i
EVALUATION OF SPATIAL DATA’S IMPACT IN MID-TERM
ROOM RENT PRICE THROUGH APPLICATION OF SPATIAL
ECONOMETRICS AND MACHINE LEARNING
Case S udy: Lisbon
Mihail Pe ko
ii
EVALUATION OF SPATIAL DATA’S IMPACT IN MID-TERM
ROOM RENT PRICE THROUGH APPLICATION OF SPATIAL
ECONOMETRICS AND MACHINE LEARNING
Case S udy: Lisbon
Disse a ion supe ised by:
P o esso Dou o Robe o And é Pe ei a Hen iques, PhD
Ins i u o Supe io de Es a ís ica e Ges ão de In o mação,
Uni e sidade NOVA de Lisboa
Lisbon, Po ugal
Joel Dinis Bap is a Fe ei a da Sil a, PhD
Ins i u o Supe io de Es a ís ica e Ges ão de In o mação,
Uni e sidade NOVA de Lisboa
Lisbon, Po ugal
P o esso Dou o Ca los G anell-Canu , PhD
Ins i u e o New Imaging Technologies,
Uni e si a Jaume I (UJI)
Cas ellón de la Plana, Spain
Feb ua y 28, 2020
iii
DECLARATION OF ORIGINALITY
I decla e ha he wo k desc ibed in his documen is my own and no om
someone else. All he assis ance I ha e ecei ed om o he people is duly
acknowledged and all he sou ces (published o no published) a e
e e enced.
This wo k has no been p e iously e alua ed o submi ed o NOVA
In o ma ion Managemen School o elsewhe e.
Lisbon, 28.02.2020
Mihail Pe ko
i
ACKNOWLEDGMENTS
An idea abou making sense o economic phenomena h ough he lenses o
geog aphical in o ma ion ha e always been a esea ch opic o in e es o
me. Though esea ch and he esul s’ a e ma h a e o en imes
unp edic able, he people ha guided me and suppo ed me h oughou his
p ocess we e o una ely no .
I wan o i s and o emos hank my main-supe iso P o esso D . Robe o
Hen iques ha was open o hea ing my ideas and helping me cons uc hem
in some hing p ac ical, o P o esso D . Joel Sil a who was always he e o
sugges imp o emen s and guide me h ough new di ec ions and doo s when
all I was seeing we e dead ends, o P o esso D . Ca los G anell o
belie ing in me, and las , bu no leas , o P o esso D . Ma co Painho o
keeping me le el-headed o push and p og ess un il I ha e eached he inish
line.
Addi ionally, my g ea es g a i ude goes o all colleagues om he
Geospa ial Technologies Mas e as a whole, wi h special shou -ou s o my
iend and la -ma e I za, o Damien, Ca los, B aund , Vicen e and Pablo.
Las bu no leas , I would like o hank all my closes amily and iends in
Macedonia. I am ex emely g a e ul o all he suppo , pa ience and
encou agemen p o ided du ing he jou ney.
EVALUATION OF SPATIAL DATA’S IMPACT IN MID-TERM
ROOM RENT PRICE THROUGH APPLICATION OF SPATIAL
ECONOMETRICS AND MACHINE LEARNING
Case S udy: Lisbon
ABSTRACT
Household p e e ences is a opic whose ele ance can be ound o domina e he applied
economics, bu whe eas u ban economies iew ci ies as p oduc ion cen e s, his hesis
aims o gi e impo ance o he ole o consump ion. P o ision o PoIs migh gi e
explana ion o wha indi iduals alue as an impo an asse o imp o emen o hei
quali y o li e in a chosen ci y. As such, unde s anding sho - e m en als and eal es a e
p ices ha e induced a ious esea ch o seek p oo o impac ing ac o s, bu analysis o
mid- e m en has aced he challenge o being an o e looked ca ego y. This hesis
consis s o an in eg a ed h ee-s eps app oach o analyze spa ial da a’s impac o e he
mid- e m oom en , choosing Lisbon as i s case s udy. The p oposed me hodology
cons i u es use o adi ional spa ial econome ic models and SVR, encompassing a
la ge se o p oxies o ameni ies ha migh be ecognized o hold a possible impac
o e en p ices. The analy ical amewo ks’ i s s ep is o c ea e a sui able HPM model
ha cap u es he da a well, so signi ican a iables can be de ec ed and analyzed as a
disc e e da ase . The second s ep applies subse s o he da ase in he c ea ion o SVR
models, in hopes o iden i ying he SVs in luencing p ice a iances. Finally, SOM
clus e s a e chosen o add ess whe he mo e na u al o de o da a di ision exis s. Resul s
con i m he impac o p oximi y o a ious ca ego ies o ameni ies, bu he en ichmen
o models wi h he p oposed p oxies o spa ial da a ailed o co obo a e a ainmen o
model wi h a highe accu acy.
(Nüs e al., 2018) p o ides a sel -assessmen o he ep oducibili y o esea ch, and
acco ding o he c i e ia gi en, his disse a ion is e alua ed as: 0, 2, 1, 2, 2 (inpu da a,
p ep ocessing, me hods, compu a ional en i onmen , esul s).

i
KEYWORDS
P edic i e Modeling
Ameni ies
Suppo Vec o Reg ession
Spa ial Econome ics
Hedonic P ice Modelling
Poin s o In e es
Mid-Te m Ren
ii
ACRONYMS
AIC – Akaike In o ma ion Re ie al
ANN – A i icial Neu al Ne wo k
API – Applica ion P og amming In e ace
BGRI - Base Geog á ica de Re e enciação de In o mação
BMU – Bes -Ma ching Uni
CCDR – Comissão de Coo denação e Desen ol imen o Regional
CML – Câma a Municipal de Lisboa
CPI – Consume P ice Index
DoF – Deg ees o F eedom
DTM – Digi al Te ain Model
GIS – Geog aphic In o ma ion Sys ems
HPM – Hedonic P ice Modelling
MAE – Mean Absolu e E o
MD – Minimum Dis ance
MLP – Mul i-Laye Pe cep on
MMH – Maximum Ma gin Hype plane
OSM – OpenS ee Map
POI – Poin o In e es
RMSE – Roo Mean Squa e E o
SD – S anda d De ia ion
SDEM – Spa ial Du bin E o Model
SLX – Spa ially Lagged X
SOM – Sel -O ganizing Map
SQL – S anda dized Que y Language
SVM – Suppo Vec o Machine
SVR – Suppo Vec o Reg ession
iii
INDEX OF THE TEXT
ACKNOWLEDGMENTS ....................................................................................................... IV
ABSTRACT .............................................................................................................................. V
KEYWORDS .......................................................................................................................... VI
ACRONYMS ......................................................................................................................... VII
INDEX OF TABLES ................................................................................................................ X
INDEX OF FIGURES ............................................................................................................. XI
1. INTRODUCTION .................................................................................................................. 1
1.1 THE MARKET AND ITS FLUCTUATIONS ............................................................................. 1
1.2 FACTORS DRIVING A MARKET – PREDICTIVE MODELLING .............................................. 2
1.3 RESEARCH OBJECTIVES ..................................................................................................... 3
1.4 DISSERTATION STRUCTURE ............................................................................................... 3
2. THEORETICAL FRAMEWORK AND RELATED WORKS ............................................. 4
2.1 THEORETICAL BACKGROUND ............................................................................................ 4
2.1.1 A i icial In elligence and Machine Lea ning ........................................................... 4
2.1.2 Spa ial Econome ics ................................................................................................ 4
2.1.3 Suppo Vec o Reg ession ........................................................................................ 6
2.1.4 Sel -O ganizing Map and GeoSOM .......................................................................... 8
2.2 RELATED WORK ................................................................................................................ 9
2.2.1 Resea ch h ough Applica ion o Spa ial Econome ics ........................................... 9
2.2.2 Resea ch h ough Applica ions o SVR ....................................................................11
2.2.3 Resea ch h ough Usage o O he Algo i hms .........................................................11
3. DATA AND METHODOLOGY ..........................................................................................12
3.1 GEOGRAPHICAL CONTEXT ...............................................................................................12
3.2 PROPOSED ARCHITECTURE ...............................................................................................13
3.3 HARDWARE AND SOFTWARE ............................................................................................13
3.3.1 Pos g eSQL ..............................................................................................................13
3.3.2 R ...............................................................................................................................14
3.3.3 Py hon ......................................................................................................................14
3.4 DATA DESCRIPTION AND RESOURCES ..............................................................................15
3.4.1 Room Ren al Sou ces ...............................................................................................15
3.4.2 Zoma o API Res au an Da a Collec ion .................................................................15
3.4.3 Google Places API Da a Collec ion .........................................................................16
3.4.4 OpenS ee Map Da a ...............................................................................................16
3.4.5 Census Da a .............................................................................................................18
3.4.6 Ancilla y Da a ..........................................................................................................18
3.4.7 O he Va iables Impo ance – Access Limi a ions...................................................19
3.5 METHODOLOGY ................................................................................................................19
3.5.1 Va iables Con e sion ...............................................................................................20
3.5.2 Spa ial Da a En ichmen ..........................................................................................21
3.5.3 Da a P ep ocessing ..................................................................................................24
3.5.4 Spa ial Econome ics and Machine Lea ning ..........................................................27
ix
4. RESULTS AND DISCUSSION ...........................................................................................28
4.1 SPATIAL DEPENDENCE .....................................................................................................28
4.2 SPATIAL LAGGED X ..........................................................................................................29
4.3 SPATIAL DURBIN ERROR MODEL .....................................................................................33
4.4 SUPPORT VECTOR REGRESSION .......................................................................................38
4.5 NATURAL DATA CLUSTERS USING SOM AND GEOSOM .................................................43
4.6 OVERVIEW OF GIVEN ANALYSIS ......................................................................................46
5. CONCLUSIONS ...................................................................................................................47
6. BIBLIOGRAPHIC REFERENCES ......................................................................................49
7. ANNEX .................................................................................................................................56
5
O dina y Leas Squa es (OLS) is a simple HPM eg ession ha akes he o m o
equa ion (1):
𝒚𝒚𝒊𝒊=𝛂𝛂+ 𝜷𝜷𝟏𝟏𝒙𝒙𝟏𝟏+𝜷𝜷𝟐𝟐𝒙𝒙𝟐𝟐+ 𝜺𝜺𝒊𝒊 (𝟏𝟏)
Whe e yi would be he i h obse a ion o he dependen a iable, x is a ec o o all
explana o y a iables, and α is an in e cep ha holds he alue o y when all
explana o y a iables a e equal o 0. 𝛽𝛽 is he ec o o he co esponding coe icien s o
he es ima es p edic o s and 𝜺𝜺 is he andom e o .
