# RECORD 67d132eca470e26b7a8c41462ca1252d056a93bb
# TITLE: Spatial and Social Media Data Analytics of Housing Prices in Shenzhen, China
# YEAR: 2016.0  VENUE: PLoS ONE  DOI: 10.1371/journal.pone.0164553
# AUTHORS: Chao Wu; Xin-Yue Ye; F. Ren; Y. Wan; Peng-Fei Ning; Qingyun Du
# PAGES: 19  CANDIDATE TABLE PAGES (1-based): [6, 13, 14]
# ABSTRACT: Housing is among the most pressing issues in urban China and has received considerable scholarly attention. Researchers have primarily concentrated on identifying the factors that influence residential property prices and how such mechanisms function. However, few studies have examined the potential factors that influence housing prices from a big data perspective. In this article, we use a big data perspective to determine the willingness of buyers to pay for various factors. The opinions and geographical preferences of individuals for places can be represented by visit frequencies given different motivations. Check-in data from the social media platform Sina Visitor System is used in this article. Here, we use kernel density estimation (KDE) to analyse the spatial patterns of check-in spots (or places of interest, POIs) and employ the Getis-Ord Gi* method to identify the hot spots for different types of POIs in Shenzhen, China. New indexes are then proposed based on the hot-spot results as measured by check-in data to analyse the effects of these locations on housing prices. This modelling is performed using the hedonic price method (HPM) and the geographically weighted regression (GWR) method. The results show that the degree of clustering of POIs has a significant influence on housing values. Meanwhile, the GWR method has a better interpretive capacity than does the HPM because of the former method’s ability to capture spatial heterogeneity. This article integrates big social media data to expand the scope (new study content) and depth (study scale) of housing price research to an unprecedented degree.


===== PAGE 1 (score 3, numeric-paren 0, table-words 0, var-words 3) =====

RESEARCH ARTICLE
Spatial and Social Media Data Analytics of
Housing Prices in Shenzhen, China
Chao Wu1, Xinyue Ye5, Fu Ren1,2,3, You Wan1,2,3, Pengfei Ning1, Qingyun Du1,2,3,4*
1 School of Resources and Environmental Science, Wuhan University, Wuhan, China, 2 Key Laboratory of
GIS, Ministry of Education, Wuhan University, Wuhan, China, 3 Key Laboratory of Digital Mapping and Land
Information Application Engineering, National Administration of Surveying, Mapping and Geoinformation,
Wuhan University, Wuhan, China, 4 Collaborative Innovation Center of Geospatial Technology, Wuhan
University, Wuhan, China, 5 Department of Geography, Kent State University, Kent, Ohio, United States of
America
* qydu@whu.edu.cn
Abstract
Housing is among the most pressing issues in urban China and has received considerable
scholarly attention. Researchers have primarily concentrated on identifying the factors that
influence residential property prices and how such mechanisms function. However, few
studies have examined the potential factors that influence housing prices from a big data
perspective. In this article, we use a big data perspective to determine the willingness of
buyers to pay for various factors. The opinions and geographical preferences of individuals
for places can be represented by visit frequencies given different motivations. Check-in
data from the social media platform Sina Visitor System is used in this article. Here, we use
kernel density estimation (KDE) to analyse the spatial patterns of check-in spots (or places
of interest, POIs) and employ the Getis-Ord G
i method to identify the hot spots for different
types of POIs in Shenzhen, China. New indexes are then proposed based on the hot-spot
results as measured by check-in data to analyse the effects of these locations on housing
prices. This modelling is performed using the hedonic price method (HPM) and the geo-
graphically weighted regression (GWR) method. The results show that the degree of clus-
tering of POIs has a significant influence on housing values. Meanwhile, the GWR method
has a better interpretive capacity than does the HPM because of the former method’s ability
to capture spatial heterogeneity. This article integrates big social media data to expand the
scope (new study content) and depth (study scale) of housing price research to an unprece-
dented degree.
Introduction
Residential property is a multidimensional and durable commodity, and its value is determined
by a combination of characteristics categorized as structural, locational and neighbourhood
attributes [1–3]. However, these attributes do not have individual market prices. Numerous
studies have explored the relationships between housing prices and specific attributes by
PLOS ONE | DOI:10.1371/journal.pone.0164553
October 26, 2016
1 / 19
a11111
OPEN ACCESS
Citation: Wu C, Ye X, Ren F, Wan Y, Ning P, Du Q
(2016) Spatial and Social Media Data Analytics of
Housing Prices in Shenzhen, China. PLoS ONE 11
(10): e0164553. doi:10.1371/journal.
pone.0164553
Editor: Kim-Kwang Raymond Choo, University of
Texas at San Antonio, UNITED STATES
Received: May 26, 2016
Accepted: September 8, 2016
Published: October 26, 2016
Copyright: © 2016 Wu et al. This is an open access
article distributed under the terms of the Creative
Commons Attribution License, which permits
unrestricted use, distribution, and reproduction in
any medium, provided the original author and
source are credited.
Data Availability Statement: POI data are from
http://open.weibo.com/. We used Open API
interface to get check-in data. The relevant data are
within the paper and its Supporting Information
files. Attributes of house data are from http://sz.
fang.com/. The relevant data are within the paper
and its Supporting Information files. Housing price
data are from Shenzhen Research Centre for Digital
City Engineering. Interested readers and
researchers can contact with Shenzhen
development research center using the
284158541@qq.com. Interested readers and
researchers can cooperate with Shenzhen


