1

Image Based Appraisal of Real Estate Properties

Quanzeng You, Ran Pang, Liangliang Cao, and Jiebo Luo, Fellow, IEEE

Images reviews [3] and so on. Generally speaking, there are several
different types of appraisal values. In particular, we are in-
terested in the market value, which refers to the trade price
in a competitive Walrasian auction setting [4]. Today, people
are likely to trade through real estate brokers, who provide
easy access online websites for browsing real estate property
in an interactive and convenient way. Fig. 1 shows an example
arXiv:1611.09180v2 [cs.CV] 27 Jul 2017

of house listing from Realtor (http://www.realtor.com/), which

is the largest real estate broker in North America. From the
figure, we see that a typical piece of listing on a real estate
property will introduce the infrastructure data in text for the
house along with some pictures of the house. Typically, a
buyer will look at those pictures to obtain a general idea of
the overall property in a selected area before making his next
move.
Traditionally, both real estate industry professionals and
researchers have relied on a number of factors, such as
Basic Property
economic index, house age, history trade and neighborhood
Detailed Information
environment [5] and so on to estimate the price. Indeed, these
Fig. 1. Example of homes for sale from Realtor. factors have been proved to be related to the house price, which
is quite difficult to estimate and sensitive to many different
human activities. Therefore, researchers have devoted much
I. I NTRODUCTION effort in building a robust house price index [6], [7], [8], [9].
In addition, quantitative features including Area, Year, Storeys,
R Eal estate appraisal, which is the process of estimating
the price for real estate properties, is crucial for both
buys and sellers as the basis for negotiation and transaction.
Rooms and Centre [10], [11] are also employed to build
neural network models for estimating house prices. However,
Real estate plays a vital role in all aspects of our contemporary pictures, which is probably the most important factor on a
society. In a report published by the European Public Real Es- buyer’s initial decision making process [12], have been ignored
tate Association (EPRA http://alturl.com/7snxx), it was shown in this process. This is partially due to the fact that visual
that real estate in all its forms accounts for nearly 20% of the content is very difficult to interpret or quantify by computers
economic activity. Therefore, accurate prediction of real estate compared with human beings.
prices or the trends of real estate prices help governments A picture is worth a thousand words. One advantage with
and companies make informed decisions. On the other hand, images and videos is that they act like universal languages.
for most of the working class, housing has been one of the People with different backgrounds can easily understand the
largest expenses. A right decision on a house, which heavily main content of an image or video. In the real estate industry,
depends on their judgement on the value of the property, can pictures can easily tell people exactly how the house looks like,
possibly help them save money or even make profits from their which is impossible to be described in many ways using lan-
investment in their homes. From this perspective, real estate guage. For the given house pictures, people can easily have an
appraisal is also closely related to people’s lives. overall feeling of the house, e.g. what is the overall construc-
tion style, how the neighboring environment looks like. These
Current research from both estate industry and academia
high-level attributes are difficult to be quantitatively described.
has reached the conclusion that real estate value is closely
On the other hand, today’s computational infrastructure is
related to property infrastructure [1], traffic [2], online user
also much cheaper and more powerful to make the analysis
Copyright (c) 2013 IEEE. Personal use of this material is permitted. of computationally intensive visual content analysis feasible.
However, permission to use this material for any other purposes must be Indeed, there are existing works on focusing the analysis of
obtained from the IEEE by sending a request to pubs-permissions@ieee.org. visual content for tasks such as prediction [13], [14], and
Manuscript received March 28, 2016; accepted February 10, 2017.
Q. You and J. Luo are with the Department of Computer Science, online user profiling [15]. Due to the recently developed deep
University of Rochester, Rochester, NY 14623 USA. E-mail: {qyou, learning, computers have become smart enough to interpret
jluo}@cs.rochester.edu. visual content in a way similar to human beings.
R. Pang is with PayPaI. E-mail: pangrr89@gmail.com
L. Cao is with Electrical Engineering & Computer Sciences, Columbia Recently, deep learning has enabled robust and accurate
University and customerserviceAI. E-mail: liangliang.cao@gmail.com feature learning, which in turn produces the state-of-the-art
 2

