See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/342761024

Real Estate Market Data Analysis and Prediction Based on Minor Advertisements
Data and Locations’ Geo-codes

Article  in  International Journal of Advanced Trends in Computer Science and Engineering · June 2020
DOI: 10.30534/ijatcse/2020/235932020

CITATION READS

1 228

2 authors, including:

Waleed Al-Sit
Mu’tah University
6 PUBLICATIONS   44 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Waleed Al-Sit on 16 July 2020.

The user has requested enhancement of the downloaded file.

 ISSN 2278-3091
Waleed T. Al-Sit et al., International Journal of Advanced Trends 9,
Volume in No.3,
Computer
MayScience
- June Engineering, 9(3), May – June 2020, 4077 – 4089
and2020

International Journal of Advanced Trends in Computer Science and Engineering

Available Online at http://www.warse.org/IJATCSE/static/pdf/file/ijatcse235932020.pdf
https://doi.org/10.30534/ijatcse/2020/235932020

Real Estate Market Data Analysis and Prediction Based on

Minor Advertisements Data and Locations’ Geo-codes
Waleed T. Al-Sit 1, Rashed Al-Hamadin 2
1
Department of Computer Engineering, Mu'tah University. Al-karak, Jordan , [email protected]
2
Princess Sumaya University for Technology, Jordan, [email protected]

 Most of economy aspects can be defined or described in

ABSTRACT terms of property sector, as it is in combination with food are of
the oldest sectors to have emerged in correspondence to sheer
Country’s economic status can derived from many complex demand imposed by the basic needs for survival for modern
and branched indicators, one of which is property prices humans [2]. Though other species do have, housing needs that
estimation. Working on such indicator changed the state of are not necessarily simple as represented with the case of
literature from many perspectives and corners. Whilst the beavers, and their well-engineered dams. Rather, humans took
scarcity of such works imposes a need for it, and demonstrates it several steps ahead, and delegated the housing and
an unutilized aspect of the economy that requires little construction to form the earliest jobs and occupations of
resources to create some business and academic opportunities. builders. In addition, it further expanded with economic
In this work, efforts evolved to address the problem of evolution with further branching and breaking down of roles to
estimating properties prices accurately, in specific apartment’s include most of early occupations and industries. For instance,
prices among The Amman City, The Capital of The Hashemite dominating engineering until the early days of the industrial
Kingdom of Jordan. Leading to shed the lights on employing revolution expanded to span over other sectors and industries
data science different techniques namely data processing, beside construction, which did not stop the property industry
analysis and predictive modeling for adopting and estimating
from being a major player in the economic scene until now [3].
the apartment’s prices based on advertisement data published
Furthermore, it displayed its dominance with the emergence of
through the web and its extracted location geocodes. In
economies of scale, banking economies and loan based
addition, the work evaluates the final analysis reported results
economies to a great degree as shown by the financial crisis in
based on selected evaluation measures, and compare them with
other five similar works on such problem conducted in other 2008 [4]. Being of this importance; i.e. property industry, lead
countries. Trying to aim to enrich the literature with valuable to trigger a lot of work on the topic as demonstrated by the
insights gained using Machine Learning and Data Mining plethora of scientific works and publications in the field.
different predictive techniques mainly, and its related Leading to form a solid base of literature around the topic
conditions, branches and requirements for other data especially in the financial aspect [5]. Such an example and
processing and analysis techniques under the data science original motivation behind this work is the property price
umbrella. prediction, having set itself with the iris dataset related work as
the default examples of the trendy and important field of data
Key words: Deep Learning, Machine Learning, Predictive science.
Modeling, Property Prices Analysis Prediction. In this view, and due to the interesting attributes of the
property problem. Alongside, how it corresponds to the
1. INTRODUCTION interrelations and different macroeconomic components [6].
Furthermore, the rich nature of data involved in house price
Property prices; a great indication of a country’s economic estimation and prediction, and the semi obvious nature of the
status. It gives an indication of the economic wellbeing and factors involved of property prices, and the existence of
stability, and being considered a clear manifestation of previous more traditional methods for the estimation process
inflation amongst other economic phenomena [1]. Usually, this [7]; all makes for a great use case for studies, serving both
field observed to be either the original sin and/or the most business related studies and academic applications. Because it
prominent presentation of most economic events, and it lead related parties to reveals the hidden aspects of a country’s
considered being so complex and branched sector. economy, whilst also demonstrating clearly the effects of data
science and predictive analysis and data manipulation

4077
Waleed T. Al-Sit et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(3), May – June 2020, 4077 – 4089

Table 1: Similar Works Summaries

Inputs Results
Dataset Dataset Used Models Data Model Evaluation Value
Dimensions Splits Criteria
UCI’s Dataset 452 X 13 SVM (1), LSSVM (2), PLS (3) 400-52 (1) MSE 10.7373
(Housing Value of Boston R. Time 0.4610s
Suburb) (2) MSE 20.3730
R. Time 20.3730s
(3) MSE 25.0540
R. Time 0.7460s
Local Data in Spain 1187 X 6 MLP [1 Hidden Layer] 952-237 MLP R2 0. 8605
RMSE 39540.36
MAE 28551.34
“Zillow.com”, 21,000 X 15 Linear Regression (1), Multivariate 80% - (1) RMSE 1.201918
”magicbricks.com” Regression (2), Polynomial 20% (2) RMSE 16,545,470
Regression (3)
(3) RMSE 11359157
bProperty.com 3505 X 15 GB-Regression (1), Random Forest 80% - All RMSE 0.1864 to 0.2340
(2), SVM Ensemble (3) 20%
random sample from 200 X 7 MLP 80% - MLP R2 0.6907 to 0.9
www.bluebook.co.nz 20% RMSE 449, 111.46 to 1,
014,721.92