===== PAGE 6 (score 29, numeric-paren 0, table-words 3, var-words 19) =====

or fake users [45]. We filter the check-in data using the following critera (i) the location of a
check-in is not in Shenzhen based on the geographical location; (ii) the name of the POI is not
correct based on the text discriminant; (iii) we believe that users who have only one check-in
record are not valid users and that their behaviours are accidental in a given time period(July
2014 and June 2015 in this article); (iiii) and based on the third criterion, we filter valid POIs
that are signed by at least one valid user. Finally, we acquire 447,778 check-in records from
216,165 users and 22,670 POI. To ensure the privacy of personally identifiable information
(PII) [46], we do not identify the information of users. This article uses 13,268 CBF sites and
1,413 GRE sites identified by researching the locations and names of POIs based on the Code
for Classification of Urban Land Use and the Planning Standards of Development Land [47].
The CBF includes commercial facilities, business facilities, entertainment and sports facilities,
and public utility outlets, and the GRE consists primarily of parks, green belts and urban
squares. POI types and aggregated information can be seen in Table 2.
Methods
Spatial hot-spot analysis method based on check-in data
In this section, we present our proposed method for using check-in data to detect hot spots.
Each check-in record includes the user ID, time, coordinates (longitude and latitude), POI ID,
Table 1. Measurement methods and housing characteristic variable signs.
Variable
Variable definition and measurement methods
Mean
Std.dev.
Expected sign
Structural Variables
Area
Square footage of the living area of house (m2)
100.487
35.689
+
Floor
Floor on which the unit is situated (floor)
18.018
10.067
+
NumBed
Number of bedrooms in unit
3.139
0.965
+
NumWash
Number of bathrooms in unit
1.725
0.643
Unknown
Fee
Property management fees (RMB)
3.868
1.069
+
RPlot
Ratio of floor area
3.976
1.404
-
RGreen
Ratio of green space area (%)
0.345
0.0729
+
Locational Variables
DCBD
Distance to CBD (central business district; km)
23.339
10.208
-
DSub
Distance to nearest metro station (km)
0.319
0.409
-
DBus
Distance to nearest bus station (km)
0.019
0.023
-
DHospital
Distance to nearest hospital (km)
0.122
0.786
Unknown
DSnursery
Distance to nearest nursery school (km)
0.597
0.464
-
DSprimary
Distance to nearest primary school (km)
0.756
0.464
-
DSmiddle
Distance to nearest middle school (km)
0.108
0.640
-
DPark
Distance to nearest park(km)
0.113
0.609
-
Neighbourhood Variables
GiZScore_G
Degree of activity in green space
0.438
2.308
-
GiZScore_C
Degree of activity in commercial centre
1.181
3.229
-
doi:10.1371/journal.pone.0164553.t001
Table 2. POI types and aggregated information.
Type
Abbreviation
Counts
Percentage
Commercial and business facilities
CBF
13,268
59%
Industrial
IND
890
3.93%
Transport facilities
TRA
1,696
7.48%
Residence communities
RES
3,569
15.74%
Green space
GRE
1,413
6.23%
Administration and public services
APS
1,834
8.09%
Total
22,670
100%
doi:10.1371/journal.pone.0164553.t002
Analytics of Housing Prices
PLOS ONE | DOI:10.1371/journal.pone.0164553
October 26, 2016
6 / 19


