy logic. Vol. 4. New
[4] Gerek, Ibrahim Halil. "House selling price assessment using two Jersey: Prentice hall, 1995.
different adaptive neuro-fuzzy techniques." Automation in Construction
41, pp. 33-39. 2014. [12] Sivanandam, S., Sumanthi, S. & Deepa, S. Introduction to Fuzzy Logic
Using Matlab. Heidelberg: Springer-Verlag Berlin. 2007.
[5] Peterson, Steven, and Albert Flanagan. "Neural network hedonic pricing
models in mass real estate appraisal." Journal of Real Estate Research [13] Sayyed Mohsen Vazirizade, Saeed Nozhati, and Mostafa Allameh
31, no. 2, pp. 147-164. 2009 Zadeh, "Seismic Reliability Assessment Of Structure Using Artificial
Neural Network," Journal of Building Engineering, vol. XI, pp. 230-
[6] Cheung, Simon KC, and Sahminan Sahminan. "A Localized Model for 235, 2017.
Residential Property Valuation: Nearest Neighbor with Attribute
Differences." International Real Estate Review 20, no. 2, pp. 221-250. [14] Larose, Daniel T. "K‐nearest neighbor algorithm." Discovering
2017. Knowledge in Data: An Introduction to Data Mining, pp. 90-106. 2005
[15] Murai,GIS Work Book. Tokyo: Institute of Industrial Science. 1999.
[7] Byeonghwa Park and Jae Kwon Bae, "Using Machine Learning
Algorithms For Housing Price Prediction: The Case Of Fairfax County, [16] Chtiara, C. Implementasi sistem sistem informasi geografis (SIG)
Virginia Housing Data," Expert Systems with Applications, vol. XLII, Universitas Indonesia (UI) berbasis Web dengan menggunakan google
pp. 2928-2934, 2015. maps api. Universita Indonesia. Jakarta: Jurnal UI. 2008

176