When dealing wi h gi en da a, he OLS model has ce ain assump ions, whose iola ion
would lead o an inapp op ia e model. One o he assump ions speci ies ha inding
spa ial au oco ela ion in he esiduals o he p edic ions shows ha he e is a spa ial
pa e n in he da a unde analysis and mo e app op ia e models exis ha can accoun
o i .
Spa ial Lagged X (SLX) po ays a spa ial eg ession model ha ex ends o include
explana o y a iables obse ed on neighbo ing c oss-sec ional uni s e ms. In o de o
imp o e he model, a spa ial exogenous au o eg essi e lag is in oduced (Elho s &
Halleck Vega, 2017). The equa ion (2) would hen ake he o m o :
𝒚𝒚𝒊𝒊𝒊𝒊 =𝐖𝐖𝒙𝒙𝒊𝒊𝒊𝒊 +𝐗𝐗𝒊𝒊𝒊𝒊𝜷𝜷+𝜺𝜺𝒊𝒊𝒊𝒊 (𝟐𝟐)
Then, 𝐗𝐗𝒊𝒊𝒊𝒊 ep esen s he au o eg essi e coe icien ha p oduces he global spa ial
spillo e e ec . I his comes o be equal o ze o, he model simpli ies o a con en ional
linea eg ession model (Lesage, 2008). Ne e heless, his app oach has i s d awbacks
because assump ions a e made ha all he spa ial dependence among esiduals is caused
by a iables one can measu e.
Spa ial Du bin Model is an ex ension o SAR ha includes he a e age o he a iables
om he neighbo ing houses, bu i ex ends o allow o global di usion o shocks o
he model dis u bances. This means ha he spillo e can be el o e mul iple egions,
bu he decay pa ame e λ con ols ha highe -o de neighbo s (indi ec ones) o ecei e
less impac han he di ec ones (Lesage, 2014). Addi ionally he E o addi ion wi hin
his model c ea es he SDEM model ha accoun s o spa ial dependence among he
e o e ms and edis ibu es his e o among he sample. The equa ion in ec o o m
akes he equa ion (3) (LeSage & Pace, 2009):
𝒚𝒚𝒊𝒊𝒊𝒊 =𝐃𝐃𝒊𝒊𝒊𝒊ծ+𝐖𝐖𝒚𝒚𝒊𝒊𝒊𝒊𝝆𝝆+𝐗𝐗𝒊𝒊𝒊𝒊𝜷𝜷+𝑾𝑾𝑾𝑾𝑾𝑾𝑾𝑾 (𝟑𝟑)
𝑾𝑾= 𝝀𝝀𝑾𝑾𝑾𝑾+𝝐𝝐
Dummy Va iables a e bina y a iables ha equal ‘1’ i he a ge a iable is desc ibed
o ei he , con ain some hing (e.g. e ace wi h seaside iew), o i belongs somewhe e
(e.g. an apa men belongs o he pa ish o Al ama). ‘0’ hen desc ibes he opposi e case.
Each o he a iables can ha e a posi i e o a nega i e ela ionship wi h he p ice. In an
ideal case, minimum dis ance o an ameni y makes a p ice ise. Ne e heless, he ma ke
can beha e unp edic ably o in e ac wi h o he a iables o yield a nega i e
ela ionship. Fo example, a en o oom in an apa men whe e he a io o numbe o

6
bed ooms di ided by numbe o ba h ooms is high migh yield a nega i e e ec .
Addi ionally, ha ing a minimum dis ance o clubs below 20 me e s migh dissuade
po en ial enan s away om a p ope y.
Di ec e ec : Change o a pa icula en i y o a lis ing (e.g. eno a ion o apa men
om 1979 o 2019) changes he p ice a ge a iable.
Indi ec e ec : Measu es he impac on he p ice o a a ge a iable om changing an
exogenous a iable in ano he da a poin .
Weigh Ma ix - A ma ix W is a ma ix o o de n x n whe e n is he numbe o egions.
Non-ze o elemen s in ow i and column j in he ma ix hen ep esen he elemen ni is
a neighbo o elemen nj. Rook and Queen each e e o wo common ways o calcula e
s a is ics o ocal cells, and hese a e known as Moo e’s and Neumann’s
neighbo hoods. These spa ial weigh s a e calcula ed, such ha each elemen in a ma ix
ep esen s a spa ial weigh be ween wo neighbo s. The spa ial weigh s Wij a e non-ze o
when i and j a e neighbo s, o ze o o he wise. The di e ence be ween he wo ypes o
neighbo hood is ha in Rook con igui y, common e ices do no make wo polygons -
neighbo s. The use o queen neighbo hood was he chosen app oach in o de o e lec
eal-li e phenomena whe e i wo neighbo hoods mee a a common co ne , hey would
s ill cause in luence be ween one-ano he .
Figu e 2.1 - Types o Weigh ed Neighbou hoods
2.1.3 Suppo Vec o Reg ession
SVM o igina ed om an algo i hm implemen a ion om 1995 by Co es and Vapnick
(Co es & Vapnik, 1995). The algo i hm had been i s in oduced in o de o sol e
pa e n ecogni ion p oblems. SVM is desc ibed as an algo i hm ha makes use o a
hype plane cons uc ed in a highe dimensional inpu space ha sepa a es he da a in
his high dimensional space op imally (Suykens & Vandewalle, 1999). Ne e heless,
wha began as a classi ica ion p oblem, ose o applica ion o sol ing unc ion
es ima ion p oblems. These indings p oceeded o addi ional inclusion o op ional
pa ame e s like epsilon’s loss unc ion and in oduc ion o di e en ypes o ke nel
unc ions in which da a can be mo e easily di ided. This way, he e could be di e en
solu ions p oduced wi h di e en complexi ies and di e en h esholds o e o s in
o de o build he mos cos -e icien model ha could be deemed a good enough i o
he pu pose i had been buil o .
7
The e a e in ini e possible ways o cons uc a hype plane, and in classi ica ion, he bes
hype plane can be de e mined as he one ha holds he la ges dis ance be ween
hype planes and he suppo ec o s. Subsequen ly, de ec ion o SVs and Lag ange
Mul iplie s demons a e he eco ds impo an o building he model (Wi en, Pal, &
Fou h, 2017).
Epsilon wi hin SVR
Wi h gi en aining da a {(x1, y1) …. (xe, ye)} whe e X ⊂ X × R, whe e X deno es he
space o he inpu pa e ns, SVR is looking o a unc ion (x) ha would ha e a mos
ε de ia ion om he ac ual a ge s o all aining da a, whils main aining he la ges
la ness (Smola & Schölkop , 2004). A linea i o a eg ession unc ion is p esen ed
in Figu e 2.2:
Figu e 2.2 - Tube wi h adius ε (Schölkop , 2002)
I all da a poin s i wi hin he ube p esen ed, he i ed model hen has an e o 0. In
his way, a model can be buil wi h an in ini ely high alue ε. This would lead o no
penal y gi en o any da a poin and an e o o 0, bu his would be a meaningless
accu acy measu e.
In a linea case he SVR unc ion akes he o m o Equa ion (4)
𝑋𝑋 = 𝑏𝑏 + ∑i α a(i)⋅a (𝟒𝟒)
The do no ion can be subs i u ed by a ious ke nel unc ions i a non-linea p oblem is
p esen ed. Mos common ke nels used a e:
• Linea ke nel Xi – Equa ion (4)
𝐾𝐾(𝑥𝑥𝑖𝑖,𝑥𝑥)=𝑥𝑥𝑇𝑇𝑥𝑥𝑖𝑖 (𝟓𝟓)
• Polynomial ke nel – Equa ion (5)
𝐾𝐾(𝑥𝑥𝑖𝑖,𝑥𝑥)= (𝑥𝑥𝑇𝑇𝑥𝑥𝑖𝑖+ 1)𝑑𝑑 (𝟔𝟔)
8
• Radial Basis Func ion ke nel – Equa ion (6)
𝐾𝐾(𝑥𝑥𝑖𝑖,𝑥𝑥)=exp �−𝛾𝛾�|𝑥𝑥𝑖𝑖−𝑥𝑥|�2�,𝑤𝑤ℎ𝑒𝑒𝑒𝑒𝑒𝑒 𝛾𝛾=1
2𝜎𝜎2 (𝟕𝟕)
Only he suppo ec o s a e impo an o he model – he dele ion o all o he ows
bea s a coe icien 0 and hence, does no change he ou come o he p edic i e model.
Ano he a iable whose alue plays impo ance in building he model is he
egula iza ion pa ame e C ha deno es he limi o he absolu e alue o he coe icien s
αi, which gi es he limi o how much i is needed o a oid misclassi ying a aining
example. The limi cases hen become he ollowing:
I. he la ge C becomes, he close he unc ion can i he da a in he hype plane
and smalle -ma gin hype plane will be needed
II. an ε o 0 c ea es a leas -absolu e-e o eg ession wi h cons ain C
III. A la ge alue o ε jus ou pu s he la es ube ha encloses all da a
Gamma (𝛾𝛾) is a speci ic pa ame e o adial basis ke nel ha de ines how a an
in luence o a single aining example eaches. A la ge gamma would co espond o
mo e suppo ec o s ha ing in luence on he hype plane.
2.1.4 Sel -O ganizing Map and GeoSOM
The Sel -O ganizing Maps a e a clus e ing algo i hm in oduced in he 1980s (Kohonen,
1982). The main idea is o map high-dimensional da a, in o dimensions om which he
human eye can unde s and and ex ac pa e ns. The uni s by hemsel es can be loosely
o closely connec ed o each o he , and he deciding clus e s a e decided by he use .
SOM algo i hm can ha e di e en dis ance me ics, bu mos o en makes use o he
Euclidean dis ance as a measu e o closeness be ween wo a bi a y uni s. Once a SOM
has been ained wi h a gi en pa e n, all he uni s mo e owa ds he bes ma ching uni
(BMU). A he end, pa e ns ha a e simila in he inpu space should also ca y his
beha iou in he ou pu space.