performance on many computer vision related tasks, e.g. digit focused on building price indexes for real properties. The
recognition [16], [17], image classification [18], [19], aesthet- seminal work in [6] built price index according to the repeat
ics estimation [20] and scene recognition [21]. These systems prices of the same property at different times. They employed
suggest that deep learning is very effective in learning robust regression analysis to build the price index, which shows good
features in a supervised or unsupervised fashion. Even though performances. Another widely used regression model, Hedonic
deep neural networks may be trapped in local optima [22], regression, is developed on the assumption that the character-
[23], using different optimization techniques, one can achieve istics of a house can predict its price [7], [8]. However, it
the state-of-the-art performance on many challenging tasks is argued that the Hedonic regression model requires more
mentioned above. assumptions in terms of explaining its target [28]. They also
Inspired by the recent successes of deep learning, in this mentioned that for repeat sales model, the main problem is
work we are interested in solving the challenging real estate lack of data, which may lead to failure of the model. Recent
appraisal problem using deep visual features. In particular, for work in [9] employed locations and sale price series to build
images related tasks, Convolutional Neural Network (CNN) an autoregressive component. Their model is able to use both
are widely used due to the usage of convolutional layers. It single sale homes and repeat sales homes, which can offer a
takes into consideration the locations and neighbors of image more robust sale price index.
pixels, which are important to capture useful features for visual More studies are conducted on employing feed forward
tasks. Convolutional Neural Networks [24], [18], [19] have neural networks for real estate appraisal [29], [30], [31], [32].
been proved very powerful in solving computer vision related However, their results suggest that neural network models
tasks. are unstable even using the same package with different run
We intend to employ the pictures for the task of real estate times [29]. The performance of neural networks are closely
price estimation. We want to know whether visual features, related to the features and data size [32]. Recently, Kontrimas
which is a reflection of a real estate property, can help and Verikas [33] empirically studied several different models
estimate the real estate price. Intuitively, if visual features can on selected 12 dimensional features, e.g. type of the house,
characterize a property in a way similar to human beings, we size, and construction year. Their results show that linear
should be able to quantify the house features using those visual regression outperforms neural network on their selected 100
responses. Meanwhile, real estate properties are closely related houses.
to the neighborhood. In this work, we develop algorithms More recent studies in [1] propose a ranking objective,
which only rely on 1) the neighbor information and 2) the which takes geographical individual, peer and zone dependen-
attributes from pictures to estimate real estate property price. cies into consideration. Their method is able to use various
To preserve the local relation among properties we employ estate related data, which helps improve their ranking results
a novel approach, which employs random walks to generate based on properties’ investment values. Furthermore, the work
house sequences. In building the random walk graph, only the in [3] studied online user’s reviews and mobile users’ mov-
locations of houses are utilized. In this way, the problem of real ing behaviors on the problem of real estate ranking. Their
estate appraisal has been transformed into a sequence learning proposed sparsity regularized learning model demonstrated
problem. Recurrent Neural Network (RNN) is particularly competitive performance.
designed to solve sequence related problems. Recently, RNNs In contrast, we are trying to solve this problem using the
have been successfully applied to challenging tasks including attributes reflected in the visual appearances of houses. In
machine translation [25], image captioning [26], and speech particular, our model does not use the meta data of a house
recognition [27]. Inspired by the success of RNN, we deploy (e.g. size, number of rooms, and construction year). We intend
RNN to learn regression models on the transformed problem. to utilize the location information in a novel way such that
The main contributions of our work are as follows: our model is able to use the state-of-the-art deep learning for
• To the best of our knowledge, we are the first to quantify feature extraction (Convolutional Neural Network) and model
the impact of visual content on real estate price esti- learning (Recurrent Neural Network).
mation. We attribute the possibility of our work to the
newly designed computer vision algorithms, in particular III. R ECURRENT N EURAL N ETWORK FOR R EAL E STATE
Convolutional Neural Networks (CNNs). P RICE E STIMATION
• We employ random walks to generate house sequences
In this section, we present the main components of our
according to the locations of each house. In this way, we framework. We describe how to transform the problem into a
are able to transform the problem into a novel sequence problem that can be solved by the Recurrent Neural Network.
prediction problem, which is able to preserve the relation The architecture of our model is also presented.
among houses.
• We employ the novel Recurrent Neural Networks (RNNs)
to predict real estate properties and achieve accurate A. Random Walks
results. One main feature of real estate properties is its location. In
particular, for houses in the same neighborhood, they tend to
II. R ELATED W ORK have similar extrinsic features including traffic, schools and
Real estate appraisal has been studied by both real estate so on. We build an undirected graph G for all the houses
industrial professionals and academia researchers. Earlier work collected, where each node vi represent the i-th house in our
 3