===== PAGE 7 (score 5, numeric-paren 0, table-words 0, var-words 5) =====

and name and category of the venue. We use address-coding technology to locate records for
our database and detect hot spots using the number of check-ins, which are obtained from POI
data [48]. The POI data include the POI ID, coordinates, check-in totals, user counts and cate-
gories. We measure the hot spots by identifying their relationship with different types of POIs
and exploring the effects of CBF and GRE.
Before detecting the hot-spot status of the different types of POIs, we analyse the overall
spatial patterns. The kernel density estimation (KDE) is used to analyse the spatial distribution
of all check-in POI data. KDE attempts to produce a smooth density surface of spatial point
events in geographic space [49]. The general form of a kernel density estimator is given by
f ðsÞ ¼
Xn
i¼1
1
ph2 k dis
h


ð1Þ
where f(s) is the KDE function at location s, h is the bandwidth of the KDE, dis, is the distance
from point i to location s, and the function k (called the kernel function) represents the spatial
weight function of the ratio between dis and the bandwidth h. Briefly, the results of the KDE
rely on kernel sizes (called the search radius) and grid sizes. In this analysis, first, individual
output cells were partitioned in 100m100m areas, which can provide greater precision in the
estimation most efficiently. Because too small of a grid size can increase the computational
cost, whereas too large of a size can result in important details being ignored. Second, we select
a search radius of 1000 m based on multiple testing rounds. Thresholds of less than 1000m,
such as 600m and 800m, make the density distribution focus on a limited extent of elemental
points, and the overall characteristic are not obvious. Conversely, thresholds of greater 1000m,
such as 1200m and 1500m, make the density distribution overly smooth, and local differences
cannot be well represented. We set 1000 m as the ideal threshold because it is the longest dis-
tance that remains comfortable to travel on foot (10–15 minutes) and obtains relatively good
results in this paper.
After kernel density analysis, we rank the POIs based on the counts of check-ins and users
for all check-in spots to detect the POIs that are popular. The power-law distribution patterns
(Fig 2) show that there are far more POIs with fewer check-ins and users than POIs with many
check-ins and users. The patterns are similar to those in a previous study [33]. The results of
Fig 2 suggest that POIs with high popularity, represented by the numbers of check-ins and
users, provide more attractions to residents and are better known to people. Differences in peo-
ples’ preferences and usages for POIs reveal that the formation of hot spots and cold spots for
POIs is reasonable. We apply the ‘average nearest neighbour’ index to judge whether the spatial
distribution of points is clustered or dispersed. This process is conducted for GRE and CBF
separately. The clustering patterns of GRE and CBF are used to determine whether they are hot
spots.
Because of the discrete distribution of points and simplification of the amount of big data
[50, 51], we use grids to replace points to analyse the spatial patterns [52]. First, we divide the
study area into 500 m x 500 m spaces and then apply the grid. Testing determined that 500 m
is the optimal choice because a grid size larger than 500 m tends to mask local differences and
because a grid size smaller than 500 m can exaggerate local characteristics. We then calculate
the total number of check-in records and the number of POIs for each grid. We explore the
possible hot spots using the grids and then use Getis-Ord G
i to identify the hot spots. The
Getis-Ord G
i statistical clustering method [53, 54] is used to detect significant clusters of high
value within a city according to locations that contain many people. Getis-Ord G
i is a local sta-
tistic used to assess each feature within the context of neighboring features. This article defines
a ‘hot spot’ as a place with a high degree of activity. The statistics are calculated as follows:
Analytics of Housing Prices
PLOS ONE | DOI:10.1371/journal.pone.0164553
October 26, 2016
7 / 19