GeoSOM is an adap a ion o SOM o accoun o he speci ici y o spa ial da a. The
sea ch o he BMU akes pa in wo cycles. Ins ead o sea ching o a BMU h oughou
he whole da ase , i ies o ind a neighbou ing one ha is limi ed in he sea ch adius
wi h he pa ame e k. The ou pu space in SOM usually akes he o m o a 2-
dimensional space (Kohonen, 2001) as he easies one o isualize.
9
Figu e 2.3 - Sel -O ganizing Map - I/O space (Hen iques, Bação, & Lobo, 2009)
The e a e a numbe o impo an pa ame e s ha need o be uned when aining a
da ase wi h SOM. These include lea ning a e, adius, a numbe o i e a ions, and hey
need o be chosen acco dingly o minimize a model’s opological e o .
2.2 Rela ed Wo k
Th oughou his subsec ion, e e ences and links o s udies ega ding p ice p edic ion
a e men ioned o gain a gene al o e iew o wha esea ch has been applied and he
co esponden indings ollow. Fo a clea e s uc u e, he ela ed wo ks a e di ided in o
h ee sec ions, one deno ing Spa ial Econome ics’ ela ed s udies, a second o SVR
ela ed s udies, and a hi d o gi ing b ie ing o o he algo i hms’ usage.
2.2.1 Resea ch h ough Applica ion o Spa ial Econome ics
Resea ch shows indispu able e idence o he impac o he cul u al he i age on he eal
es a e p ices in he ci y. A case s udy o he ci y o Lisbon inds p oo ha a p o ec ed
zone can p oduce a nega i e impac on he house alue bu i also p oduced indings
ha when accoun ing o he he e ogenei y o he a eas unde ques ion, he e ec s had
seemingly disappea ed (F anco, Macdonald, F anco, & Macdonald, 2016).
Ano he s udy has e ealed ha o e all, his o ic ameni ies gene ally gi e ise o
dwelling p ices in o de o 4.2%, bu once he adius b oadened, a high concen a ion
o hese his o ical ameni ies s a ed o yield a sligh ly nega i e e ec o 0.1% (F anco
& Macdonald, 2016).
When analyzed, hese pape s b ing abou conclusions ha li ing in a p oximi y o an
ameni y wi h some deg ee o cul u al impo ance gi es exposu e and iden i y o an
apa men being in he cen e o a local clus e (a desi able esiden ial a ea). Li ing in
close p oximi y o many o hese objec s on he o he hand had shown a co ela ion o
10
i being a highly ou is ic a ea, which in u n is alued as a e y desi able dwelling o
one’s day- o-day li e.
O he esea ch has also ound ax exemp ions o d i e posi i e spillo e s o nea by
p ope ies ( an Duijn and Rouwendal 2012, Ahl eld , Holman, and Wendland 2012,
Coulson and Lah 2005).
Wha ’s mo e, using a HPM on eal es a e p ices ha e ound Lisbon’s ‘Me o S a ion
P oxy’ a iable coe icien o gi e be ween 3.49% and 5.18% ise in p ices wi h a ed
posi i e accessibili y o a single me o line, o be ween 4.62% and 6.17% o
accessibili y o wo me o lines and e en la ge o he ail accessibili y (Ma ínez &
Viegas, 2009).
Addi ionally esea ch ha e also used addi i e hedonic eg ession models o accoun o
he he e ogenei y in he en ma ke (B unaue , Lang, Wechselbe ge , & Biene , 2010)
o eal es a e p ices in Vienna. This had led o conclusions abou disco e ies o
subma ke s which ha e as a signi ican ac o a ‘Belonging’ o lis ings in he speci ic
ci y’s dis ic s.
O he sou ces (Mo o, Mayo , & Lyons, 2013) include a s udy o e he ci y o Dublin
as a case s udy whe e he i age si es wi hin he ci y we e ca ego ized as ac o s using a
limi ed numbe o ca ego ies. This s udy had gi en a small bu signi ican posi i e sco e
o ce ain ypes o ca ego ies (his o ic buildings and memo ials) as a spillo e e ec o
apa men s in he icini y, al hough o he ca ego ies (a cheological si es) had p oduced
a nega i e one. The s udy had been made h ough p oducing dummy a iables as lags
o whe he a ce ain ype o building was in a adius o a chosen ameni y. Va iables
pe aining o me os and bus s op as means o public anspo had also been used and
i had been no ed ha li ing wi hin 100 me e s o a me o s a ion esul ed in a posi i e
co ela ion wi h a p ice inc ease o an apa men (F anco e al., 2016).
(Poo , 2015) ha e mo eo e a gued ha he p esence o cul u al he i age a ac s highly
educa ed households using hei case s udy o he Ne he lands. This ac o had also
caused a spillo e o indus ies in he coun y in es ing and p e e ing o eside whe e
he eli e households o he coun y eside (Ma le and Woe kens, 2005).
Las , bu no leas , o he s udy indings include cul u al he i age si es i sel ca ying a
mul iplie e ec o o he ameni ies including es au an s and shops ( an Duijn, 2013).
This gi es an idea o linea dependencies be ween exogenous a iables could clu e o
bias an analysis and make o an o e ly complex model.
These s udies, o he mos pa , decons uc many spa ial a iables seg ega ed and
analyzed sepa a ely (as objec i e o unde s anding single a iable’s impo ance) a he
han i ing one model using many. The majo i y o hem use Euclidean dis ance o
calcula e p oximi y, whe ein a PoI can eside anywhe e wi hin a p oposed adius
h eshold o an ameni y in ques ion.

11
2.2.2 Resea ch h ough Applica ions o SVR
Finding using he SVR include (Chen, 2010) ha discuss ha s a i ica ion o
Shanghai’s ma ke in o mo e homogeneous subse s p o ides conside able bene i s as
con as o aking he agg ega e o he whole Shanghai’s en ma ke . The app oach
used a hedonic app aisal model as a p ep ocessing s ep o choosing he igh a iables
be o e building he SVR. Ano he pape has buil a eal es a e o ecas ing model based
on pa icle swa m op imiza ion (PSO) be o e applying he SVR (Wang, Wen, Zhang,
& Wang, 2014). Op imiza ion h ough HPM seems o be a ecu ing heme when using
SVR and SVM. In his way, he signi ican a iables need o be de ec ed i s be o e
building he model so he ime complexi y also lowe s subs an ially, whils also making
su e he model’s accu acy does no su e om he mul icolinea i y p oblem.
2.2.3 Resea ch h ough Usage o O he Algo i hms
A ecen s udy o eal-es a e p ice e alua ion was done h ough he usage o bo h NN
and HPM as a way o compa ing he wo algo i hms. Ne e heless, his esea ch was
concen a ed on p oducing minimum MAE accu acy model o p edic ions and does
no gi e u he in o abou he singula in luence o unique a iables in building he
model (Sa ono , 2017). This is because o he “black box” na u e o ANN in which
mul iple neu ons a e exchanging in o ma ion and a e upda ing hei weigh s, bu o he
ou e wo ld no a lo o subs an ial in o ma ion is gi en h ough which he obse e can
ge many conclusions. Ano he s udy was done using ANN on he island o Cabo Ve de
using 1092 da a poin s wi h a yea ly empo al scale be ween 2009 and 2014. This s udy
also came abou as a compa ison be ween using he algo i hm o ANN and Random
Fo es and d aws conclusions o MLP p oducing highe e o s han he la e algo i hm
(Es e & Ma ins, 2016). The mos impo an a iables he e we e ound wi h he model
i , and among hem, he loca ion o he apa men and he squa e me e o he apa men
seemed o cause bo h highes signi ican and posi i e in luence, wi h he closeness o
public ins i u ions and an exis ing balcony in he apa men compu ed as ones wi h leas
impo ance. ML s udies o en ha e a lo o esea ch concen a ed solely on inding he
bes model wi h he leas e o (C is ina and Teixei a 2009, McCluskey e al. 2013,
Limsombunchai e al. 2004) wi h using NN and MLP’s as ML s anda ds ha manage
o build models ha cap u e he da a’s beha io he bes , p oducing he smalles e o .
One can also no e usage and compa ison o Random Fo es in eal es a e p edic ions in
Ljubljana (Kiliba da, 2018), whe e indings sugges in he s udy a ea unde e alua ion,
i pe o med signi ican ly be e han o he ML algo i hms. These a e jus a ew small
e e ences o he as usage o many di e en algo i hms wi hin he ML ield ha could
be used o assess his ype o da a, al hough i mus be no ed ha e e y algo i hm a ies
in he ou pu in o ma ion and hence, he ype o analysis i allows o .
12
3. DATA AND METHODOLOGY
This chap e explains in de ail he unde aken s eps o p oduce his disse a ion. Fi s , a
geog aphical con ex is gi en in Sec ion 3.1 whe ein gene al in o ma ion abou he
encompassing s udy a ea is in oduced. In Sec ion 3.2 a b ie o e iew o he p oposed
a chi ec u e. The chap e con inues wi h Sec ion 3.3 he eupon gi ing a synopsis o he
so wa e and algo i hms ha a e impe a i e o p oducing new spa ially consis en da a.
Sec ion 3.4 hen desc ibes he da a sou ces and a iables choices. The inal, Sec ion 3.5
po ays how he di e en unde aken pa s o he me hodology come oge he o
encapsula e he analysis.
3.1 Geog aphical Con ex
Lisbon is he capi al and la ges ci y o Po ugal, wi h a bounding box desc ibing a
la i ude ange be ween -9.2379 and -9.0863 and a longi ude ange o 38.6800 and
38.7967. The adminis a i e a ea o he ci y co e s app oxima ely 100.05km2, and he
ci y’s las census had eco ded 505,526 ci izens. A clea dis inc ion is made be ween
Lisbon’s adminis a i e a ea and he u ban a ea ha ex ends a beyond hese limi s.
The posi i e ele a ion goes om 0 a i s minimum up o 227 me e s o al i ude.
The ci y has a ci cula physical con igu a ion co e ing app oxima ely 12 kilome e s
bo h eas o wes and no h o sou h. O icial di isions di ide i in o 24 pa ishes, o in
3623 subsec ions a he lowes mapping le el. Maps o he pa ishes’ polygons and he
le el 4 subsec ions a e shown in Figu e 3.1 below.