data set. The similarity sij between house hi and house hj is RNN with gradient descent. Fig. 2 shows the details of a
defined using the Gaussian kernel function, which is a widely single Long Short-Term Memory (LSTM) block [38]. Each
used similarity measure1 : LSTM cell contains an input gate, an output gate and an forget

dist(hi , hj )
 gate, which is also called a memory cell in that it is able to
sij = exp , (1) remember the error in the error propagation stage [39]. In this
2σ 2
way, LSTM is capable of modeling long-range dependencies
where dist(hi , hj ) is the geodesic distance between house hi than conventional RNNs.
and hj . σ is the hyper-parameter, which controls the similarity
decaying velocity with the increase of distance. In all of our

experiments, we set σ to 0.5 miles so that houses within the
1.5 (within 3σ) miles will have a relatively larger similarity.
Input Gate
The -neighborhood graph [34] is employed to build G in our ht-1
implementation. We assign the weight of each edge eij as the it
Cell
similarity sij between house hi and the house hj .
ct  ht
Given this graph G, we can then employ random walks Forget Gate
to generate sequences. In particular, every time, we randomly ft
choose one node vi as the root node, then we proportionally xt
jump to its neighboring nodes vj according to the weights
between vi and its neighbors. The probability of jumping to
node vj is defined as ot
Output Gate
eji
pj = P , (2)
k∈N (i) eki Fig. 2. An illustration of a single Long Short-Term Memory (LSTM) Cell.
where N (i) is the set of neighbor nodes of vi . We continue
to employ this process until we generate the desired length For completeness, we give the detailed calculation of ht
of sequence. The employment of random walks is mainly given input xt and ht−1 in the following equations. Let
motivated by the recent proposed DeepWalk [35] to learn W.i , W.f , W.o represent the parameters related to input,
feature representations for graph nodes. It has been shown that forget and output gate respectively.  denotes the element-
random walks can capture the local structure of the graphs. In wise multiplication between two vectors. φ and ψ are some
this way, we can keep the local location structure of houses selected activation functions and σ is the fixed logistic sigmoid
and build sequences for houses in the graph. Algorithm 1 function. Following [38], [27], [40], we employ tanh for both
summarizes the detailed steps for generating sequences from φ in Eq.(6) and ψ in Eq.(8).
a similarity graph.
it = σ(Wxi xt + Whi ht−1 + Wci ct−1 + bi ) (4)
We have generated sequences by employing random walks.
In each sequence, we have a number of houses, which is ft = σ(Wxf xt + Whf ht−1
+ Wcf ct−1 + bf ) (5)
related in terms of their locations. Since we build the graph on ct = ft  ct−1 + it  φ(Wxc xt + Whc ht−1 + bc ) (6)
top of house locations, the houses within the same sequence ot = σ(Wxo xt + Who ht−1 + Wco ct + bo ) (7)
are highly possible to be close to each other. In other words,
ht = ot  ψ(ct ) (8)
the prices of houses in the same sequence are related to
each other. We can employ this context for estimating real
estate property price, which can be solved by recurrent neural C. Multi-layer Bidirectional LSTM
network discussed in following sections. In previous sections, we have discussed the generation of
sequences as well as Recurrent Neural Network. Recall that
B. Recurrent Neural Network we have built an undirected graph in generating the sequences,
With a Recurrent Neural Network (RNN), we are trying to which indicates that the price of one house is related to all
predict the output sequence {y1 , y2 , . . . , yT } given the input the houses in the same sequence including those in the later
sequence {x1 , x2 , . . . , xT }. Between the input layer and the part. Bidirectional Recurrent Neural Network (BRNN) [41]
output layer, there is a hidden layer, which is usually estimated has been proposed to enable the usage of both earlier and
as in Eq.(3). future contexts. In bidirectional recurrent neural network, there
is an additional backward hidden layer iterating from the last
ht = ∆(Whi ht−1 + Wx xt + bh ) (3) of the sequence to the first. The output layer is calculated by
∆ represents some selected activation function or other com- employing both forward and backward hidden layer.
plex architecture employed to process the input xt and ht . Bidirectional-LSTM (B-LSTM) is a particular type of
One of the most widely deployed architectures is Long Short- BRNN, where each hidden node is calculated by the long
Term Memory (LSTM) cell [36], which can overcome the short-term memory as shown in Fig. 2. Graves et al. [40] have
vanishing and exploding gradient problem [37] when training employed Bidirectional-LSTM for speech recognition. Fig. 3
shows the architecture of the bidirectional recurrent neural
1 http://en.wikipedia.org/wiki/Radial basis function kernel network. We have two Bidirectional-LSTM layers. During the
 4

