Available via license: CC BY 4.0
Content may be subject to copyright.
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2946401, IEEE Access
Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1100/ACCESS.2000.DOI
Analyzing the Leading Causes of Traffic
Fatalities Using XGBoost and
Grid-based Analysis: A City
Management Perspective
JUN MA1,2, YUEXIONG DING2, JACK C.P. CHENG1, YI TAN3, VINCENT J.L. GAN1, and
JINGCHENG ZHANG4
1Department of Civil and Environmental Engineering, The Hong Kong University of Science and Technology, Hong Kong, China
2Department of Research and Development, Big Bay Innovation Research and Development Limited, Hong Kong, China
3College of Civil Engineering, Shenzhen University, Shenzhen, China
4Shenzhen Mariocode Science and Technology Co. Ltd., Shenzhen, China
Corresponding author: JINGCHENG ZHANG (e-mail: zhangjingcheng0306@outlook.com).
ABSTRACT Traffic accidents have been one of the most important global public problems. It has caused
a severe loss of human lives and property every year. Studying the influential factors of accidents can help
find the reasons behind. This can facilitate the design of effective measures and policies to reduce the traffic
fatality rate and improve road safety. However, most of the existing research either adopted methods based
on linear assumption or neglected to further evaluate the spatial relationships. In this paper, we proposed
a methodology framework based on XGBoost and grid analysis to spatially analyze the leading factors on
traffic fatality in Los Angeles County. Characteristics of the collision, time and location, and environmental
factors are considered. Results show that the proposed method has the best modeling performance compared
with other commonly seen machine learning algorithms. Eight factors are found to have the leading impact
on traffic fatality. Spatial relationships between the eight factors and the fatality rates within the Los Angeles
County are further studied using the grid-based analysis in GIS. Specific suggestions on how to reduce the
fatality rate and improve road safety are provided accordingly.
INDEX TERMS XGBoost; Factors analysis; GIS; Grid-based analysis; non-linear machine learning; traffic
fatality.
I. INTRODUCTION
TRAFFIC accidents are considered as one of the most
critical and dangerous problems all around the world.
According to World Health Organization (WHO), about 1.3
million people die each year due to traffic accidents. An
additional 20-50 million are injured or disabled [1]. In the
United States, traffic fatalities increased by 15.3 percent from
2011 to 2016 (29,867 to 37,461) [2]. How to reduce the
occurrence of fatal traffic accidents and improve road safety
has been a significant problem for both governments and
research institutions. To this end, many scholars have con-
ducted different kinds of research to study traffic accidents.
Some focused on the road safety management and education
[3], [4], some paid attention to the improvement of vehicles
[5], [6], some studied the emergency medical service [7], [8],
and some focused on the factors that influence the severity of
traffic accidents [9]–[12].
Knowing what the influential factors are and how they
affect the accidents can help us better understand the cause-
effect behind. This is beneficial to improve the estimation of
the accident severity and the preparation of countermeasures.
Many factors have been studied by previous literature. For
example, Adanu et al. [10] studied the factors that influ-
ence the severity of single-vehicle accidents that happen
on weekdays and weekends. They found that factors such
as driver unemployment, driving with invalid license, no
seatbelt, fatigue, have a high correlation with the severity
of the accidents. Mohamed et al. [11] studied the influential
factors of rear-end crashes. They found that factors like tail-
gating, driving too fast, years of experience, number of lanes,
are significantly affecting the severity of rear-end crashes.
Lee et al. [13] analyzed the impact of rainfall intensity and
VOLUME 4, 2016 1
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2946401, IEEE Access
MA et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
water depth on traffic accident severity. Schneider et al.
[14] examined the statewide motorcycle crash data in a five
year range in Ohio and found that female, the presence of
alcohol and high degree curve would increase the likelihood
of severe traffic accidents. Hashimoto et al. [15] studied the
correlations between traffic accidents and environmental fac-
tors like population and road condition. Aziz, Ukkusuri, and
Hasan [16] concluded that low-lighted-roads and pedestrian
crossing intersections would increase the likelihood of severe
pedestrian-vehicle accidents.
When analyzing the impact factors, these literature had
adopted different kinds of statistical methods. For example,
Russo and Savolainen [17] applied three regression models
namely negative binomial, order logit and multinomial logit
to ascertain the relationship between different median barrier
types and freeway median crash frequency and severity.
Dapilah et al. [18] examined how motorcyclist characteristics
influence the road traffic accidents using chi-square analysis.
Karimi and Kashi [19] investigated the effect of geometric
parameters on accident reduction using sensitivity analysis.
Onozuka et al. [20] studied whether a full moon contributes
to the road traffic accidents using conditional Poisson regres-
sion. However, most of these methods are based on linear
assumptions. Their performances are limited since the real
cases are very complicated, and the relationships between the
factors and accidents are non-linear. Therefore, non-linear
machine learning algorithms were applied in some recent
research to address this problem. For example, Mussone et
al. [21] predicted the severity of the accidents in urban road
intersections using Artificial Neural Network (ANN). Li et al.
[22] used Support Vector Machine (SVM) models for crash
injury severity analysis. However, algorithms like ANN and
SVM are referred to as black box processes because they do
not provide a direct explanation of the variable importance
[23]. This does not help when analyzing the cause-effect
behind traffic accidents. Therefore, a non-linear methodology
framework with the ability to calculate the variable impor-
tance needs to be proposed.