Figu e 3.1 - S udy A ea Di isions: Pa ishes e sus Le el 4 Census Blocks
13
3.2 P oposed A chi ec u e
The pu pose o he me hodology is o y o inco po a e many di e en sou ces,
acco ding o he a ailabili y o he da a and he ca ego ies o da a poin s hey o e . The
policy o anspa ency o he EU gi es space o use open da a eely which helps in he
p ocess o ga he ing an abundance o da a ha could be in e -checked o eliabili y
be ween he di e en sou ces and gi es he choice be ween using sou ces wi h he
bigges quan i y, bigges quali y o in eg a ions. Ne e heless, his also b ings abou
weigh and complexi y o compu a ions needed o eplica ing he analysis. Fo easie
p ocesses, all da a is impo ed o a Pos g eSQL da abase, om whe e i could be easily
expo ed and ead in bo h QGIS and RS udio. This in eg a ion o en i onmen s p o ides
a obus b idge o easie da a con e sion, manipula ion and addi ional pa alleliza ion
o compu a ional p ocesses. The p oposed a chi ec u e is ound in Figu e 3.2.
The Figu e shows li le de ail o he R analysis, al hough he packages used a e o be
men ioned in he ollowing sec ion.
Figu e 3.2 – P oposed A chi ec u e
3.3 Ha dwa e and So wa e
Conce ning ha dwa e, he esea ch used an In el Co e i7-7700HQ wi h 16 GB o RAM
capaci y, unde a Windows 10 ope a ing sys em.
Rega ding so wa e, he majo i y o he unde aken analysis was done wi hin
Pos g eSQL and R. Addi ionally ancilla y so wa e was used o isualiza ion including
QGIS, A cGIS P o, as well as mino checks and la iles o he da ase o e sioning
his o y as excel iles.
3.3.1 Pos g eSQL
This is a ee and open-sou ce ela ional da abase managemen sys em ha ex ends he
SQL language whe e use s can sa ely s o e and scale complica ed da a wo kloads (The
Pos g eSQL Global De elopmen G oup., 2014). When dealing wi h spa ial-da a, he
14
da abase capabili ies can be accompanied by addi ional add-ons. pgAdmin4 e sion is
used, alongside he ollowing ex ensions:
• Pos GIS: Allows o geog aphic loca ion que ies o be un in SQL (Pos g eSQL
Global De elopmen G oup, 2020).
• pgRou ing: Enables a oad ne wo k analysis o be pe o med; I is used o
calcula e he sho es pa h o places h ough pg _d i ingDis ance unc ion
(OSGeo Founda ion, 2018)
3.3.2 R
R is a sys em o s a is ical compu a ion and g aphics. I p o ides, among o he hings,
a p og amming language, high-le el g aphics, in e aces o o he languages and
debugging acili ies (R Co e Team, 2019). The language is able o be execu ed on many
ope a ing sys ems ee o cha ge. The RS udio, subsequen ly, ep esen s an in e ace o
he execu ion o R code. Packages ha make he analysis possible a e lis ed as ollows:
• s : Package ha allows o eading, w i ing, manipula ing and isualizing
spa ial da a and geome ies, o and om R (Edze Pebesma, 2019)
• RPos g eSQL: Da abase in e ace and ‘Pos g eSQL’ d i e o R (Conway e
al., 2019). Wi h his package di ec connec ion o he Pos g eSQL da abase is
made, and ables and que ies o he da abase can be di ec ly impo ed and sa ed
as a iables wi hin he RS udio en i onmen
• dply : A ool used o easie manipula ion o da a ames (Wickham, 2019)
• mag i : Ope a o ha o wa ds a esul s o an exp ession o he nex unc ion
ha comes a e (S e an Mil on Bache and Hadley Wickham, 2014)
• ca e : A comp ehensi e amewo k o machine lea ning models (Kuhn, 2020)
• e1071: A package used o in ui i e applica ion and analysis o he SVR
algo i hm, one among many p obabilis ic clus e ing and eg ession models
(Meye , 2019)
• spa ial eg: A collec ion o es ima ion unc ions o spa ial c oss-sec ional
models (Roge Bi and 2019). Lib a y used o applying HPM on he da a
• spdep: A collec ion o unc ions o c ea e spa ial weigh s ma ix objec s om
polygon ‘con igui ies’ by dis ance and essella ions (Roge Bi and 2019). This
package is used o applying unc ions in he HPM spa ial eg ession analysis.
3.3.3 Py hon
Py hon is an in e p e ed, high-le el, gene al-pu pose p og amming language. The
language can be used o any compu a ional pu poses, al hough o he means o he
disse a ion, i was used o inding speci ic-pu pose API W appe s. The API W appe
hen p o ides an in e ace be ween an API and a p og amming language, so all p o ided
API me hods o da a collec ion can be made h ough he desi ed p og amming
language’s unc ion calls and w appings.
21
dummy a iables we e, in u n, no included. The mos equen wo ds a e shown in
Figu e 3.5 abo e.
3.5.2 Spa ial Da a En ichmen
Da a om many o he sou ces o e laps. The e o e, decisions o uni ica ion o hese
da a in o one sou ce we e unde aken.
The da a om Zoma o holds he la ges quan i y o es au an s, and con ains
in o ma ion pe aining o e iews and he cos liness o he es au an s. Hence, i was
decided o use i as he only sou ce o ood- ela ed PoIs. O he PoI ca ego ies om a
single sou ce (CML) we e me ged depending on whe he he di e en ia ion be ween
hem ailed o appea as a esponse in simila s udies o his ype (e.g. p i a e and public
schools o i s , second deg ee and p e-cycle a e me ged). Medical p ac i ione s we e
subsequen ly me ged wi h independen doc o s. Ne e heless, hospi als and medical
cen es a e le as a sepa a e ca ego y.
I wo used sou ces a e ound o hold a subs an ial amoun o di e en unique da a,
A cGIS P o Analysis – “Gene a e Nea Fea u es” unc ion is used o de ec ion o all
ea u es be ween he wo laye s ha a e wi hin 2 me e s one o he . These we e
p esumed o ep esen he same da a poin and we e dele ed om he second sou ce
be o e me ging.
The ca ego ies o ameni ies chosen a e lis ed in Figu e 3.6 and Figu e 3.7 below.
Figu e 3.6 - Coun s o Fea u es pe Ca ego y A
Figu e 3.7 - Coun s o Fea u es pe Ca ego y B

22
Fi y- i e new a iables hus exis ; 54 o hem a e shown in he Figu es abo e, and he
las one deno es ele a ion ex ac ed om a DTM (Lisboa - Digi al Te ain Model,
2010).
The spa ial da a en ichmen hen makes use o he unc ion pg _d i ingDis ance() o
de ec ion o eachable ameni ies om a eco d’s lis ing.
pg _d i ingDis ance('SELECT id, sou ce_osm, a ge _osm, cos _s, FROM ways',
s a ing_ e ices, dis ance, di ec ed := alse)
To be able o execu e his command, a lis ing needs i s node o igin poin , and a
des ina ion poin . In he case o he ooms, compu a ions gene a e he closes node o
each oom as well as he closes nodes o each o he des ina ions. Figu e 3.8 below
shows sample ables wi h gene a ed closes nodes’ ield using he ways able.
Figu e 3.8 - O igin (Room), Des ina ion (PoI) and Edge Ne wo k DB Schema
Fo less cos ly compu a ion, an addi ional in e media y able ca chmen _a ea is
c ea ed, which desc ibes all po en ially eachable nodes om all o igin nodes o he
ooms (gi en a desi ed me e s uppe h eshold).
WITH
nodes AS (
SELECT a ay_agg(node_id) AS nodes om ap _ oom.node_id)
SELECT om_ as s a _node, node as end_node, agg_cos as cos om nodes,
pg _d i ingdis ance(
'SELECT gid as id, sou ce as sou ce, a ge as a ge , cos _s as
cos , di ec ed FROM public.ways':: ex , nodes, 2000, alse)
This code, in u n, p oduced a able ha holds all he eachable nodes wi hin 2km. A
isualiza ion o he eachable a ea om a sample poin is shown in Figu e 3.9 below.
23
Figu e 3.9 - Road Ne wo k Accessibili y om a Sample Room
The p oxies o each ca ego y o PoI hen cons i u e o :
• Minimum dis ance (in me e s) om he oom o he closes ameni y o a
p ede ined ype
• Coun pe ca ego y o PoIs wi hin a dis ance h eshold. Fi e dis inc h esholds
a e p oposed as ollows: 2km, 1.5km, 1km, 0.5km, and 0.25km.
Two a iables ha we e speci ic in hei seg ega ion p ocess we e he ‘Pa ks’, many o
which ha e mul iple closes nodes as being 0 me e s away om i s en ance, and he
‘Res au an s’ ha ha e a e y high quan i y and densi y h oughou he whole s udy
a ea. The e o e, a sepa a e able o pa ks was cons uc ed ha holds all nodes less han
10 me e s away om he pa k o in e es , as possible des ina ion nodes om an o igin
poin . Unique p oxies a e cons uc ed o he es au an s acco ding o hei
expensi eness (1-4 s a s), and addi ionally, acco ding o hei a e age e iew a e
(s a ing om 0 as no- e iews, 1 as bad, up o 5 as excellen a e).
Figu e 3.10 below, p esen s an example wi h an in e pola ed su ace o he minimum
dis ance needed o walk om a sample oom o he i s egis e ed s a ue in he da ase .
24
Figu e 3.10 - Example: Minimum Dis ance o a Mall (in me e s)
3.5.3 Da a P ep ocessing
The goal o his sec ion is o p o ide an o e iew o he desc ip i e da a s a is ics, as
well as some p ep ocessing decisions o he in eg a ed da ase and he easoning behind
hem.
Ou lie de ec ion – A ecommended app oach when dealing wi h da ase s is o emo e
ou lie s. These can ep esen e oneous da a and in his speci ic case, alse en ies can
be pa o o e p iced o e s i a oom’s acancy is a in he u u e, o addi ional
nego ia ions migh be p esen be ween cus ome -clien , which canno be accoun ed o .
The s a is ics o he endogenous a iable a e shown in Table 3.4 below:
S a is ic
Value
Mean
473
S anda d E o
3.72
Median
420
Mode
400
S anda d De ia ion
173
Range
900
Minimum
100
Maximum
1000
Coun
2149
Table 3.4 – Desc ip i e S a is ics o Dependen Va iable
The in e qua ile ange inds all alues om 870 abo e as ou lie s and no ou lie s below
he minimum alue, which amoun s o a o al o 105 ou lie s ou side o he 1.5
in e qua ile anges below he i s and abo e he hi d qua ile. Ne e heless, his da a
25
and scena io depic s a eal-li e Lisbon en ma ke a a speci ic empo al ins ance in
Sep embe 2019. Hence, i is decided agains he emo al o hese ou lie s.