techniques due to the straightforwardness of the problem and provides the predictive modeling analysis different
the size of the available work on it [8]. Starting with the experiments conducted using machine learning. In Section 5,
benefits yielded of an accurate estimation for property buyers, the work provides the reported results alongside the needed
alongside the assumption related to if a certain property priced discussions, before concluding the presented work in section 6
fairly, and to have a better idea about the possible impacts of and its future next steps.
each of the attributes on the price and in what way. All are
leading to facilitate the process of making a purchase 2. SIMILAR WORK
decision and budget and priorities setting for them. In addition,
helping the property investors in knowing if a purchase deal is As a heavily covered topic especially in the recent years, the
a bargain with a high margin of profit or not. On the other side options to compare to and benefit from are plenty. Hence, we
of the same transaction type for describing the property sales chose a sample of few papers that addressed the price
process, we have property sellers and potential investors in the estimation problem that felt closest and most relevant to this
property sector whether they are individuals or corporate sized work.
parties. Who are looking forward to understand the market Starting with hedonic regression, where we performed
prices, and how much they expect to charge for their properties, several regression to each attribute individually. Then, trying to
and what aspects related to the market in order to concentrate observe the change in a target attribute (Price), so constructing
on, and what to dismiss, all are a common use case for data the final equation of the predicted variable of the weighted
science some techniques. estimation attributes. With an origin out of the real estate,
The contributions of the work presented in here are pricing [9] and some interesting use cases [10]. A lot of
threefold: (1) demonstrates utilizing data science to figure literature on this problem is studying the alternative methods
more countries’ economic aspects, which requires huge for the estimation with a concentration on the use of machine
resources to create some business and academic opportunities learning. Table.1 summarizes five works followed with a
and insights. (2) Presenting a comparative predictive modeling discussion for each further more in this comparative work.
and analysis study for Jordanian property market against five In the first work [11], using a UCI dataset with housing prices
similar works conducted in different countries. (3) Finally, data in Boston, and 13 attributes in the data with a relatively
reporting the analysis resulted insights upon adopting different low number of instance at 452. Authors presented their work
machine learning techniques and evaluation measures for the as comparison of three different models; namely Partial Least
presented problem. Squares (PLS) regression, Support Vector Machine (SVM),
The rest of the work organized as follows. In section 2, and Least-Squares Support-Vector Machine (LSSVM), where
presenting similar works for other countries and their reported the third is somewhat a kernelled version of regular SVMs
results. Section 3 provides the methodology adopted to conduct with an optimization included by design. The results shows
end-to-end data preprocessing and Explanatory Data Analysis that SVMs where superior to the other two methods both on
(EDA) considering real estate apartments prices prediction Mean Squared Error (MSE) and fitting time of the model,
based on advertisement data and locations geocodes. Section 4 though LSSVM would have smaller fitting time if the

4078
Waleed T. Al-Sit et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(3), May – June 2020, 4077 – 4089

Table 2: Web Scraped Data Attributes

Attributes Type Attributes ... Type ... Attributes ... Type ... Attributes ... Type ...
ID INT Ad Images Count INT Air Conditioning BIN Nearby Facilities BIN
Title STR City STR Heating BIN Security BIN
Date STR Location STR Balcony BIN Built-in Wardrobes BIN
Real Estate BIN No. Rooms INT Elevator BIN Swimming Pool BIN
Paid Ad Feature.1 BIN No. Bath Rooms INT Garden BIN Solar Panels BIN
Paid Ad Feature.2 BIN Area INT Garage Parking BIN Double pane Windows BIN
Paid Ad Feature.3 BIN Floor INT Maid Room BIN Ad Post STR
Paid Ad Feature.4 BIN Age STR Laundry Room BIN
Price INT Payment Type STR Is Furnished BIN

parameter optimization were dropped, while all models seems that has shown good results, exceeding the performance of
to yield acceptable results. other models that do not utilize the ensemble paradigm like
The next work presented in [12] provides a good outlook on ANNs. With 3505 instances in the data set and 19 attributes,
the Spanish real estate market. It utilizes data that spans over a the dataset considered is features rich and should allow for
long time interval by making use of a one hidden layer better estimation. As it reduces, the complexity of the fitted
Multi-Layer Perceptron (MLP), in order to yield an estimation model due to the higher dimensionality that is at the same time
model of the real estate price based on some exogenous might hinder the fitting and learning process due to the larger
variables. Authors claiming the superiority of one hidden layer search space for the solution. The results is good yet could
over two, where the results obtained are to some degree a benefit from a clearer presentation of the results, this while
supporter of the claim. However, further examination of such lacking in clarity demonstrated the potential benefit of
work should tested, considering the various architectures and ensemble methods for this type of regression. In addition, the
parameters that an MLP could take. While also taking into data preprocessing and transformation effect on the
account the benefits of the two hidden layers models [13] regression’s final output, and this with the respect to the data
beyond the general function approximation abilities that categories and how each might effect on the learning process in
artificial neural networks have [14]. with a somewhat rich data terms of model and training/fitting performance, while
on vertically with 1187 instance the data is slim horizontally elaborative on the effects of tuning machine-learning
with only 6 attributes, demonstrating with the calculated R2, ensembles with parameters such as depth and number of
RMSE and Median Absolute Error (MAE) the potential estimators [17].
Artificial Neural Networks (ANNs) have for solving such Following in the trend of neural models for price estimation,
problems and improvement over traditional hedonic models while going further with the hedonic versus ANN house price
which is an observation that is shared between a good fraction estimation. Where they share a similar definition of the
of literature. problem addressed to solve. Authors in [18] dives deeper into
Coming more on the technical side. Authors in [15] makes both the hedonic and artificial neural network theories.
use of rich data with 21000 instances and 15 attributes and Alongside the histories and the inflection of the
tries several regression models, namely linear, multivariate and aforementioned theories on the corresponding models. Rather,
polynomial regression. However, the evaluation method leaves scarce in the data used with 200 instances and 7 attributes.
a lot to be desired, going with RMSE solely leaves a need for Nevertheless, the models produced varies highly in prediction
data exploration to understand the results further more. But the performance. That is rather obvious in terms of R-squared,
use of such evaluation criterion could be understandable due to which it clearly explained in the different architectures of the
the nature of the experiment where most regression models rely used neural models in terms of number of neurons. However,
on minimizing some sort of error or residual to produce the the scarcity of data raises some questions regarding the top
model and with the iterative nature of the tuning both the obtained performance, that explained by the more complex
models and the data. model producing it. In addition, the number of test instance
Considering the ensemble way, authors in [16] based on the indicates less support for the results. Nonetheless the statistical
assumption that several models should yield better results. As nature of this work and the emphasis on the hedonic-neural
in the aggregated results of several models should present more comparison and attributes contribution. The analysis
support to a certain decision. This work utilizes several models conducted gives an indication, that these results are just