Output Layer y1 yt yt yt1 yT

2nd Backward Layer h1b 2  htb21 htb 2 htb21  hTb 2

2nd Forward Layer h1f 2  ht f12 ht f 2 ht f12  hTf 2

1st Backward Layer h1b1  htb11 htb1 htb11  hTb1

1st Forward Layer h1f 1  ht f 11 ht f 1 ht f11  hTf 1

Input Layer x1 xt 1 xt xt 1 xT

Fig. 3. The Multi-layer Bidirectional Recurrent Neural Network (BRNN) architecture for real estate price estimation. There are two bidirectional recurrent
layers in this architecture. For real estate price estimation, the price of each house is related to all houses in the same sequence, which is the main motivation
to employ bidirectional recurrent layers.

forward pass of the network, we calculate the response of both generated i-th sequence and ŷij is the corresponding estimated
the forward and the backward hidden layers in the 1st-LSTM price for this house.
and 2nd-LSTM layer respectively. Next, the output (in our When training our Multi-Layer B-LSTM model, we employ
problem, the output is the price of each house) of each house the RMSProp [42] optimizer, which is an adaptive method for
is calculated using the output of the 2nd-LSTM layer as input automatically adjust the learning rates. In particular, it normal-
to the output layer. izes the gradients by the average of its recent magnitude.

Algorithm 1 RandomWalks Algorithm 2 Training Multi-Layer B-LSTM

Require: H = {h1 , h2 , . . . , hn } geo-coordinates of n houses Require: H = {h1 , h2 , . . . , hn } geo-coordinates of n houses
σ hyper-parameter for Gaussian Kernel X = {x1 , x2 , . . . , xn } features of the n house
t threshold for distance Y = {y1 , y2 , . . . , yn } prices of the n houses
M total number of desired sequences 1: S = RandomWalks (see Algorithm 1)
1: Calculate the Vincenty distance between any pair of 2: Split S into mini-batches
houses 3: repeat
2: Calculate the similarity between houses according to the 4: Calculate the gradient of L in Eq.(9) and update the
Gaussian kernel function (see Eq.(1)). parameters using RMSProp.
3: repeat 5: until Convergence
4: Initialize sc = {} 6: return The learned model M
5: Randomly pick one node hi and add hi to sc
6: set hc = hi We conduct the back propagation in a mini-batch approach.
7: while size(sc ) < L do Algorithm 2 summarizes the main steps for our proposed
8: Pick hc ’s neighbor node hj with probability pj algorithm.
defined in Eq.(2)
D. Prediction
9: add hj to sc
10: set hc = hj In the prediction stage, the first step is also generating
11: end whileadd sc to S sequence. For each testing house, we add it as a new node into
12: until size (S) = M our previously build similarity graph on the training data. Each
13: return The set of sequence S testing house is a new node in the graph. Next, we add edges
to the testing nodes and the training nodes. We use the same
settings when adding edges to the new -neighborhood graph.
The objective function for training the Multi-Layer Bidirec-
Given the new graph G0 , we randomly generate sequences and
tional LSTM is defined as follows:
keep those sequences that contain one and only one testing
N
1 XX node. In this way, for each house, we are able to generate
L= k ŷij − yij k2 (9) many different sequences that contain this house. Fig. 4 shows
N n=1 j
the idea. Each testing sequence only has one testing house.
where W is the the set of all the weights between different The remaining nodes in the sequence are the known training
layers.yij is the actual trade price for the j-th house in the houses.
 5

a) Average: The above strategy implies that we are able TABLE I

to build many different sequences for each testing house. T HE AVERAGE PRICE PER S QFT AND THE STANDARD DEVIATION ( STD ) OF
THE PRICE OF THE TWO STUDIED CITIES .
To obtain the final prediction price for each testing house,
one simple strategy is to average the prediction results from City # of Houses Avg Price std of Price
different sequences and report the average price as the final San Jose 3064 454.2 132.1
prediction price. Rochester 1500 76.4 21.2