The Clus e and Ou lie Analysis was un in o de o iden i y clus e s and spa ial
ou lie s. False Disco e y Ra e Co ec ion (FDR) pa ame e is checked, o educe he
c i ical p- alues o accoun o mul iple es ing and spa ial dependence. I is no iceable
ha he e a e many clus e s adjoining high-p iced ooms wi h low-p iced ooms nea by,
as well as low-p iced ooms wi h many high-p iced ooms nea by. The high-low p ice
clus e s end o exhibi a pa e n om cen al o he no he n bo de s o he s udy a ea,
whe eas he low p icings wi h many high-p iced ooms end o ha e a gene alized
localiza ion nea he coas .
Clus e Type
Coun
High-High
185
High-Low
64
Low-High
86
Low-Low
232
Non-Signi ican
542
To al
1109
Table 3.5 – Coun s o Signi ican Ho -Spo Analysis Clus e s
Figu e 3.11 - Anselin Local Mo an’ I
When i comes o numbe s o ooms a ailable pe Pa ish, mos o he a ailable ooms
a e ound in pa ishes whe e he signi ican clus e s domina e. As such, A enidas No as,
A eei o, Al alade, San a Ma ia Maio and Mise icó dia hold he la ges quan i ies.
Addi ionally, a high coun o bed ooms p edomina es he cen al pa ishes o Lisbon, as
26
he bed ooms end o dec ease apidly owa ds he no h bo de s and dec ease
ma ginally owa ds he sou hwes and sou heas bo de s.
Following his app oach, a ew bina y a iables we e pu unde inqui y and esul s o a
Ho Spo Analysis a e subsequen ly shown in Table 3.6.
Va iable
Pa ish – Ho spo
Pa ish – Coldspo
Male Landlo d
Lumia , San a Cla a, Ben ica,
Al alade, Bea o, Penha de F ança,
Belém
San o An ónio,
A oios
Female Landlo d
Ca nide, Al alade, Lumia , San a
Cla a, Pa que das Nações
São Domingos de
Ben ica, A eei o
No Landlo d
A enidas No as, Campolide,
A oios, San o An ónio
Al alade, Lumia ,
Ben ica
Has Ou doo A ea
Ajuda, Ma ila, Oli ais, Es ela,
Campo de Ou ique
A enidas No as
No Ou doo A ea
Alcán a a, San a Ma ia Maio ,
Ma ila, Mise icó dia, São Vicen e,
A eei o
Lumia , Al alade,
Oli ais, São
Domingos de Ben ica
Ou doo A ea N/A
Campolide, Ca nide, Lumia ,
Al alade, A enidas No as
Oli ais, Ma ila,
Belém
Table 3.6 – Ho -Spo s and Cold-Spo s o Pa ishes
These ho spo s can indi ec ly in oke knowledge abou he pa e n – and i can be seen
ha mos ooms ha a e en ed in an apa men wi h a hos appea in all pa ishes bu
he cen al ones. This signi ies he a ea in ques ion is likely o be comp ised o s uden -
domina ed apa men s and i includes he pa ishes o Al alade, A eei o, A enidas
No as and San o An ônio. ‘Male Landlo d’ ho -spo s a e ound bo h in he Pa ishes in
he no h and on he i e coas , bu ‘Female Landlo ds’ do no appea as ho spo s on
any coas al pa ish. An exis ing ‘Ou doo A ea’ ho -spo s appea equen ly in Ajuda
and Alcán a a, bu ho -spo s o apa men wi hou i appea equen ly in all pa ishes
alongside he coas (in which Ajuda and Alcán a a also belong). Rooms belonging in
apa men wi h ‘Non-speci ied Ou doo A ea’ also show a ew ho spo s in he mos
cen al pa ishes.
Iden i ica ion o Co ela ed P edic o s and Linea Dependencies - The da ase
con ains many a iables, and a pai wise check is done o emo e a iables ha migh
bias he eg ession. All p edic o s ha a e co ela ed wi h abo e 85% a e hus emo ed.
Da a No maliza ion – The da ase la gely con ains da a in e y di e en me ics
(me e s, quan i ies and bina y a iables) – he min-max ange is applied so he e exis s
a common scale ha would no lead o any dis o ing di e ences in he ange o alues.
Da a Agg ega ion – As he spa ial applica ion o hedonic p ice modelling h ough he
‘spa ial eg’ R package equi ed polygons neighbou s o calcula ing he lagged
in luence o he exogenous a iables, di e en da ase is c ea ed o applying HPM. A
spa ial join is pe o med using he laye o he Census Blocks a he 4 h le el wi h he
laye o ooms and i s ea u es. Fo each polygon, he a e ages s a is ics o he lis ings
belonging in he mapping uni is calcula ed. The esul ing shape ile hen con ains 987
polygons ha deno e only census blocks o which he e a e lis ings p esen .

27
Coun o Dwellings pe Block
# o Census Blocks
To al Rooms
1
503
503
2
218
436
3
113
339
4
63
252
5
36
180
6
15
90
7
12
84
8
11
88
9
7
63
11
3
33
14
2
28
10
1
10
12
1
12
13
1
13
18
1
18
987
2149
Table 3.7 – Coun o Agg ega ed Dwellings in Blocks
T aining and Tes ing Spli – The HPM is o be buil o measu e how well he da a can
i accoun ing o he e o o he esiduals h ough he spa ial-speci ic chosen models,
and he whole da ase is used, as he knowledge o he neighbou polygons a iables is
needed.
The SVR model, on he o he hand, needs o be i s ained o be able o es he
pe o mance on a di e en unseen se . In o de o a oid he o e i ing p oblem (Zang,
Be a di, & Rei e mano á, 2010), a sepa a e es ing da ase is alloca ed o check whe he
he algo i hm pe o ms well only because i has memo ized he pa e ns on he aining
se oo well.
The p ocess o o e i ing can also be waned by K-Fold c oss- alida ion, which can be
in oked as a con ibu ing pa ame e wi hin he ‘e1071’ lib a y wi h which he SVR
model is buil . The choice o k is usually 5 o 10, al hough a o mal bes - alue o K
does no exis . In gene al, as k ge s la ge , he di e ence in size be ween he aining
se and he esampling subse s ge s smalle . As his di e ence dec eases, he bias o he
echnique becomes smalle (Kuhn & Johnson, 2013).
3.5.4 Spa ial Econome ics and Machine Lea ning
As a las s ep in he me hodology, di e en subse s o he clean da ase a e o be es ed
o compa isons o a iables’ in luence, and he assessmen will be done wi h he
da ase s con aining:
• All a iables (applied o HPM and SVR)
• Only non-spa ial a iables (applied o HPM, SVR, SOM)
• Only spa ial a iables (applied o HPM and SVR)
• All signi ican a iables om HPM (applied o SVR and SOM)
28
4. RESULTS AND DISCUSSION
Fo compa ison o he di e en eg ession models applied, Adjus ed R2 and Akaike
In o ma ion C i e ion (AICs) me ics a e used.
4.1 Spa ial Dependence
OLS was pe o med on he da ase , and on ha model, he Global Mo an I s a is ic was
compu ed o e he esiduals. The null hypo hesis assumes no spa ial dependence, and
he esul s show he ollowing:
• S anda d de ia ion = 1.751
• p- alue = 0.03583
Al e na i e hypo hesis: GREATER
Sample Es ima es:
Obse ed Mo an Index
Expec ed Index
Va iance
-0.0624
-0.1157
0.0008
Table 4.1 - Local Mo an’s I
The p- alue below 0.01 p esen s a s a is ically signi ican ejec ion o he ini ial null
hypo hesis. The da a unde analysis leads o spa ial esiduals, and spa ial models a e
deemed app op ia e in o de o make he e ec s o spa ially dependen e o s disappea .
The LaG ange Mul iplie es diagnos ics addi ionally gi es di ec ions o whe he he e
is a signi ican imp o emen o he model i i spa ial models we e applied. Resul s can
be seen in Table 4.2 below:
Model
p- alue
SAR
0.9457
SDEM
0.0042
SARMA
0.0166
Table 4.2 – Models Imp o emen
The p- alue sugges s ei he Spa ial Du bin E o Model (SDEM) o Spa ial
Au o eg essi e Mo ing A e age (SARMA) can b ing a signi ican imp o emen ,
al hough SAR is deemed un i (insigni ican p- alue). Acco ding o Luc Anselin, he
bes app oach is o con inue wi h he one wi h he lowes p- alue, which in his case
was he SDEM (Anselin, Be a, Flo ax, & Yoon, 1996). Ano he model in oduced is
SLX, in o de o be able o compa e models ei he wi h local (SLX) o global (SDEM)
spillo e s in he dis u bances.
29
4.2 Spa ial Lagged X
When applying an HPM eg ession, he con ibu ing in luence o indi idual a iables
is compu ed. The numbe o s a s assigned nex o a a iable indica es i s signi icance.
Coun s o how many o each signi icance ype is ound o he model ha includes all
a ailable a iables is shown in Table 4.3.
Code
P obabili y (> |)
Coun o Va iables
***
>0.001
3
**
>0.01
6
*
>0.05
38
X
>0.1
41
Table 4.3 – Signi icance Pe cen age
The esul s ha e a esidual s anda d e o o 114.7€ which amoun s o a MAPE o 15%
on p edic ing a oom p ice wi hin census blocks ha has agg ega e desc ip i e s a is ics
o maximum p ice o 950€ and a minimum o 200€ among hem.
The a iables ha a e deno ed as impo an excluding he X in e cep amoun o 97
a iables. The mos signi ican di ec e ec ones (abo e 95% signi icance) a e p esen
in Table 4.4 below:
Va iable
Es ima e
SE
- alue
P (>| |)
Sig.