4079
Waleed T. Al-Sit et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(3), May – June 2020, 4077 – 4089

indicative of the statistical conclusions of a potential

superiority for the neural models over hedonic ones, in the real
estate price regression demonstrated with the use of measures
like White Heteroscedasticity Test and confidence intervals.
While also showing that, even aggregated features may lose
some information, but are still a feasible option if fitting
performance is of an issue [19]. An interesting observation
here would the different contributions to each attribute and its
importance level for the regression in the two models
presenting the low interpretability of them in terms of their
produced model represented by patterns seen in data to produce
the regression or classification shall not always comply with
the higher level semantics observed by humans [20].
Most of literature including the discussed examples above use
Figure 1: Advertising Percentage (%) per City.
an 80% and 20% for the training and test data splits. In
addition, the most addressed the hedonic regression as a
common method for such a problem and use case. while also
describing it as the method to be replaced by machine learning,
due to its more dynamic and complex nature [21], [22]. Which
allow capturing more intricate trends and adapting more
through time relying on online machine learning methods [23].
In the meantime, facilitating the time based adaptation and
therefore performing better for economic analysis [24], either
presenting data over a temporal axis implicitly or explicitly,
which will accommodated by some machine learning models
in later stages of analysis. For instance, Hidden Markov
Models (HMMs) and Recurrent Neural Networks (RNNs)
[25], with some future anticipated and currently demonstrated
cases of image based price estimations [26]. As much other
type of estimators in different areas that rely on machine
learning [41]-[43]. Figure 2: No. of Advertisements Listed per Advertisements’ Dates.

3. JORDANIAN MARKET: REAL ESTATE

APARTMENTS PRICES PREDICTION 3.2 Data Attributes
3.1 Methodology
Starting from acquiring the data from online sources as no data
is available for the Jordanian real estate market, as most data
with these specs mostly aggregated from many sources. For
instance, adopting web-scraping approach. Then, this work
proceeds to dealing with data issues, though the actual
workflow was more iterative than sequential it will listed latter
for convenience. After handling the data work, this work
continues to the predictive modelling. Where, several
experiments conducted using several models, namely Linear
and Gradient Boosting for simple and complex
regression-based models. Support Vector Machines and
Random Forests, which are both ensembles and rely on the use Figure 3: Apartment Area in M2 – The Whole Data (Before applying
of decision trees [27], [28], as their weaker models to build the splitting and outlier's removal).
ensembles. Finally, MLP as advanced predictive modeling
technique. Each of the aforementioned predictive modeling The work analysis dataset obtained from an online sales
techniques has its own unique features and their fitting use websites that include listings for apartment for sales in Jordan,
cases. Finally, these models are then fitted with the prepared which is the category this work is going to use. The original
data after manipulation then tuned if necessary to yield data contained 3697 instances with 34 attributes ranging in
acceptable results. type and benefit to the apartment price estimation process.
Several manipulation carried over the data in-order to reach a
4080
Waleed T. Al-Sit et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(3), May – June 2020, 4077 – 4089

state where models are able to capture the trend in the data in in the final outcomes of the data whether it being predictive
order to produce a sane estimation. Then, exploring the data models or simple statistics. Such steps varies from dropping
visually to understand the nature of the data, also it is a NULLs in the main anticipated features to dealing with
necessary for some preprocessing steps, for instance, inconsistent values, duplicates and obvious noise. In addition,
discovering outliers in order to remove and the data to fix data types, correcting, and unifying the values of
distribution to better understand the results obtained from each attributes as the data collected from online sources that are
model. Rather, the visualization for the data can provide open for contribution from any seller and non-seller parties,
valuable insights about the Jordanian real estate market from who wants to make an apartment listing. due to the fair number
with the visual aids as they can substitute even partially for the of instances in the data, and the assuming that the distribution
more defined metrics of the interactions between the attributes and coverage of the data is to be preserved due to size a lot of
of the data [29] including price. messy and unclean data were dropped, that will also be
As mentioned earlier, the obtained data set contains 34 contributing to the speed of the fitting and learning process for
attributes, and usage to those attributes that listed in Table.2 the predictive modelling part.
can varies to conduct different types of analysis other than
3.4 Data Preprocessing
estimating real estate prices. However, the data instances
collected are for advertisements related to apartments in Keeping in mind that the data should be too clean and perfect
specific in overall Jordan for a short period (5, March 2020 to in order to leave room for the models to generalize over unseen
data. Several preprocessing steps done, each serving some
28, April 2020).
purpose, overlapping with preprocessing some transformations
3.3 Data Cleaning applied to the features serving purposes.
Some essential steps are to be done before any further process
should take place to accommodate for the tools to be used
nature and to reduce the bias and errors that may be perceived

Figure 4: Apartment Area in M2 – The Training Data Split (After

applying outlier's removal).

Figure 6: Average of Listed Apartments' Area in M2

Figure 5: Apartment Area in M2 – The Test Data Split (After

applying outlier's removal). Figure 7: Apartments Prices Distribution

4081
Waleed T. Al-Sit et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(3), May – June 2020, 4077 – 4089

Figure 10: Graphical Heat Maps before Outliers Removal.

Figure 8: Price and Area Regression Relationship.