testing
the houses in our collected data from San Jose and Rochester
h1  ht-1 ht ht+1  hT using the returned geo-locations from Bing Map API.
According to these coordinates, we are able to calculate the
distance between any pair of houses. In particular, we employ
known Vincenty distance (https://en.wikipedia.org/wiki/Vincenty’s
formulae) to calculate the geodesic distances according to the
Fig. 4. Testing sequence h1 → h2 → · · · → hT . In each testing sequence, coordinates. Fig. 7 shows distribution of the distance between
there is one and only one testing node in that sequence. The remaining nodes
are all come from training data. any pair of houses in our data set. The distance is less than
4 miles for most randomly picked pair of houses. In building
our -neighborhood graph, we assign an edge between any pair
IV. E XPERIMENTAL R ESULTS of houses, which has a distance smaller than 5 miles ( = 5
miles).
In this section, we discuss how to collect data and evaluate
the proposed framework as well as several state-of-the-art
approaches. In this work, all the data are collected from
Realtor (http://www.realtor.com/), which is the largest realtor B. Feature Extraction and Baseline Algorithms
association in North America. We collect data from San Jose, In our implementation, we experimented with GoogleNet
CA, one of the most active cities in U.S., and Rochester, NY, model [43], which is one of the state-of-the-art deep neural
one of the least active cities in U.S., over a period of one architectures. In particular, we use the response from the last
year. In the next section, we will discuss the details on how avg − pooling layer as the visual features for each image. In
to preprocess the data for further experiments. this way, we obtain a 1, 024 dimensional feature vector for
each image. Each house may have several different pictures
A. Data Preparation on different angles of the same property. We average features
The data collected from Realtor contains description, school of all the images of the same house (also known as average-
information and possible pictures about each real property pooling)2 to obtain the feature representation of the house.
as shown in Fig. 1 show. We are particularly interested in We compare the proposed framework with the following
employing the pictures of each house to conduct the price algorithms.
estimation. We filter out those houses without image in our 1) Regression Model (LASSO): Regression model has been
data set. Since houses located in the same neighborhood seem employed to analyze real estate price index [6]. Recently, the
to have similar price, the location is another important features results in Fu et al. [3] show that sparse regularization can ob-
in our data set. However, after an inspection of the data, we tain better performance in real estate ranking. Thus, we choose
notice that some of the house price are abnormal. Thus, we to use LASSO (http://statweb.stanford.edu/∼tibs/lasso.html),
preprocess the data by filtering out houses with extremely high which is a l1-constrained regression model, as one of our
or low price compared with their neighborhood. baseline algorithms.
TABLE I shows the overall statistics of our dataset after 2) DeepWalk: Deepwalk [35] is another way of employing
filtering. Overall, the city of San Jose has more houses than random walks for unsupervised feature learning of graphs. The
Rochester on the market (as expected for one of the hottest main approach is inspired by distributed word representation
market in the country). The house prices in the two cities learning. In using DeepWalk, we also use -neighborhood
also have significant differences. Fig. 5 shows some of the graph with the same settings with the graph we built for
example house pictures from the two cities, respectively. From generating sequences for B-LSTM. The learned features are
these pictures, we observe that houses whose prices are above also fed into a LASSO model for learning the regression
average typically have larger yards and better curb appeal, and weights. Indeed, deepwalk can be thought as a simpler version
vice versa. The same can be observed among house interior of our algorithm, where only the graph structure are employed
pictures (examples not shown due to space). to learn features. Our framework can employ both the graph
Realtor does not provide the exact geo-location for each structure and other features, i.e. visual attributes, for building
house. However, geo-location is important for us to build regression model.
the -neighborhood graph for random walks. We employ
Microsoft Bing Map API (https://msdn.microsoft.com/en-us/ 2 We also tried max-pooling. However, the results are not as good as
library/ff701715.aspx) to obtain the latitude and longitude for average-pooling. In the following experiments, we report the results using
each house given its collected address. Fig. 6 shows some of average-pooling.
 6

(a) Rochester (b) San Jose

Fig. 5. Examples of house pictures of the two cities respectively. Top Row: houses whose prices (per Sqft) are above the average of their neighborhood.
Bottom Row: houses whose prices (per Sqft) are below the average of their neighborhood.