C. o Bed ooms
-199.908
54.823
-3.646
0.00001
***
Male Landlo d
-135.44
20.312
-6.668
0.00031
***
MD o Club
20.344
7.799
-2.609
0.00958
**
MD o Viewpoin
89.951
33.764
2.664
0.00817
**
MD o Thea e
51.507
18.864
2.73
0.00672
**
Lag Cen alHea _F
-271.162
97.123
-2792
0.00560
**
Lag C. o Bed ooms
-126.196
46.517
-2.713
0.00708
**
Lag Ele a o N/A
350.502
159.086
2.203
0.00283
**
Table 4.4 – Signi ican Va iables – Model using All Va iables
F om hese a iables, i is in e ed ha he exis ence o a male landlo d li ing inside
he apa men makes he p ice o he oom be a subjec o nega i e in luence. This may
esul om he ac ha he ooms a e mos ly used be ween 6 mon hs and 1 yea , and
as such a e pe ec o ei he s uden s o people ha ha e ecen ly s a ed wo king. In
his sense, o eigne s ha wan o expe ience Lisbon may end o ei he g a i a e
owa ds an apa men ha he e is no landlo d li ing in, o expe ience mo e eedom
du ing hei s ay – o hey s a is ically alue a emale pe son li ing o aking ca e o an
apa men much mo e han a male pe son. S ill, on a e age ooms wi hou landlo ds a e
p iced highe . This signi icance can be ein o ced by s udies ha con i m ha young
adul hood has become a dis inc new li e phase (Be houd, Ge shuny, & B i ish
Household Panel Su ey., 2000). Du ing his phase, people alue and choose o li e in
pee -sha ed households and quasi-communes wi h people likely o be hei own age,
a he han li ing wi h a landlo d (Hea h, 2004).
30
Howe e , simila s udies ha measu e a hos ’s gende signi icance in he p ices o
sho - e m s ays applied on an Ai bnb da ase co obo a e no impac on i , as opposed
o de ec ion o signi ican acial disc imina ion (Kaka , Voelz, & Wu, 2017).
The numbe o bed ooms a iable impac in decline o p ice comes also as gene ally,
apa men s end o ha e one ki chen o use and hus, he spaces would be subjec o
sha ed use among people ha a e no necessa ily amilia wi h each o he .
F om he spa ial ones, he mos signi ican ones deno e he minimum dis ance om a
iewpoin as a posi i e coe icien . This means apa men s close o a iewpoin would
be lowe -p iced. This migh be a esul as hese iewpoin s a e o en o -g id o whe e
mos daily ac i i ies happen, and li ing close o a iewpoin can be o en connec ed o
an upwa d slope when looking a he oom as a des ina ion. Coun s o uni e si ies
wi hin 500 me e s and ams wi hin 1km also all yield posi i ely in luencing a iables
(T oy & G o e, 2008). The anspo a ions’ e ec s a e no unique o mid- e m en als
as i has been p o en hey a e ound in esiden ial p ope y alue indices wi hin he
same s udy a ea, and u he hey can be used o o ecas possible upwa d p ice changes
i an a ea is a subjec o new anspo a ion in es men s (Ma ínez and Viegas, 2009).
A lag o he dummy a iable signi ying he exis ence o an ele a o wi hin he buildings
o he apa men s lis ed wi hin a census a ea close o he one being en ed would seem
a e ched o ha e an in luence in p edic ing p ice – bu i migh be ha he i s ule o
nea by hings being mo e simila han a ones comes o play he e – and buildings wi h
no ele a o s migh beha e in a clus e ed way. Howe e , he in luencing s eep
coe icien o 350 aises a ques ionable impac .
F om he a iables wi h less signi icance, he ollowing coun s a e p esen :
Type o a iable
Signi icance “*”
Signi icance “X”
Spa ial
14
13
Spa ial Lagged
8
19
Census
5
2
Census Lagged
3
3
Non-Spa ial
2
1
Non-Spa ial Lagged
3
2
To al
35
39
Table 4.5 - Coun o Signi ican Va iables in SLX
Adding each o he a iables’ e ec combined wi h i s lagged coun e pa ’s indi ec
e ec yields he o al signi icance o a iable. In his way, a s a is ically insigni ican
a iable wi h i s s a is ically insigni ican lagged coun e pa migh yield a s a is ically
signi ican a iable o he model. The di e ences be ween he wo models include a
diminishing amoun o a iables being signi ican , bu spa ial ones ha appea o be
impo an in he model include a uni e si y wi hin 0.5km, a ou is ou (coun s wi hin
bo h 1km and 0.5km), and a non-spa ial posi i e one deno ing ha he hos has disclosed
whe he he apa men has cen al hea ing. This, howe e , migh lead o di e en
knowledge – apa men s wi h mo e missing da a migh be less desi able o en .
When subse s o he a iables a e applied he esul s di e ma ginally, bu he AIC’s
me ic does no imp o e. I can be no ed ha he a iables signi ican wi h con idence
abo e 95% s ay he same in bo h he non-spa ial model and he spa ially en iched one.
37
o eseeable u u e, and la e e ise he oom o a less expensi e p ice when he e is less
ime o ind a sui able enan o he p oposed loca ion. This is a s a egy ha is
commonly used in sho - e m en als by mo e expe ienced hos s ha manage mo e
lis ings as i helps in maximizing p o i s (Gibbs, Gu en ag, G e zel, Yao, & Mo on,
2018).
These dynamics could be o e come i he e is a mul i- empo al egis e ing on p ice
modi ica ions o e he s udy a ea aken h oughou di e en mon hs wi hin he yea .
Figu e 4.7 - Boxplo s o 4 Highly Signi ican P oxies - SDEM
Addi ionally, boxplo s o signi ican a iables a e shown in Figu e 4.7 abo e. These
boxplo s ep esen agg ega ions o anges o he p ice on he X-axis , and a signi ican
a iable o choice on he Y-axis, wi h he ed line deno ing he 10 h and 90 h pe cen ile
o i s alues wi hin he gi en p ice ange.
F om he boxplo s, i can be no ed ha coun o uni e si ies wi hin 1km ha e a sligh
endency o co ela e o a aise in a lis ing’s p ice. As is, his is a desi able geog aphic
loca ion o a s uden . As la e indings sugges he housing ma ke in Lisbon is
exhibi ing an accommoda ion sho age and housing c isis (Cocola-Gan and Gago,
2018), he hos s migh eel he highe p ice will no impede hem om inding a enan .
This inding ollows and ea i ms s uden i ica ion li e a u e claims ha s ess
p oximi y o campuses as a main d i ing o ce in choice p ocesses o s uden clus e s
wi hin a ci y (Sage, Smi h, & Hubba d, 2012).

38
F om he o he a iables, i is ound ha cheape ooms ha e e y ew, i a all, cul u al
acili ies a ound hem. The minimum dis ance o a hea e also eco ds a sligh
downwa d spi al wi h he lis ings wi h highes p ices esiding closes o his ameni y.
Ano he a iable ha ollows a simila s uc u e is he coun o elec onic s o es wi hin
1km. Ne e heless, wi h 427 elec onics s o es h oughou he s udy a ea, i is likely
ha mos o hem ha e clus e ed ho spo s in cen al pa ishes whe e his in luence s ems
om. The coun o elec onic s o es wi hin 1 km o he lis ing dis ibu ed among he
pa ishes o he s udy a ea a e shown on a cho ople h map in Figu e 4.8 below and ha
assump ion can be empi ically con i med.
Figu e 4.8 – Elec onic S o es pe Pa ish
To con inue he analysis, a da ase con aining only he signi ican a iables as deno ed
by SDEM is ex ac ed and u he applied o c ea ing an SVR model.
4.4 Suppo Vec o Reg ession
Be o e applying SVR, he da ase s a e di ided in o a aining (85%) and es ing da ase
(15%) wi h andomized ows aken as aining samples. The ou pu s models’ e o s
ange be ween 80 and 150, al hough a MAE ha di e s a lo be ween a aining and
es ing se poin s o an o e i ed model, so esul s need o be analysed ca e ully and
pa ame e choices subsequen ly uned.
In Table 4.11, a g id sea ch o he hype pa ame e op imiza ion is shown. The minimal
alue o he accu acy me ic o RMSE as gi en by e1071 inds he model ha bes i s
he da a.
39
Pa ame e
Sea ch G id Values
Coun
C
0.001,0.01,0.1,1,5,10,50,250,500
9
ε
0.001, 0.01, 0.1, 0.3, 0.5, 1.5
6
𝜸𝜸
.0001,.001,.01,.1,1, 2
6
Models Tes ed
324
Table 4.11 – Hype pa ame e Op imiza ion
Va Inclusion
C
ε
𝜸𝜸
SV
SV%
RMSE (€)
RBF All (M1)
10
0.3
0.001
1109
65%
133
RBF Spa ial (M2)
1
0.5
0.01
909
52%
147
RBF Non-Spa (M3)
5
0.3
0.01ssd
1079
62%
130
RBF Sign (M4)
5
0.5
0.1
888
67%
143
Lin All (M5)
0.1
0.5
/
907
52%
155
Lin Spa ial (M6)
0.001
0.5
/
906
52%
157
Lin Non-Spa (M7)
0.1
0.5
/
905
52%
146
Lin Sign (M8)
0.01
0.5
/
888
52%
141
Table 4.12 – Bes Pe o ming Models
F om he p oposed hype pa ame e op imiza ion used on he SVR models (Table 4.11)
and he esul ing RMSE and SVs (Table 4.12), i can be no ed ha all models equi e a
subs an ial amoun o eco ds o be used (upwa d o 50%) o de ine he ma gins o he
hype planes. This is a con ibu ing p oo he e is no much egula i y in he da ase .
Mo eo e , a small sample o 2149 da a poin s o e 312 a iables migh be deemed as
oo li le aining da a o disce n a meaning ul hype plane.
The bes pe o ming model s ill shows he non-spa ial inclusi e as he one ha bes i s
he da a. Ne e heless, he usage o spa ial a iables in addi ion seems o p oduce a
model wi h a compa able beha io in bo h he coun s o suppo ec o s and RMSE.
Mo eo e , he linea ke nel pe o ms compa a i ely wo se o all he da a subse
choices.
In Table 4.13, he MAE me ic o he bes models is shown (see Table 4.12 o model
names). In model M1, i can be no ed ha he ain and es pe o mance di e
subs an ially. Addi ional g id sea ch o e he same da ase wi h a lowe alue o he
pa ame e C (0.1) was done, and a new app op ia e model is calcula ed. In Table 4.13,
his model is deno ed as M9:
Model
MAE T ain (€)
MAE – Tes (€)
R2 – T ain
R2 - Tes
M1
60.78
102.93
0.80
0.44
M2
84.95
114.46
0.66
0.32
M3
83.96
98.96
0.55
0.37
M4
104.78
110.15
0.42
0.26
M9
115.08
114.35
0.33
0.23
Table 4.13 – MAE on Bes RBF Ke nel Models
40
The new aining and es ing e o s in model M9 a e much mo e compa able al hough
hey p oduce a much wo se i . The a iabili y o he da a is no cap u ed well, no in
he aining, no in he es ing se .