 Filtering none ’Amman’ City Ads: Because of the

imbalance, in advertisements’ listings for cities in Jordan, as
Amman constitute the majority of listings followed by
Al-Aqaba and Irbed. The non-Amman listings were
eliminating the need to deal with the imbalance problem and
focusing the efforts on predicting and exploring Amman’s
listings due to the apparent importance of it visible in the
number of listings as appears in Figure.1.
 Obtaining Listings Coordinates: As previously
mentioned, the location attribute is a string of the address of
Figure 11: Graphical Heat Maps before Outliers Removal.
the apartment in the listing. Although it can be beneficial as
a categorical feature, to allow for a more range of trends
supported by the nature of the dummy variables produced
Table 3: Regression Models Reported Model-Based Score Results
from it. But the required manipulation and the high number Trail Dataset Models Results
of different locations in the data set instances gives an edge Type Generated Train Test
to getting the geographical coordinates (longitude, latitude) Linear Regular Data,
1 0.67 0.71
serving also in better performance in comparison to the Regression Scaled Data
Ensemble
higher dimensionality to be produced by using the one-hot Method –
Regular Data,
1 0.97 0.81
encoding and better serving in continuous and numerical Scaled Data
GB-Regression
models like artificial neural networks and linear regression.
 Processing
Figure 9: Discovering OutliersDate
Advertisement using the
andCorrelation
Building Matrix.
Age:
Though not used in estimating prices, this attribute yields Table 4: Random Forest Best Estimator Searching Techniques
some insights about the listings when coupled with other Trail Dataset Results
attributes indicating what makes a listing stays for long Type Train Test
without it ever sold. For instance, Figure.6 shows the RF-Standard Regular 0.92 0.82
number of advertisements listed per day for the data RF-Randomized Regular 0.92 0.8
collection period. Such result would communicate some Search
indicators for economists about when such market Three Types Used : Standard [12], Grid [16], and Randomized [17]
movement for demands and offers may appear during the
different seasons the study and regulates [30]. Such attribute
transformed from its described form as a categorical and a
date variable to a continuous variable representing the days
since the listing was posted, a similar transformation was

4082
Waleed T. Al-Sit et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(3), May – June 2020, 4077 – 4089

Table 5: SVM Reported Results Table 6: MLP Reported Results

Trail Dataset Results Modeling Models Results
Type Train Test Mechanism Used Generated Train Test
SVM Regular -0.05 -0.04 L-BFGS 1 0.76 0.71
SVM-Normalized (y) Scaled 0.9 0.800.72 (1 Hidden Layer 9
SVM-Randomized Search Scaled 0.84 0.77 Neurons)
ADAM 1 0.71 0.71
(1 Hidden Layer
100 Neurons)
applied to the Building age attribute although on a smaller
L-BFGS 40 0.76 0.7
interval.
(1 Hidden Layer
 Creating Price per Meter Attribute: To be able to fairly 120 Neurons)
judge the listings for further steps like outlier removal, price ADAM 20 0.66 0.67
and area might not be enough as in the case of large (1 Hidden Layer
apartment with a matching price tag and a small apartment 10-200 Neurons)
with a higher than actual value price. ADAM 134 0.93 0.67
 Binary Attributes Creation: Dealing with attributes that (2 Hidden Layers
still include categorical values is not ideal especially for the 10-200 X 10-200
Neurons)
mostly continuous and numeric implementations of machine
learning models. So, and in order to facilitate the use of such
learning models, outliers were detected and removed using 5
attributes (Is Furnished and Payment Type). Each were
attributes (Price ,Area ,Price Per Meter, Longitude And
transformed into two attributes that each can take 0 or 1
Latitude) using the Interquartile Rule for Outliers, while the
allowing for 4 permutations and preserving the other values
z-score method didn’t yield as good results.
of the attributes like the ‘not specified’ or ‘both’, while also
The resulting dataset after the preprocessing is smaller in
being easier to use in computing processes due to its numeric
nature. size (2253 X 26) and uniform in distribution (Train: 1572 X
 Data Splits: The data divided with a ratio of 70\% to 30\% 26, Test: 681 X 26), though the data is still noisy, the early
for training and test sets respectively; this done before further predictive models fitting results indicate an enhancement in
processing and exploring data to reduce bias in the fitting comparison to fits that conducted after. The other attributes
and learning processes, hence, avoid skewing the prediction dropped since not relevancy to the prices predictive analysis
results. The splitting process resulted with the 2044 records conducted. Alongside this type of trimming using quartiles,
for training dataset and 877 records for test dataset another type of trimming using z-score were experimented and
considering the aforementioned preprocessing steps. yields a less satisfactory results. I.e. it does not solve the
 Outlier Removal Using Quartiles: Outliers removed out of outlier’s problem because only one record trimmed when using
necessity based on the distribution observed by the Figures Area in square meter feature. Therefore, we stick with quartiles
3-5. This might hinder the performance of some machine method.
Table 7: Evaluation Measures
Measure Usage Best Value Worst Value
R-squared (R2) Represents proportion of variance of y 100%, the model explain all Lower and Negative Values, i.e.
that explained by independent attributes the variability of the response the model explain nothing about
in the model. Which indicate strength and data around the mean. the response data variability
goodness of fit of the data to the around the mean.
regression line.
Mean Squared Risk metric based on the expected value Lower values as there would Higher values in addition to the
Error (MSE) of the squared error or loss. be excellent match between the lower values with no excellent
[38] actual and predicted dataset. matching between the actual and
the predicted dataset.
Median 1. Outliers robust (unaffected by values at Lower values as there would Higher values in addition to the
Absolute the tails). be excellent match between the lower values with no excellent
Error (MdAE) 2. Error or loss function. actual and predicted dataset. matching between the actual and
[39] 3. Calculate univariate variability. the predicted dataset.
Mean Absolute 1. Risk metric. Lower values as there would Higher values in addition to the
Error (MeAE) 2. Scale-Dependent Accuracy Measure. be excellent match between lower values with no excellent
[40] the actual and predicted matching between the actual and
dataset. the predicted dataset.