TABLE II
P REDICTION DEVIATION OF DIFFERENT MODELS FROM THE ACTUAL SALE PRICES . N OTE THAT RNN- BEST IS THE UPPER - BOUND PERFORMANCE OF THE
RNN BASED MODEL PROPOSED IN THIS WORK .

LASSO DeepWalk RNN-best RNN-avg

City
MAE MAPE MAE MAPE MAE MAPE MAE MAPE
San Jose 70.79 16.92% 68.05 16.12% 17.98 4.58% 66.3 16.11%
Rochester 14.19 24.83% 13.68 23.28% 5.21 9.94% 13.32 22.69%

800

700

600
Number of House Pairs

500

400

300

200

100

(a) San Jose, CA

0
0 1 2 3 4 5 6 7
Distance in Miles

Fig. 7. Distribution of distances between different pairs of houses.

houses only to build our training sequences for Multi-layer

B-LSTM.

For both cities, we build 200, 000 sequences for training,

with a length of 10. Similarly, we also generate testing
(b) Rochester, NY sequences, where each sequence contain one and only one
Fig. 6. Distribution of the houses in our collected data for both San Jose and testing house (see Fig. 4). On the average, we randomly
Rochester according to their geo-locations. generate 100 sequences for each testing house. The B-LSTM
model is trained with a batch size of 1024. In our experimental
settings, we set the size of the first hidden layer to be 400 and
C. Training a Multi-layer B-LSTM Model the size of the second hidden layer to be 200.

With the above mentioned similarity graph, we are able to The evaluation metrics employed are mean absolute error
generate sequences using random walks following the steps (MAE) and mean absolute percentage error (MAPE). Both
described in Algorithm 1. For each city, we randomly split of them are popular measures for evaluating the accuracy of
the houses into training (80%) and testing set (20%). Next, prediction models. Eq.(10) and Eq.(11) give the definitions for
we generate sequences using random walks on the training these two metrics, where pi is the predicted value and ti is
 7

the true value for the i-th instance. D. Confidence Level

N For each testing house, the proposed model can give a group
1 X of predictions. We want to know whether or not the proposed
MAE = |ti − pi | (10)
N i=1 model can distinguish the confidence level of its prediction.
N In particular, we group the testing houses evenly into three
1 X ti − pi
MAPE = | | (11) groups for each city. The first group has the smallest standard
N i=1 ti deviation of the prediction prices. The second group is the
middle one and the last group is the one with the largest
We use the same training and testing split to evaluate all standard deviation.
the approaches. TABLE II shows the regression results for all Fig. 8 shows the MAE and MAPE for the different groups.
the different approaches in the two selected cities. For each The results show that standard deviation can be viewed as a
testing house, we generate about 100 sequences. In TABLE II, rough measure of the confidence level of the proposed model
we report both the best and the average price of the predicted on the current testing house. Small standard deviation tends
price. For Rochester, the average standard deviation of the to indicate a high confidence of the model and overall it also
predicted prices over all the houses is 5.6, which is 7.33% of suggests a smaller prediction error.
the average price in Rochester (see TABLE I). Comparably,
the average standard deviation for San Jose is 34.64, which is V. C ONCLUSION
7.63% of the average price in San Jose. The best is the price
In this work, we propose a novel framework for real estate
closest to the true price among all the available sequences for
appraisal. In particular, the proposed framework is able to take
each house3 . Overall, our B-LSTM model outperforms other
both the location and the visual attributes into consideration.
two baseline algorithms in both cities. All of the evaluation
The evaluation of the proposed model on two selected cities
approaches perform better in San Jose than in Rochester in
suggests the effectiveness and flexibility of the model. Indeed,
terms of MAPE. This is possible due to the availability of
our work has also offered new approaches of applying deep
more training data in the city of San Jose. DeepWalk shows
neural networks on graph structured data. We hope our model
slightly better performance than LASSO, which suggests that
can not only give insights on real estate appraisal, but also can
location is relatively more important than the visual features
inspire others on employing deep neural networks on graph
in the realtor business. This is expected
structured data.
80
R EFERENCES
60 [1] Y. Fu, H. Xiong, Y. Ge, Z. Yao, Y. Zheng, and Z.-H. Zhou, “Exploiting
Price ($)