===== PAGE 13 (score 30, numeric-paren 1, table-words 8, var-words 9) =====

with the findings of previous studies [58–60]. The structural attributes, such as the number of
washrooms, floor height, area, property management fees and green ratio, all have effects on
housing prices and add 9.4%, 6.3%, 29.3%, and 2.3%, respectively, of extra value to the price of
housing with a 1- unit increase. The effects of location variables are similar as expected with
the exception of distance to primary schools. A 1- km increase in the distance to parks, hospi-
tals, bus stations and subway stations reduces housing prices by 11.1%, 9.4%, 4.9% and 4.9%,
respectively. Primary schools are an exception because of the policy of school district housing
in China. The effect of school location on housing prices is determined by distance as well as
school quality.
Geographicallyweighted regression results. The results of the local regression analysis
(GWR) are summarized in Table 5. As previously mentioned, this method evaluates spatial
variations in the relationships between the variables and housing prices. In this model, the R2
value increases from 0.724 (HPM) to 0.9399 (GWR), thereby showing that the GWR method
has a better interpretive capability than the HPM. The GWR can explain 93.99% of the varia-
tion according to the R2 value. We also measure the spatial auto-correlation (Moran’s I
index = 0.041, z-score = 3.86) by calculating Moran’s I index for the residuals of the GWR to
ensure that the validity of the model for the residuals is random. Additionally, the AICc of the
GWR is -74703.46207, and the bandwidth equals 0.337, which are less than that of the HPM
(AICc = -36082.16912). The difference in the AICc of greater than 3 suggests that the GWR is
more representative of reality [56]. Table 5 shows that all the variables are significant at 1% or
5% with the exception of DPark (p-value = 0.855). Additionally, the outstanding achievement
of the GWR method is its ability to visualize local parameters. Fig 7 presents the GiZscore_C
and GiZscore_G. Because of the discrete distribution of points, we use the inverse distance
weighted (IDW) interpolation method to obtain a better visual effect. The output cell size of
the IDW processes is set to 100 m, and a natural breaks (Jenks) classification is applied. Con-
sidering that the coefficientshave positive and negative values, we set a zero value as a class
limit to distinguish the positive and negative effects.
Table 4. Hedonic price method (HPM) parameter estimate summary.
Explanatory variable
Estimated coefficient
t-ratio
p-value
Constant
0.060
19.388
.000***
DCBD
-0.326
-55.863
.000***
Fee
0.293
43.604
.000***
DPark
-0.111
-27.376
.000***
NumWash
0.094
21.743
.000***
GiZScore_C
0.083
17.104
.000***
DHospital
-0.094
-24.942
.000***
SPrimary
0.080
21.966
.000***
Floor
0.063
20.077
.000***
GiZScore_G
0.134
24.466
.000***
DBus
-0.049
-14.273
.000***
DSub
-0.049
-11.417
.000***
Area
0.031
6.907
.000***
RGreen
0.023
6.854
.000***
R2
0.724
*** Significant at the 1% level.
doi:10.1371/journal.pone.0164553.t004
Analytics of Housing Prices
PLOS ONE | DOI:10.1371/journal.pone.0164553
October 26, 2016
13 / 19


===== PAGE 14 (score 25, numeric-paren 1, table-words 3, var-words 9) =====

The regression coefficientfor GiZscore_C is greater than zero because major concentrations
are observedin Futian, Longhua, central and northeast Nanshan, the southern part of Bao’an
and Guangming and central Dapeng. The CBF is more strongly clustered, which is associated
with more check-ins, and the housing prices are higher in these locations, which represent the
most prosperous areas of each district and have more complete facilities and more workplaces.
These areas provide convenient amenities to housing buyers, such as employment, cafes, shop-
ping and entertainment. Therefore, the housing prices are high. However, Fig 7(a) presents
unexpectedresults for Longgang, Luohu and Yantian, with lower housing prices despite the
greater clustering (with more check-ins) of the commercial and business facilities. GiZscore_C
Table 5. Geographically weighted regression (GWR) parameter estimate summary.
Explanatory Variable
Minimum
Lower quartile
Median
Upper quartile
Maximum
p-value
Constant
-12.8186
-1.38137
-0.2675
0.04614
10.28509
.000***
Floor
-0.35291
0.031246
0.040169
0.056628
0.339561
.000***
NumBed
-0.66773
-0.02149
0.040887
0.128011
0.493299
.000***
NumWash
-0.64065
-0.0432
0.022523
0.06847
0.358014
.000***
Area
-0.68338
-0.01717
0.074358
0.142633
1.335995
.000***
RPlot
-10.9418
-0.5452
-0.1074
0.145695
1.900368
.000***
RGreen
-3.84742
-0.0312
0.089782
0.399657
1.858314
.000***
Fee
-2.68112
0.015945
0.25459
0.621551
4.052185
.000***
DSub
-12.302
-0.40319
-0.02085
0.563174
15.22164
.000***
DBus
-4.17797
-0.35394
-0.03028
0.509983
23.5197
.004**
DSnursery
-18.9376
-0.1124
-0.03278
0.158163
4.604339
.000***
DSprimary
-12.4242
-0.13258
0.017652
0.261109
2.282185
.000***
DSmiddle
-3.55686
-0.41095
0.017384
0.249635
14.10076
.000***
DHospital
-4.868
-0.62254
-0.12827
0.300655
5.447187
.000***
DPark
-2.27624
-0.07367
0.012992
0.328673
11.93626
.855
DCBD
-16.3229
-0.74499
-0.09705
0.430907
1.94271
.000***
GiZScore_G
-11.9198
-2.41877
-0.1252
0.179357
4.508242
.000***
GiZScore_C
-9.16523
0.015567
0.463302
2.229864
11.68491
.000***
R2
0.9399
Bandwidth
0.337
***, ** Significant at the 1% and 5% levels, respectively.
doi:10.1371/journal.pone.0164553.t005
Fig 7. Spatial variation of the local parameters of the GWR. (a) and (b) are GiZscore_C and GiZscore_G, respectively.
doi:10.1371/journal.pone.0164553.g007
Analytics of Housing Prices
PLOS ONE | DOI:10.1371/journal.pone.0164553
October 26, 2016
14 / 19