4083
Waleed T. Al-Sit et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(3), May – June 2020, 4077 – 4089

Table 8: Experimental Environment HW Specifications Tableby

and removal MLP Models’
10:visualizing theAverage Evaluation
correlation matrixScores
alongside
Item Description Data Split Used ADAM L-BFGS
keeping eyes on the normal distribution for the given attributes
… (1)
and discover that, the early aforementioned … (1) used for
attributes
PC Type Laptop
R2-Train 0.895829998 0.763841224
removing the outliers from the collected dataset are affecting
Brand ASUS ROG G703GX Notebook
R2-Test relationships
clear regression 0.674536562
between each0.705861926
other, which
CPU Intel i9-8950HK clearly depicted in Figure.8. Eventually, leads this work to
MSE-Train 203345745.1 460995310.1
RAM 64GB discover more clear regression relationships between the
MSE-Test 662487173.3 598723783.7
HD 3x1TB NVMe SSD RAID attributes as it not possible to be discovered through the results
MeAE-Train
in Figure.9. 6596.689703 12485.69136
GPU NVIDIA GeForce RTX 2080 8GB
Finally, MeAE-Test
Heat Graphical Maps employed 14651.50066
14093.88889 also in the EDA
Screen 17.3" FHD 144Hz 3ms G-Sync activities in order to show which area got the highest
MdAE-Train 9198.765418 16161.98873
OS Windows 10 Pro advertising density through overall areas in Amman before and
MdAE-Test 18868.66751 18622.71286
after outlier’s removal as presented in the Figure.10 and
Adam = (∑ Score)=155; L-BFGS = (∑ Score)=41 … (1)
Figure.11.
3.5 Visualization and EDA
The main difficulty in here is not coming up with ideas to 4. PREDICTIVE MODELLING EXPERIMENTS
test and evaluate on the dataset; it is coming up with ideas that 4.1 Scaling the Attributes
are likely to turn into insights and valuable indicators about the
Non-binary attributes were scaled using z-score
trends and patterns hidden in the data [31]. In specific, those
normalization to reduce variance and reduce peak values
related to the property business. Accordingly, we have to ask
effects by reducing spread allowing some models to fit the data
questions that lead to deep understanding for the observations
better [32]. The predictor variable, the price in this case was
collected from the online different sources. Therefore,
also scaled to account for the use of distance in SVMs [33],
implementing the concept of Explanatory Data Analysis [29],
using the z-score normalization as well, as was discovered
which is a graphical analysis technique, that employs a variety
through the experiments.
of techniques in order to maximize insights, uncover the
underlying patterns, important features extraction; detect 4.2 Learning Models Employed and Experiments Setup
anomalies and outliers, and further determining the optimal Several types of machine learning models used and
tuned model settings. employed in the comparative process, whilst they all
Adopting EDA mechanism for discovering normal implemented and evaluated according to SCIKIT Learn
distributions for the attributes, or finding the relationships Package [34] provided functionalities on the basis of scaled
between the attributes of different types, namely categorical data and regular data, where the late type means no scaling or
and/or continuous, will better lead to find valuable insights in transformation on the data sets used before learning processes.
the data. For instance, the average area in square meter for the Firstly, one of the simplest machine learning models used in
listed apartments in Amman city was around 150 M2 as addition, evaluated in this work was Linear Regression, which
appears in Figure.6. Whereas the average prices for the better suited for linear data, as the degree of the regressor is
apartments appear to be around 74,647$ as appears in limited to one. It implemented under several assumptions as
Figure.7. the mentioned data linearity, independence and several others,
From another perspective mostly related to outliers detection
Table 9: MLP Top-10 R2-Ranked Models
MLP ID MLP solver Layers Structure R2 MSE MdAE MeAE
# Solver Layer1 Layer2 Train Test Train Test Train Test Train Test
168 LBFGS 7 0 0.74 0.7 510320333 603349036 13284.7 14313.1 17158.3 18909.7
155 LBFGS 6 0 0.74 0.72 512527565 571225948 13098.8 14093.8 16984 18154.8
135 LBFGS 5 0 0.74 0.72 515600513 576319329 13741.5 15218.9 17261.6 18609.6
136 LBFGS 5 0 0.73 0.72 521154916 574018441 13071.1 14527.5 17213.6 18367.5
85 ADAM 180 0 0.73 0.73 522058827 554305503 13099.4 13982.5 17178.6 17895.2
87 ADAM 190 0 0.73 0.72 530135262 563064858 12943.2 13909.9 17294.4 17992.9
92 ADAM 200 0 0.73 0.72 530948320 561928876 13009.8 13994.8 17334.2 17987.2
81 ADAM 160 0 0.73 0.72 535067377 572433080 13245.1 14357.1 17425.3 18199.8
75 ADAM 130 0 0.73 0.72 535396669 564242215 13326.5 14501.8 17366.4 18064.9
79 ADAM 150 0 0.73 0.72 536741651 568369222 13220.1 14319.3 17422.7 18132.1