geographic dependencies for real estate appraisal: a mutual perspective

MAE ROC of ranking and clustering,” in SIGKDD. ACM, 2014, pp. 1047–1056.
40 [2] K. Wardrip, “Public transits impact on housing costs: a review of the
MAE SJ
literature,” 2011.
20 [3] Y. Fu, Y. Ge, Y. Zheng, Z. Yao, Y. Liu, H. Xiong, and N. Yuan,
“Sparse real estate ranking with online user reviews and offline moving
0 behaviors,” p. 120129, 2014.
[4] A. Beja and M. B. Goldman, “On the dynamic behavior of prices in
1 1.5 2 2.5 3 disequilibrium,” The Journal of Finance, vol. 35, no. 2, pp. 235–248,
Group by SD 1980.
[5] E. L’Eplattenier, “How to run a comparative market analysis (cma) the
(a) MAE right way,” http://fitsmallbusiness.com/comparative-market-analysis/,
2016.
40 [6] M. J. Bailey, R. F. Muth, and H. O. Nourse, “A regression method for
MAPE ROC real estate price index construction,” Journal of the American Statistical
Percentage

MAPE SJ Association, vol. 58, no. 304, pp. 933–942, 1963.

30 [7] R. Meese and N. Wallace, “Nonparametric estimation of dynamic
hedonic price models and the construction of residential housing price
indices,” Real Estate Economics, vol. 19, no. 3, pp. 308–332, 1991.
20 [8] S. Sheppard, “Hedonic analysis of housing markets,” Handbook of
regional and urban economics, vol. 3, pp. 1595–1635, 1999.
[9] C. H. Nagaraja, L. D. Brown, L. H. Zhao et al., “An autoregressive
approach to house price modeling,” The Annals of Applied Statistics,
10 vol. 5, no. 1, pp. 124–149, 2011.
1 1.5 2 2.5 3 [10] T. Lasota, Z. Telec, G. Trawiński, and B. Trawiński, “Empirical compar-
Group by SD ison of resampling methods using genetic fuzzy systems for a regression
problem,” in Intelligent Data Engineering and Automated Learning-
(b) MAPE IDEAL 2011. Springer, 2011, pp. 17–24.
[11] O. Kempa, T. Lasota, Z. Telec, and B. Trawiński, “Investigation of
Fig. 8. Performance of B-LSTM-avg in different groups. All the testing bagging ensembles of genetic neural networks and fuzzy systems for
houses are grouped by the predicted standard deviation. real estate appraisal,” in Intelligent Information and Database Systems.
Springer, 2011, pp. 323–332.
[12] W. Di, N. Sundaresan, R. Piramuthu, and A. Bhardwaj, “Is a picture
really worth a thousand words?:-on the role of images in e-commerce,”
3 This is the upper bound of the prediction results. We choose the closest in Proceedings of the 7th ACM international conference on Web search
price using the ground truth price as reference. and data mining. ACM, 2014, pp. 633–642.
 8