===== PAGE 15 (score 12, numeric-paren 0, table-words 2, var-words 8) =====

has a negative effect on housing prices, and because of the development of Futian and Nan-
shan, the position of Luohu has been declining. Luohu is an old urban district that is currently
facing the problem of an old-age demographic bump. Senior citizens tend to choose housing in
locations with a good environment characterized by peace and quiet. The clustering of resi-
dents can cause noise and traffic jams, which may affect the residents’ daily lives and rest. Yan-
tian has relatively poor traffic conditions and no subway system. Thus, the clustering of CBF in
Yantian may result in worse traffic conditions. Similarly, GiZscore_C has a negative effect on
housing prices in Yantian. Longgang is a relatively poorly developed area in Shenzhen and has
comparatively low housing prices. Housing buyers in Longgang pay more attention to trans-
portation and the convenience of getting to work. Thus, it is not difficultto imagine that GiZs-
core_C would have a negative effect on housing prices in Longgang.
The GWR model reveals that the relationship between GiZscore_G and housing prices is
not stationary across the study area, as illustrated in Fig 7(b). Most major areas show a positive
effect of GiZscore_C on housing prices, with the exception of the south of Bao’an, the west
and centre of Longgang, and most of Yantian. Yantian was the first ‘national biome’ in South
China, and the air quality is good, which attracts many people. However, Yantian’s traffic con-
ditions are relative poor, and it is located in a fault zone in Shenzhen. Therefore, it is likely that
people prefer to visit this scenic spot for recreational purposes rather than live there.
Conclusions
This article suggests a framework for applying a GWR model and a big data perspective for
evaluating the influence of urban hot spots on housing prices in Shenzhen, China. This work
contributes to the field of housing price research by identifying POI hot spots through the anal-
ysis of check-in data and by exploring the influence of a POI’s degree of activity (from the
perspective of popular perception) on housing prices. Despite the widespread use of crowd-
sourced geographic data (CSGD) [49, 61–64], few studies have used these data to examine
housing prices, which are closely tied to public opinion and sentiments. We apply GWR
method to test the hypothesis that hot spots (identified using check-in data) have significant
effects on housing values. Our results are consistent with expectations. In this article, we intro-
duce the use of check-in data, which represent a novel data source in the field of housing price
research. Additionally, we take spatial heterogeneity into consideration and explore the factors
that influence housing prices for different locations. Finally, we chose Shenzhen, China as our
case study because housing prices in Shenzhen have skyrocketed in recent years, which is an
area of continued concern [61]. This article integrates big social media data and expands the
scope (new research content) and depth (study scale) of housing price research to an unprece-
dented degree.
This article applies the Getis-Ord G
i technique to detect hot-spot areas of CBF and GRE.
We then explore the relationship between housing prices and hot spots using hedonic price
modelling and geographical regression modelling. The R2 of the HPM is 0.724, which indicated
that the HPM can be used to analyse the effects of hot spots from a global perspective. CBF and
GRE hot spots both increase housing prices. The R2 of the GWR is 0.9399, which indicates a
stronger interpretive capability compared with the HPM. The results of the GWR reveal that
the effects of influential factors vary over space, as represented by different coefficients. Finally,
we select typical and special areas for our interpretation.
Most of the data values require innovative methods of analysis to identify their potential
uses [65]. Although this article provides a new perspective for the application of big social
media data to the study of housing prices, it still has several limitations.
Analytics of Housing Prices
PLOS ONE | DOI:10.1371/journal.pone.0164553
October 26, 2016
15 / 19