4084
Waleed T. Al-Sit et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(3), May – June 2020, 4077 – 4089

is usually expressed as minimizing error/residual or some

transformation of it, to be more precise the linear regression in
this case where multiple explanatory variable exist it is called
multiple linear regression [35]. Next Ensemble learning based
method – Gradient Boosting Regression (GB-Regression) [27]
analysis conducted on the features sets in terms of the regular
data and the scaled data. The configurations for such model
include 400 estimator trees with number of five levels for the
max depth and minimum two samples required to split a node
at learning rate 0.1 with loss function to be optimized namely
least squares regression. The results yield over fitted model in
acceptable range according to the standard score function Figure 12: Random Forest 500 Estimators Overfitting [Training:
included in the model implementation, whilst higher results 0.96, Test: 0.82].
when considering R2 (R-Squared) that is coefficient of
determination provided by the same learning model scored results.
implementation [34]. However, used regression models’ The best estimator configurations [34] founded for Random
alongside the standard model-based evaluation [36] scores Forest model were 1500 estimators with max depth 100 and
yielded on the scaled data where summarized in the Table.3. minimum sample leaves of 2 and 6 minimum samples splits
Another Ensemble learning based method employed with evaluated using the criterion MSE [38].
almost the same number of estimators set before for On the other hand, Non-linear SVMs [33] to aim to fit a
GB-Regression is Random Forest [37]. The analysis conducted more complex and non-linearly separable data using the RBF
on the features sets in terms of the regular data, yielded worst Kernel conducted on the features sets in terms of the regular
results in terms of model overfitting for training data with no and scaled data. Results yielded bad results with no any
any significance changes in the prediction results based on the significance changes in the prediction results to overcome the
standard model score function, the results reported in previous trails and models for the regular data, whilst it
Figure.12. indicated that SVMs perform better with the scaled data as it
Furthermore, investigations following acquiring the uses distances, so a drastic difference in distances will cloud
overfitting results from the previous trail, lead in initiating two the finer trends in the classifier (or used regression model).
search processes to figure out the best configurations for the Experiments shows the overfitting is high due to the difference
decision tree, i.e. what are the best attributes or parameters for for the R2 score between the training and the test datasets, yet
the Random Forest method to better fit the data and create a its acceptable range for the model. Leading the work to
good price estimator on unseen data. Total 450 runs for the continue its way toward exhaustive figuring out the
different configurations provided that could divided into three best-optimized model’s hyper parameters. Nevertheless, the
similar runs of each configuration based on a different data reported results show minor improvement over the almost
split (fold). The reported results show minor improvement over default scaled model (overfitting went down, hence better
the almost default model, and this can be judged due to the results for the model. i.e. more generalized model). The results
theory of the ensemble methods that several models shall be reported based on the standard R2 score function for this stage
better than one. To provide more convenience, the results in the Table.5.
acquired upon using such learning methodology depicted
clearly in Table.4 using the standard model-based evaluation

Table 11: Other Models’ Evaluation Scores

Model Type R2 MSE MdAE MeAE
Train Test Train Test Train Test Train Test
Linear Regression 0.67 0.71 637955720.47 587972332.82 15565.15 15893.58 19509.84 19227.21
GB-Regression 0.98 0.82 43477365.96 371593383.38 2510.74 9485.87 4220.98 13674.42
SVM 0.90 0.72 0.10 0.30 0.01 0.30 0.16 0.39
SVM – Best Estimator 0.84 0.77 0.16 0.24 0.12 0.28 0.25 0.36
Random Forest 0.97 0.82 67627086.20 360681397.17 3925.50 9529.00 5671.71 13455.18
Random Forest – 0.92 0.82 146483386.29 367440117.15 5719.66 9973.90 8415.14 13727.93
Best Estimator

4085
Waleed T. Al-Sit et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(3), May – June 2020, 4077 – 4089

since all the trained and tested models backed-up on external

files, providing the chance to try different evaluation measures,
and assessing the conduction of the final models evaluation
and selection for both training and testing datasets. For
instance, measures depicted in Table.7 are the chosen criteria
to amend this comparative work with the evaluation results
needed.
5.2 Predictive Modeling Evaluation
Considering the evaluation criteria shown early in this work,
in here, listing for the evaluation results will take place, where
it will be organized in showing the MLP results first and the
best evaluated architectures followed by the discussions
related, then the rest of the models and the best accuracy
achieved.
Figure 13: MLP Evaluated Models Using R2. Starting with MLP trained and tested models. The resulted
evaluation scores shown in Table.9 depict that L-BFGS solvers
Finally, this work continuous to the most complex learning out performs well over the Adam solvers in terms of overfitting
methods. Usually an MLP or a neural model should exceed on smaller datasets as all of them ranked in terms of R2
other models in performance. Given that, if it tuned well, a evaluation measure for the top ten the trained and the twisted
process that is high in cost, and could be infeasible considering MLP models for both solvers.
the way simpler models can give satisfactory results for some
problems. Here we tried two types of solvers (ADAM and The best achieved results for the MLP models that got the
L-BFGS), each is different in degree and where it excels. highlighted architectures where all have one layer with number
However, the most complicated parameter to tune is the of neurons less than or equal to 180 neurons for the first. One
architecture size, due to it being not a single parameter and the hundred sixty nine different models trained on longer periods
high dynamicity of it. That was addressed by fitting around 196 measured in hours, all evaluated and the best accuracy
MLPs different in architecture and solvers while sticking with achieved in terms of R2 measure was (0.72) for both training
mostly default parameters except for the activation where we and test data sets. Further results shown by the MSE measures
went with the ReLU as we're not constrained in computing that depict training got loss score greater than test score with a
power to a degree that we should try something simpler as the little difference on unseen data, yet still acceptable results for
ReLU provides the flexibility in activation and scaling. Each the models. Another interested result for MLP models
model where allowed to go up to 100000 epochs while evaluated related to the MdAE and MeAE measures that
constrained with the tolerance over the enhancement in loss appeared to measure the well fit considering the MdAE’s
over a validation split of 20\% of the training data. Exhaustive robustness for the outliers, and this proof the correctness of the
trails and stochastic optimizations applied to the learning methodology adopted in this work for removing the outliers.
models, whilst each got an ID and exported to external Whilst the Table.10 shows the average results for each
serializable files for further analysis, evaluation and future evaluation, measure used for different MLP solvers among the
usage, for instance Figure.13. For convenience, all the used trained and tested models. It appears clearly that L-BFGs
MLP models’ standard model-based evaluation [34] scores solvers perform well on smaller datasets while Adam solvers
yielded on the scaled data where summarized in the Table.6. suffer from learning overfitting for such type of datasets. Other
models shown in Table.11, where Random Forest ensemble
4.3 Experiential Machine Used method, which used for conducting complicated predictive
All the analysis early mentioned and the results for all analysis for the given data. Evaluation results using standard
experiments conducted on machine that got the hardware score function for random forest trained and tested models on
specifications as depicted in Table.8. regular data basis, i.e. data not transformed or scaled, shows
clear overfitting which listed early in this work early in
5. RESULTS AND DISCUSSIONS Figure.12. The initiated two search strategies, grid and
5.1 Predictive Modeling Evaluation Criteria random, where grid search used to test configurations on lower
searching rates, which then supplement random search for the
Fitting time dismissed here as apparent by the experiments of
sake of figuring out best configurations for the random forest
the MLP and the grid search optimization for the SVM and
algorithm. This process resulted in number of estimators 1500
Random Forest. While the R-squared was the main evaluation
of max depth of 100; error function MSE for minimum samples
criterion through the experiments. Other common regression
split and leafs of six, and two respectively as all the best prices
evaluation metrics added later on like mean squared error,
estimator configurations. Even the primary evaluation results
median absolute error and mean absolute error. Moreover,
4086
Waleed T. Al-Sit et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(3), May – June 2020, 4077 – 4089