[13] X. Jin, A. Gallagher, L. Cao, J. Luo, and J. Han, “The wisdom of social [40] A. Graves, N. Jaitly, and A.-R. Mohamed, “Hybrid speech recognition
multimedia: using flickr for prediction and forecast,” in Proceedings of with deep bidirectional lstm,” in Workshop on Automatic Speech Recog-
the international conference on Multimedia. ACM, 2010, pp. 1235– nition and Understanding (ASRU). IEEE, 2013, pp. 273–278.
1244. [41] M. Schuster and K. K. Paliwal, “Bidirectional recurrent neural net-
[14] Q. You, L. Cao, Y. Cong, X. Zhang, and J. Luo, “A multifaceted ap- works,” Signal Processing, IEEE Transactions on, vol. 45, no. 11, pp.
proach to social multimedia-based prediction of elections,” Multimedia, 2673–2681, 1997.
IEEE Transactions on, vol. 17, no. 12, pp. 2271–2280, Dec 2015. [42] T. Tieleman and G. Hinton, “Lecture 6.5 - rmsprop, coursera: Neural
[15] Q. You, S. Bhatia, and J. Luo, “A picture tells a thousand words?about networks for machine learning,” University of Toronto, Tech. Rep., 2012.
you! user interest profiling from user generated visual content,” Signal [43] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,
Processing, vol. 124, pp. 45–53, 2016. V. Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,”
[16] Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, in CVPR, June 2015.
W. Hubbard, and L. D. Jackel, “Backpropagation applied to handwritten
zip code recognition,” Neural computation, vol. 1, no. 4, pp. 541–551,
1989.
[17] G. E. Hinton, S. Osindero, and Y.-W. Teh, “A fast learning algorithm for
deep belief nets,” Neural computation, vol. 18, no. 7, pp. 1527–1554,
2006.
[18] D. C. Cireşan, U. Meier, J. Masci, L. M. Gambardella, and J. Schmid-
huber, “Flexible, high performance convolutional neural networks for
image classification,” in IJCAI. AAAI Press, 2011, pp. 1237–1242.
[19] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks.” in NIPS, vol. 1, no. 2, 2012,
p. 4.
[20] X. Lu, Z. Lin, H. Jin, J. Yang, and J. Z. Wang, “Rapid: Rating pictorial
aesthetics using deep learning,” in ACM MM. ACM, 2014, pp. 457–466.
[21] B. Zhou, A. Lapedriza, J. Xiao, A. Torralba, and A. Oliva, “Learning
deep features for scene recognition using places database,” in NIPS,
2014, pp. 487–495.
[22] G. Hinton, “A practical guide to training restricted boltzmann machines,”
Momentum, vol. 9, no. 1, p. 926, 2010.
[23] Y. Bengio, “Practical recommendations for gradient-based training of
deep architectures,” in Neural Networks: Tricks of the Trade. Springer,
2012, pp. 437–478.
[24] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning
applied to document recognition,” Proceedings of the IEEE, vol. 86,
no. 11, pp. 2278–2324, 1998.
[25] D. Bahdanau, K. Cho, and Y. Bengio, “Neural machine translation by
jointly learning to align and translate,” ICLR, 2014.
[26] O. Vinyals, A. Toshev, S. Bengio, and D. Erhan, “Show and tell: A
neural image caption generator,” in CVPR, 2015, pp. 3156–3164.
[27] A. Graves, A.-r. Mohamed, and G. Hinton, “Speech recognition with
deep recurrent neural networks,” in ICASSP. IEEE, 2013, pp. 6645–
6649.
[28] F. T. Wang and P. M. Zorn, “Estimating house price growth with repeat
sales data: what’s the aim of the game?” Journal of Housing Economics,
vol. 6, no. 2, pp. 93–118, 1997.
[29] E. Worzala, M. Lenk, and A. Silva, “An exploration of neural networks
and its application to real estate valuation,” Journal of Real Estate
Research, vol. 10, no. 2, pp. 185–201, 1995.
[30] P. Rossini, “Improving the results of artificial neural network models for
residential valuation,” in Fourth Annual Pacific-Rim Real Estate Society
Conference, Perth, Western Australia, 1998.
[31] P. Kershaw and P. Rossini, “Using neural networks to estimate constant
quality house price indices,” Ph.D. dissertation, INTERNATIONAL
REAL ESTATE SOCIETY, 1999.
[32] N. Nghiep and C. Al, “Predicting housing value: A comparison of
multiple regression analysis and artificial neural networks,” Journal of
Real Estate Research, vol. 22, no. 3, pp. 313–336, 2001.
[33] V. Kontrimas and A. Verikas, “The mass appraisal of the real estate by
computational intelligence,” Applied Soft Computing, vol. 11, no. 1, pp.
443–448, 2011.
[34] U. Von Luxburg, “A tutorial on spectral clustering,” Statistics and
computing, vol. 17, no. 4, pp. 395–416, 2007.
[35] B. Perozzi, R. Al-Rfou, and S. Skiena, “Deepwalk: Online learning of
social representations,” in SIGKDD. ACM, 2014, pp. 701–710.
[36] F. Gers, “Long short-term memory in recurrent neural networks,” Un-
published PhD dissertation, École Polytechnique Fédérale de Lausanne,
Lausanne, Switzerland, 2001.
[37] R. Pascanu, T. Mikolov, and Y. Bengio, “On the difficulty of training
recurrent neural networks,” in ICML, 2013, pp. 1310–1318.
[38] F. A. Gers, N. N. Schraudolph, and J. Schmidhuber, “Learning precise
timing with lstm recurrent networks,” The Journal of Machine Learning
Research, vol. 3, pp. 115–143, 2003.
[39] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural
computation, vol. 9, no. 8, pp. 1735–1780, 1997.