Model M3 using only he non-spa ial a iables is s ill he bes p edic o and a di e ence
o 16 eu os be ween aining and es ing da a p o ides p oo ha his model does no
o e i he da a.
The la ges a iables’ con ibu ion owa ds models M3 and M4 a e shown in Table 4.14
below.
M3 Va iable Impo ance
Con ibu ion (%)
Numbe o bed ooms
18.9
Numbe o ba h ooms
17.3
No P e e ed Gende
5.50
Double Bed - Dummy
5.20
No speci ied - Ele a o
4.90
Cen al Hea ing – T ue
4.90
Ou doo A ea – T ue
4.40
Single Bed- Dummy
4.30
Male Landlo d
3.50
Female Landlo d
2.40
Table 4.14 – Va iables Con ibu ions; Model M3 and M4
M4 – Va iable Impo ance
Con ibu ion (%)
Numbe o bed ooms
6.45
Bed oom - Single
3.26
Pa k Coun s – 250m
2.34
Room Squa ed – No Speci ied
2.17
Ele a ion
1.95
MD - Lodging
1.94
Tou is Tou Coun s – 500m
1.91
Viewpoin Coun s – 500m
1.80
MD – Real Es a e
1.79
Supe ma ke – 250m
1.74
Resul s sugges ha Model M3 gi es li le in luence o unique a iables, al hough i
uses a la ge ange o hem. Ne e heless, when he a iables a e sepa a ed in o spa ial,
and non-spa ial, i is compu ed ha he models gains con ibu ions cons i u e only 25%
om he non-spa ial ela ed a iables, and 75% impo ance om a iables ha a e
pe aining o he su oundings o he a ea. As he spa ial a iables a e only supposed o
be e he p edic ion o close he e o gap enough o jus i y hei use, i seems his
model has lea n o sepa a e he da a using hem as he p ima y exogenous a iables.
Ex ac ion o he Lag ange Mul iplie ’s coe icien s o he SV o he models did no
bene i he s udy as coun s o 852 and 601 SV’s coe icien s in model M3 and M4
acco dingly a e gi en he maximum mul iplie ’s alue as he o ce equi ed o en o ce
he cons ain s o he models.
41
Figu e 4.9 – SVM Reg ession – Residual E o s abo e SD Th eshold
F om mapping he high esiduals in Figu e 4.9, he wo models demons a e a di e en
e o beha io , whe ein model M3 in cen al Lisbon and ollow a high esidual pa e n
owa ds he sou h-wes bo de s o he a ea, whe eas M4 seems o ha e i s e o s sp ead
h oughou mos pa ishes. Mo eo e , he model pe o ms qui e well in ce ain
eguesias in he no h-eas e n egions o he case s udy whe ein no da a poin s wi h
la ge e o s a e ound. Howe e , he cen al a eas expe ience e sa ile p ices o which
a ma gin canno be ound – he spa ial a iables a e hus no help ul, bu hey a e
c ea ing close clus e s in he mos cen al pa ishes whe ein la ge coun s o many
ameni ies a e si ua ed. To ha e a be e o e iew o he aul s o each model, he wo
models a e compa ed h ough sca e plo s in Figu e 4.10 and Figu e 4.11 below:
Figu e 4.10 – Model Compa ison in Residual E o
42
Figu e 4.11 – Compa ison o E o s h ough B eaks’ Agg ega ions
Model M3 has a mo e s abilized e o beha io h oughou he anges when compa ed
o M4, acco ding o he b akes gi en. Model M4, on he o he hand, has a e y la ge
absolu e di e ence be ween he p edic ed p ice and he ac ual p ice o he oom o en
in he b eaks deno ing he highe anges. Howe e , bo h models ail o ind easoning
o p ices ising owa ds he uppe ange in gene al (and his is whe e he la ges
esiduals a e p oduced). The maximum absolu e e o s a e egis e ed in model M3 on
he posi i e scale (o e es ima ion o 395€) on an a e age p iced oom (450€), and M1
wi h -426€ unde es ima ion on a high p iced lis ing o 900€.
Figu e 4.12 – Residual E o Compa ison

43
4.5 Na u al Da a Clus e s using SOM and GeoSOM
Bo h SOM and GeoSOM clus e ing o he da ase s we e applied on Models M3 and
M4, o check whe he he s uc u e o he a iables would make simila -p iced ooms
join oge he , o i would jus o ce he o de o na u al geog aphic coo dina es o
o e ule he non-spa ial a iables impac . Di e en sizes o neu ons a e e i ied as
ollows: 10x1, 15x1 and15x10. Manual sepa a ion o clus e s a e elec ed isually using
he GeoSOM so wa e, wi h an icipa ions o clea e sepa a ion o he da a whils also
main aining a small opological e o .
F om he desc ip i e s a is ics o each clus e , unde s andings o a a iable’s agg ega e
beha io is o be ou lined. The SOM algo i hm clus e s o M4 we e chosen wi h a
opological e o a a alue o 0.01 and 14 numbe o clus e s. The clus e s show a e y
spa se beha io h oughou he s udy a ea. E en hough Clus e 3 neu ons on he U-
Ma ix showed a simila i y be ween hese eco ds, e y li le spa ial pa e n was
de ec ed when di iding hem be ween he pa ishes.
Figu e 4.13 – SOM Clus e s – Model M4
When i comes o he p ice a iable, he chosen clus e s’ minimum p ices anged om
a minimum a e age o 398€ in Clus e 2, up o a maximum a e age p ice o 607€ in
Clus e 10. Bo h o hese clus e s a e ound pa ially in he majo i y o he pa ishes.
When GeoSOM is used, mo e na u al o de o geog aphic clus e s appea , as imposed
by he limi ing adius pa ame e . In Figu e 4.14 below, he bina y a iables om he
GeoSOM clus e o M3 a e p esen ed on pa allel coo dina e plo s.
44
Figu e 4.14 – GeoSOM – M3 Bina y Va iables
Va iable/Clus e
1
2
3
4
5
C. o Bed ooms
3.08
4.55
2.78
3.36
3.07
C. o Ba h ooms
1.27
1.58
1.21
1.26
1.29
Minimum S ay
43.01 75.48 21.55 75.45 43.59
A e age P ice
506
454
403
434
420
Minimum P ice
110
200
200
225
100
Maximum P ice
1000
940
890
900
920
Table 4.15 – Clus e S a is ics o Nume ic Va iables in GeoSOM M3
Figu e 4.15 – GeoSOM – Model M3
45
Lis ings in Clus e 3 a e ela i ely indisce nible wi h hose om Clus e 4 in he no h-
eas bounda ies as shown in Figu e 4.15, and hey also blend wi h Clus e 2 in he
cen al a ea. A huge disc epancy can be no ed in he ‘Minimum-S ay’ a iable, wi h a
subs an ially low alue in Clus e 3 deno ing ha he e is no minimum-s ay ule o
hese lis ings. This clus e also desc ibes a high chance o a landlo d li ing in he
icini y o he apa men , co ela es wi h a smalle numbe o bed ooms, and missing
in o ma ion abou he exis ence o cen al hea ing. Al hough geog aphically hese
ooms ie in oge he wi h he wo o he clus e s, his clus e deno es an a ea whe e i is
mo e likely he owne is suble ing a oom o hei own li ing space o inc ease hei
li ing income. Unsu p isingly, he oom p ice he e is he lowes , al hough he la ges
di e ence be ween he minimum a e aged oom p ice Clus e and he maximum is only
103 eu os (12% o he p ice ange). Clus e 2 desc ibes ooms which a e no gende -
assigned, and con ain he la ges numbe o bed ooms wi hin one apa men . Clus e 1
acqui es he coas al pa ishes geog aphically, bu non-spa ial a iable wise he a e age
o he a iables a e mo e compa able o Clus e 2 wi h a sligh ly smalle pe cen age o
lis ings wi hou landlo ds, and a smalle a e age o ooms wi hin he apa men . Clus e
4 seems o ha e a high p obabili y o double beds wi hin he oom, and abou an equal
p obabili y o a landlo d possibly li ing in he lis ing. Clus e 5 is he one wi h he
second-lowes p ice a e age and a la ge lack o in o ma ion abou he ype o bed. This
a iable can p esumably be co ec ed by image p ocessing/analysis o exis ing images
o he oom, o de ec he bed size – an a ea ha could be u he explo ed.
Figu e 4.16 – SOM, Model M4
A he end, a SOM di ision using he SDEM signi ican a iables is made, wi h a choice
o 14 Clus e s o easie compa ison wi h Model M3. Al hough he e is a la ge
pe cen age o some clus e s being cons i u ed o one pa ish i is s ill ha d o pe cei e a
g ea di ision, as many clus e s ha e a he e ogenei y ha sp eads h oughou he whole
s udy a ea (e.g Clus e 2 and Clus e 7). The Mise icó dia a ea is he pu es pa ish
desc ibed by only 2 Clus e s (8 and 12), ollowed by San a Ma ia Maio wi h 3 Clus e s
(8 and 11 wi h mul iple lis ings and 1 lis ing in Clus e 5). Clus e 11 and 12 desc ibe
46
he mos expensi e lis ings om wo coas al pa ishes: San a Ma ia Maio and
Mise icó dia, wi h a e ages o 685€ and 721€ espec i ely. Clus e 8 con ains cheape
lis ings om he wo pa ishes mixed oge he . Clus e 9 desc ibes he cheapes lis ings
(a e aged), comple ely exclusi e o lis ings om he pa ishes o A oios and San o
An onio. These pa ishes by hemsel es, howe e a e no cheap, as lis ings wi h a wide
a ie y o p ices om he wo pa ishes can be ound in 7 and 8 clus e s o he p oposed
14, acco dingly.
4.6 O e iew o Gi en Analysis
The da ase was analyzed using spa ial econome ics, and wo ypes o HPM models
we e es ed (SDEM and SLX) as judged by he mos signi ican p- alue om he
Lag ange Mul iplie es diagnos ics. The spa ial speci ica ions o he models made use
o da a agg ega ion wi h 987 esul ing polygons desc ibing a e ages p ices wi hin he
blocks o he s udy a ea ha con ain a leas one lis ing.