shows minor improvement over the default model, but the R2 inside the local and/or global markets different areas, trying to
remained showing higher deviation (i.e. Overfitting) achieved provide insights about, enrich, and curate the local and/or
as MLP modeling techniques used for random configuration global economic aspects, which had better enhance the
lookup. These results can attributed to either low instances understanding and aid specialists to draw insightful
searched and/or the efficiency of the default models to begin conclusions about the market different indicators using
with on smaller dataset. different data science techniques
Finally, Linear and GB-Regression evaluated well using R2
score function with a small indication for minor trendy model REFERENCES
overfitting. Whilst the non-linear SVMs using RBF Kernel 1. M. Hoesli, C. Lizieri, and B. MacGregor. The inflation
yielded in two very different standard model score-based hedging characteristics of US and UK investments: a
results, giving an indication that SVMs perform better with the multi-factor error correction approach. The Journal of
scaled data as it uses distances, so a drastic difference in Real Estate Finance and Economics, 36(2), 183-206.
distances will cloud the finer trends in the classifier (or used 2008.
regression model). Rather, best estimator configurations https://doi.org/10.1007/s11146-007-9062-6
founded shows minor improvement over the default model 2. S. J. A. P. Zavei, and M. M. Jusan. Exploring housing
even reporting with R2 evaluation metric, which shows better attributes selection based on Maslow's hierarchy of
results in the other models employed. needs. Procedia-Social and Behavioral Sciences, 42,
311-319. 2012.
6. CONCLUSION 3. S. Hudson-Wilson, F. J. Fabozzi, and J. N. Gordon,. Why
real estate?. The Journal of Portfolio Management, 29(5),
Not all the previously mentioned aspects and benefits to 12-25. 2003.
work on such problem changed the state of literature and the 4. J. Aizenman, and Y. Jinjarak. Real estate valuation,
work on the Jordanian market. Whilst the scarcity of such current account and credit growth patterns, before
works imposes a need for it, and demonstrates an unutilized and after the 2008–9 crisis. Journal of International
aspect of the economy that requires little resources to create Money and Finance, 48, 249-270. 2014.
5. A. Anari, and J. Kolari. House Prices and InflationReal
some business and academic opportunities. In addition,
Estate Economics, 30(1), 67–84. 2002.
helping in further understanding of Jordanian economy, and
6. K. E. Case, E. L. Glaeser, and J. A. Parker. Real estate
potentially diagnosing some problems and recommending
and the macroeconomy. Brookings Papers on Economic
actions to be taking to revert the declination of that economy
Activity, 2000(2), 119-162. 2000.
and identify where correction should be concentrated. Going 7. S. B. Billings. Hedonic amenity valuation and housing
further to cover all the potential inductions and renovations. Real Estate Economics, 43(3), 652-682.
recommendations amongst other possible extractions and 2015.
solutions to this problem start with methodological data https://doi.org/10.1111/1540-6229.12093
science’s use case. The use case shown is offering a 8. N. Shinde, and K. Gawande. Survey on predicting
comparative study of several predictive and descriptive property price. In 2018 International Conference on
methods over an online collected data of apartments for sale in Automation and Computational Engineering (ICACE)
Jordan and their prices alongside listed features. The focus of (pp. 1-7). IEEE. October 2018.
the work concentrate mainly on data mining and machine 9. B. Sopranzetti, Hedonic Regression Models. pp.
learning different techniques. Some anecdotal insights were 2119–2134 10 1007 978–1–4614–7750–1 78.)
listed that relates to the apartments market and on the 10. D. Harrison Jr, and D. L. Rubinfeld. Hedonic housing
techniques used in this work, where the experimental results prices and the demand for clean air. 1978.
might indicate some relations, the consolidation of the 11. J. Mu, F. Wu, and A. Zhang. Housing value forecasting
observed interrelations still needs a more thorough study with based on machine learning methods. In Abstract and
more comprehensive tests on bigger datasets. However, the Applied Analysis (Vol. 2014). Hindawi. January 2014.
work focuses on exploring some of the attributes interactions 12. J. M. N. Tabales, J. M. Caridad, and F. J. R. Carmona, F.
while preparing them for the predictive modelling, and J. R. Artificial neural networks for predicting real
reshaping, transforming and filtering the data for a better estate price. Revista de Métodos Cuantitativos para la
learning by addressing some of basic data problems like Economía y la Empresa, 15, 29-44. 2013.
distribution, outliers and incompatible attribute types with 13. E. D. Sontag. Feedback stabilization using
two-hidden-layer nets. In 1991 American Control
some models. Then and based on a selected subset of machine
Conference (pp. 815-820). IEEE. June 1991.
learning models that vary in complexity and behavior fitted
14. Y. Li, and Y. Yuan. Convergence analysis of two-layer
with while tuning both the models and data in atrial to enhance
neural networks with relu activation. In Advances in
performance, judging based on several criteria (R2, MSE,
neural information processing systems (pp. 597-607).
MdAE, and MeAE) and trying to explain and evaluate the final 2017.
results obtained. The next steps for this work will go deeply