The mos signi ican a iable in bo h SDEM and SLX, and u he in SVR was he coun
o bed ooms – and i bea s a nega i e coe icien . The dummy a iable ‘Male Landlo d’
also le eled a con idence o abo e 99% bea ing a nega i e coe icien . The SDEM
model ound 118 a iables o bea signi icance opposed o 44 om he SLX model.
Va iables deno ing minimum dis ance and coun s o ameni ies up o 250m ha e a high
equency in he mo e signi ican le els. 500m and 1km sha e lowe , bu also simila
equencies be ween one ano he . The impo ance o a iables deno ing a p oximi y o
ameni y abo e 1000m a e a ely a case o in e es , and only appea a signi icance le els
‘X’ and ‘*’.
The da a a iabili y wi hin he bes SDEM model was cap u ed wi h an R Squa ed o
0.92 deno ing 35.7€ e o . This p esen s 8% imp o emen o e he non-spa ially
en iched model. Usage o a high amoun o a iables aised ques ion o model’s da a
o e i ing. Sensi i i y analysis illus a ed as changes wi hin he coe icien s adjoined
o he a iables, hough he a iables signi icance yielded ma ginal di e ences.
Hype pa ame iza ion o he SVR pa ame e s allowed o c ea ion o a model ha i s
he o iginal da a poin s wi h an a e aged pe cen age e o o 11% o he p ice ange.
The spa ially en iched model esul ed in wo se pe o mance wi h a 12% mean absolu e
e o . All SVR models needed mo e han 50% o he da a as SVs, and he maximum
Lag ange Mul iplie coe icien was gi en o he majo i y o he SVs.
The bigges aul s in building his p edic i e model we e he bigge p iced lis ings
which a e ound o be as ly unde es ima ed. The models hus ail o ind a disce nable
pa e n o wha allows a highe p ice ag in a eco d.
A SOM model wi h all he signi ican a iables (as deno ed by SDEM) demons a ed
some geog aphically unde s andable clus e s – as wo ooms belonging o he same
node (o one e y close) would sha e e y simila p oxies s a is ics. Ne e heless, he
bigges denomina o s in he clus e di isions we e s ill he non-spa ial a iables which
we e illus a ed on а co esponding PCP igu e wi h a sho discussion he ea e .
53
P edic ion accu acy in mass app aisal: A compa ison o mode n app oaches. Jou nal o P ope y
Resea ch, 30(4), 239–265. h ps://doi.o g/10.1080/09599916.2013.781204
MESSONIER, M. L., & LUZAR, E. J. (1990). A Hedonic Analysis o P i a e Hun ing Land A ibu es
Using an Al e na i e Func ional Fo m. Jou nal o Ag icul u al and Applied Economics, 22(2),
129–135. h ps://doi.o g/10.1017/s1074070800001887
MEYER, D. (2019). Misc Func ions o he Depa men o S a is ics, P obabili y Theo y G oup.
MIT. (2016). Gi Hub - Py hon Zoma o API W appe . Re ie ed Janua y 23, 2020, om
h ps://gi hub.com/ a ihsucu/pyzoma o
MONTERO, J. M., MÍNGUEZ, R., & FERNÁNDEZ-AVILÉS, G. (2018). Housing p ice p edic ion:
pa ame ic e sus semi-pa ame ic spa ial hedonic models. Jou nal o Geog aphical Sys ems,
20(1), 27–55. h ps://doi.o g/10.1007/s10109-017-0257-y
MORO, M., MAYOR, K., & LYONS, S. (2013). Does he housing ma ke e lec cul u al he i age ? A
case s udy o G ea e Dublin. 45(1), 2884–2903. h ps://doi.o g/10.1068/a45524
NATIONAL OFFICE FOR STATISTICS. (2011). Census 2011 - Gene al Repo o Po ugal.
NEIS, P., & ZIPF, A. (2012). Analyzing he con ibu o ac i i y o a olun ee ed geog aphic in o ma ion
p ojec - The case o OpenS ee Map. ISPRS In e na ional Jou nal o Geo-In o ma ion.
h ps://doi.o g/10.3390/ijgi1020146
NÜST, D., GRANELL, C., HOFER, B., KONKOL, M., OSTERMANN, F. O., SILERYTE, R., &
CERUTTI, V. (2018). Rep oducible esea ch and GIScience: An e alua ion using AGILE
con e ence pape s. Pee J, 2018(7), e5072. h ps://doi.o g/10.7717/pee j.5072
OPENSTREETMAP, C. (2020). OpenS ee Map S a is ics. Re ie ed Janua y 23, 2020, om
h ps://www.opens ee map.o g/s a s/da a_s a s.h ml
OSGEO FOUNDATION. (2018). pgRou ing Manual Release 2.6.
PEREIRA, S. M. DA C. (2017). Explo ing he Rela ionship be ween Residen ial Ren s and House P ices
in he Po uguese Residen ial Es a e Ma ke .
POORT, J. (2015). Cul uu en c ea i i ei naa waa de gescha . (Sep embe ).
POSTGRESQL GLOBAL DEVELOPMENT GROUP. (2020). Pos GIS 3.0.1de Manual. Re ie ed
Janua y 22, 2020, om h ps://pos gis.ne /docs/index.h ml
PRESIDENT, M. M., & MARQUES, P. (2018). THE GLOBAL RENT GAP OF LISBON ’ S HISTORIC.
13(4), 683–694. h ps://doi.o g/10.2495/SDP-V13-N4-683-694
R CORE TEAM. (2019). R Language. Re ie ed Janua y 22, 2020, om h ps://c an. -
p ojec .o g/doc/manuals/ - elease/R-lang.h ml
ROGER BIVAND. (2019a). CRAN - Package spdep. Re ie ed Janua y 22, 2020, om h ps://c an. -
p ojec .o g/web/packages/spdep/index.h ml

54
ROGER BIVAND. (2019b). Spa ial Reg ession Analysis o R. Re ie ed Janua y 22, 2020, om
h ps://c an. -p ojec .o g/web/packages/spa ial eg/index.h ml
ROSEN, S. (1974). Hedonic P ices and Implici Ma ke s: P oduc Di e en ia ion in Pu e Compe i ion.
Jou nal o Poli ical Economy, 82(1), 34–55. h ps://doi.o g/10.1086/260169
SAFRONOV, O. (2017). NON-RESIDENTIAL REAL ESTATE PRICE EVALUATION OF LISBON ,
PORTUGAL NON-RESIDENTIAL REAL ESTATE PRICE EVALUATION OF LISBON ,
PORTUGAL.
SAGE, J., SMITH, D., & HUBBARD, P. (2012). The apidi y o s uden i ica ion and popula ion change:
The e goes he (S uden )hood. Popula ion, Space and Place. h ps://doi.o g/10.1002/psp.690
SCHÖLKOPF, B. (2002). Lea ning wi h ke nels. P oceedings o 2002 In e na ional Con e ence on
Machine Lea ning and Cybe ne ics. h ps://doi.o g/10.7551/mi p ess/4175.001.0001
SMOLA, A. J., & SCHÖLKOPF, B. (2004). A u o ial on suppo ec o eg ession. S a is ics and
Compu ing. h ps://doi.o g/10.1023/B:STCO.0000035301.49549.88
STEFAN MILTON BACHE AND HADLEY WICKHAM. (2014). mag i : A Fo wa d-Pipe Ope a o
o R. Re ie ed Janua y 22, 2020, om h ps://c an. -
p ojec .o g/web/packages/mag i /index.h ml
SUYKENS, J. A. K., & VANDEWALLE, J. (1999). Leas squa es suppo ec o machine classi ie s.
Neu al P ocessing Le e s. h ps://doi.o g/10.1023/A:1018628609742
THE POSTGRESQL GLOBAL DEVELOPMENT GROUP. (2014). Pos g eSQL: The wo ld’s mos
ad anced open sou ce da abase. Re ie ed Janua y 22, 2020, om h ps://www.pos g esql.o g/
TOBLER, W. (2004). On he i s law o geog aphy: A eply. Annals o he Associa ion o Ame ican
Geog aphe s. h ps://doi.o g/10.1111/j.1467-8306.2004.09402009.x
TROY, A., & GROVE, J. M. (2008). P ope y alues, pa ks, and c ime: A hedonic analysis in Bal imo e,
MD. Landscape and U ban Planning, 87(3), 233–245.
h ps://doi.o g/10.1016/j.landu bplan.2008.06.005
UNIPLACES. (2019). Uniplaces - Gene al. Re ie ed Janua y 25, 2020, om
h ps://www.uniplaces.com/abou -us
VAN DUIJN, M. (2013). Loca ion Choice , Cul u al He i age and House P ices.
VAN DUIJN, M., & ROUWENDAL, J. (2012). Cul u al he i age and he loca ion choice o Du ch
households in a esiden ial so ing model. Jou nal o Economic Geog aphy, 13(3), 473–500.
h ps://doi.o g/10.1093/jeg/lbs028
WANG, X., WEN, J., ZHANG, Y., & WANG, Y. (2014). Real es a e p ice o ecas ing based on SVM
op imized by PSO. Op ik, 125(3), 1439–1443. h ps://doi.o g/10.1016/j.ijleo.2013.09.017
WICKHAM, H. . & M. K. (2019). R Da abase In e ace [R package DBI e sion 1.1.0].
WITTEN, I. H., PAL, C. J., & FOURTH, M. (2017). Maximum Ma gin Hype plane Ex ending ins ance-
55
based and linea models.
ZANG, G., BERARDI, V., & REITERMANOVÁ, Z. (2010). Time se ies o ecas ing wi h neu al
ne wo k ensembles: An applica ion o exchange a e p edic ion. The Jou nal o he Ope a ional
Resea ch Socie y.
ZOMATO. (2019). Zoma o API. Re ie ed Janua y 23, 2020, om h ps://de elope s.zoma o.com/api
56
7. ANNEX
• Sample Que y – C ea ion o all eachable nodes om he lis ings wi hin 5KM
o hei icini y using he OSM Ne wo k and pgRou ing
• Sample Que y - Gene a ing Closes Nodes o Pa ks
57
• Sample Que y – C ea ion o Va iable: Coun s o T ams wi hin 2KM
58
• Sample Que y – C ea ion o Va iable: Minimum Dis ance o T ams

59
EVALUATION OF SPATIAL DATA’S IMPACT IN MID-TERM ROOM RENT PRICE THROUGH APPLICATION OF SPATIAL ECONOMETRICS AND MACHINE LEARNING
2020
Case S udy: Lisbon
Guia pa a a o ma ação de eses Ve são 4.0 Janei o 2006