4087
Waleed T. Al-Sit et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(3), May – June 2020, 4077 – 4089

15. R. Manjula, S. Jain, S. Srivastava, and P. R. Kher. Real 29. M. F. De Oliveira, and H. Levkowitz. From visual data
estate value prediction using multivariate regression exploration to visual data mining: A survey. IEEE
models. In IOP Conference Series: Materials Science and transactions on visualization and computer graphics, 9(3),
Engineering (Vol. 263, p. 042098). November 2017. 378-394. 2003.
16. A. A. Neloy, H. S. Haque, and M. M. Ul Islam. Ensemble 30. S. Lima, A. M. Gonçalves, and M. Costa. Time series
learning based rental apartment price prediction forecasting using Holt-Winters exponential
model by categorical features factoring. In Proceedings smoothing: An application to economic data. In AIP
of the 2019 11th International Conference on Machine Conference Proceedings (Vol. 2186, No. 1, p. 090003).
Learning and Computing (pp. 350-356). February 2019. AIP Publishing LLC. December 2019.
https://doi.org/10.1145/3318299.3318377 https://doi.org/10.1063/1.5137999
17. A. Singh, and R. Lakshmiganthan. Impact of different 31. C. Lee, O. Kwon, M. Kim, and D. Kwon. Early
data types on classifier performance of random forest, identification of emerging technologies: A machine
naive bayes, and k-nearest neighbors algorithms. learning approach using multiple patent indicators.
2018. Technological Forecasting and Social Change, 127,
18. V. Limsombunchai. House price prediction: hedonic 291-303. 2018.
price model vs. artificial neural network. In New 32. R. Berwick. An Idiot’s guide to Support vector
Zealand agricultural and resource economics society machines (SVMs). Retrieved on October, 21, 2011. 2003.
conference (pp. 25-26). June 2004. 33. A. Abanda, U. Mori, and J. A. Lozano. A review on
19. G. J. McKee, and D. Miljkovic. Data Aggregation and distance based time series classification. Data Mining
Information Loss (No. 381-2016-22080). 2007. and Knowledge Discovery, 33(2), 378-412. 2019.
20. H. YILDIRIM. Property Value Assessment Using 34. F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B.
Artificial Neural Networks, Hedonic Regression and Thirion, O. Grisel, ... and J. Vanderplas. Scikit-learn:
Nearest Neighbors Regression Methods. Selçuk Machine learning in Python. The Journal of machine
Üniversitesi Mühendislik, Bilim ve Teknoloji Dergisi, Learning Research, 12, 2825-2830. 2011.
7(2), 387-404. 2019. 35. R. A. Bottenberg, and J. H. Ward. Applied multiple
21. M. Miyamoto, and H. Tsubaki. Measuring technology linear regression (Vol. 63, No. 6). 6570th Personnel
and pricing differences in the digital still camera Research Laboratory, Aerospace Medical Division, Air
industry using improved hedonic price estimation. Force Systems Command, Lackland Air Force Base.
Behaviormetrika, 28(2), 111-152. 2001. 1963.
22. H. Xu, and F. Mueller. Work-in-progress: Making 36. E. W. Steyerberg, A. J. Vickers, N. R. Cook, T. Gerds, M.
machine learning real-time predictable. IEEE Gonen, N. Obuchowski, ... and M. W. Kattan. Assessing
Real-Time Systems Symposium (RTSS) (pp. 157-160). the performance of prediction models: a framework
IEEE. December 2018. for some traditional and novel measures. Epidemiology
23. N. H. Abroyan, and R. G. Hakobyan. A review of the (Cambridge, Mass.), 21(1), 128. 2010.
usage of machine learning in real-time systems. 37. L. Breiman. Random Forests. Machine learning, 45(1),
Bulletin of the National Polytechnic University of 5–32. 2001.
Armenia. Information Technology, Electronics, Radio 38. N. J. Nagelkerke. A note on a general definition of the
Engineering, (1), 46-54. 2016. coefficient of determination. Biometrika, 78(3),
https://doi.org/10.22606/fsp.2017.12002 691-692. 1991.
24. R. Bellazzi, C. Larizza, P. Magni, S. Montani, and G. De 39. Z. Wang, and A. C. Bovik. Mean squared error: Love it
Nicolao. Intelligent analysis of clinical time series by or leave it? A new look at signal fidelity measures.
combining structural filtering and temporal IEEE signal processing magazine, 26(1), 98-117. 2009.
abstractions. In Joint European Conference on Artificial 40. C. J. Willmott and K. Matsuura. Advantages of the mean
Intelligence in Medicine and Medical Decision Making absolute error (MAE) over the root mean square error
(pp. 261-270). Springer, Berlin, Heidelberg. June 1999. (RMSE) in assessing average model performance.
25. X. Chen, L. Wei, and J. Xu. House price prediction Climate research, 30(1), 79-82. 2005.
using LSTM. arXiv preprint arXiv:1709.08432. 2017. 41. R. Tripathi, and D. P. Rai. Comparative Study of
26. E. Ahmed, and M. Moustafa. House price estimation Software Cost Estimation Technique. International
from visual and textual features. arXiv preprint Journal of Advanced Research in Computer Science and
arXiv:1609.08399. 2016. Software Engineering, 6(1). 2016.
27. F. Zhang, B. Du, and L. Zhang. Scene classification via a 42. V. Kale, and F. Momin. Video Data Mining Framework
gradient boosting random convolutional network for Surveillance Video. International Journal of
framework. IEEE Transactions on Geoscience and Advanced Trends in Computer Science and Engineering,
Remote Sensing, 54(3), 1793-1802. 2015. 2(3). 2013.
28. L. Breiman. Using iterated bagging to debias 43. A. A. Mahule, and A. J. Agrawal. Hybrid Method for
regressions. Machine Learning, 45(3), 261-277. 2001. Improving Accuracy of Crop-Type Detection using
Machine Learning. International Journal, 9(2). 2020.

4088
Waleed T. Al-Sit et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(3), May – June 2020, 4077 – 4089

https://doi.org/10.30534/ijatcse/2020/209922020
44. M. Akour, O. Al Qasem, H. Alsghaier, and K.
Al-Radaideh. The effectiveness of using deep learning
algorithms in predicting daily activities. International
Journal, 8(5). 2019.

4089

View publication stats