Furthermore, geographical information system (GIS) is
gradually becoming more popular due to its ability to capture,
store, manipulate, analyze, manage and present spatial or
geographical data [24]–[26]. Many previous studies have
used GIS to analyze the data spatially and temporally. For ex-
ample, Kazmi and Zubair [27] estimated the vehicle damage
cost involved in road traffic accidents based on GIS. Din et
al. [28] assessed the level of service of public transportation
using GIS. Zangeneh et al. [29] conducted a spatial-temporal
cluster analysis of mortality from road traffic injuries using
GIS. However, few studies have combined GIS and machine
learning algorithms to analyze the influential factors on traffic
accident fatality.
The objective of this paper is to propose a Grid-based
non-linear machine learning framework to analyze the crit-
ical factors that influence traffic accidents fatality. A non-
linear machine learning algorithm namely XGBoost is imple-
mented to process the data and distinguish the variable with
higher importance. Its performacne is compared with five
other machine learning algorithms, including multiple linear
regression (MLR), logistic regression (LR), multi-layer per-
ceptron (MLP), support vector machine (SVM) and random
forest (RF). Factors including crash characteristics, time and
locations, and environmental characteristics are considered in
the model. Whether a traffic accident involves fatality is the
classification target. Results show that XGBoost outperforms
other algorithms with higher classification accuracy. Eight
factors such as alcohol use and lighting conditions are found
to be the leading causes of fatal accidents based on the
proposed framework. GIS is then used to conduct the spatial
analysis on the relationships between these eight factors and
the fatality rate.
The remaining paper is organized as follows: Section II
describes the proposed methodology framework, Section III
presents a case study in Los Angeles County, Section IV and
Section V show the results and discussion, and Section VI
concludes the work.
II. METHODOLOGY FRAMEWORK
Figure 1 shows the proposed methodology framework. It
consists of three parts. The first part is data preprocessing.
Data collection, cleaning, formatting, and balancing are con-
ducted. The second part is model training and optimization.
XGBoost is applied to build the classification model and
calculate the feature importance. The last part is post-mining.
This is accomplished by conducting the grid-based spatial
analysis in GIS.
A. XGBOOST
The classification modeling using Extreme Gradient Boost-
ing Decision Tree (XGBoost) is an essential component
in the proposed framework. It plays the role of modeling
between the urban features and the outputs. This algorithm
has been reported to achieve good performance in differ-
ent research domains. For example, Zhang and Zhan [30]
adopted XGBoost to classify the rock faces. Torlay et al.
[31] utilized XGBoost to identify atypical language patterns
and differentiate patients with epilepsy. Ma and Cheng [32]
improved the identification accuracy using XGBoost when
modeling features that affect the green building markets.
Zheng et al. [33] applied XGBoost to evaluate the feature
importance in short-term electric load forecasting.
XGBoost is an ensemble technique developed based on the
Gradient Boosting proposed by Friedman [34]. It learns a set
of regression trees (CARTs) in parallel and obtains the result
by summing up the score of each CART. Compared with the
original GBDT algorithm [34], Chen and Guestrin [35] added
some improvements in 2016 and named it as XGBoost. One
of special improvements is the regularized objective to the
loss function. Calculation of the regularized objective Lkfor
the kth iteration is shown in Equation (1).
Lk=
n
X
i=1
l(y(i),ˆy(i)
k) +
k
X
j=1
Ω(fj)(1)
2VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2946401, IEEE Access
MA et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
FIGURE 1: Methodology Framework.
where nrepresents the number of samples, ˆy(i)
krepresents the
prediction of the sample iat iteration k,l(·)represents the
original loss function. Ωis the regularization term, and is
calculated by Equation (2).
Ω(f) = γT +1
2λ
T
X
j=1
ω2
j(2)
where Trepresents the number of leaf nodes. γand λare two
constants used to control the regularization degree.
Another improvement of XGBoost is the utilization of
additive learning strategy [36]. Instead of applying stochastic
gradient descent method to complement the corresponding
optimization procedure, XGBoost adds the best tree model
fk(x(i))into the current classification model to give predic-
tion result for the mth iteration. In this case, Equation (1) can
be further formulated as follows:
Lk=
n
X
i=1
l(y(i),ˆy(i)
k−1+fk(x(i)))+Ω(fk)+
k−1
X
j=1
Ω(fj)(3)
Furthermore, XGBoost adopts the Taylor Expansion sec-
ond order to the objective function, and Equation (3) can be
further expanded to Equation 4.
Lk=
n
X
i=1
[l(y(i),ˆy(i)
k−1) + gi·fk(x(i))
+1
2hi·fk(x(i))] + Ω(fk) + C
(4)
where gi=∂ˆyk−1l(y(i),ˆyk−1)and hi=∂2
ˆyk−1l(y(i),ˆyk−1)
are the first and second order derivatives on the loss function
respectively, Cis a constant.
These advantages make XGBoost potentially to obtain
better results than traditional GBDT. According to the results
of the experiments in our case study, it also outperforms other
commonly-used machine learning algorithms. More details
will be provided later.
B. VARIABLE IMPORTANCE
Except the ability in modeling non-linear classification and
regression problems, XGBoost is also capable of ranking
the variable importance by using the weight. It is given by
the frequency of a feature used in splitting the data across
all CARTs. Importance of a variable v(Wv) based on the
weights can be calculated by Equation 5 and Equation 6.
Wv=
K
X
k=1
L−1
X
l=1
I(Vl
k, v)(5)
I(Vl
k, v) = (1,if Vl
k=v
0,otherwise (6)
where Krepresents the number of trees or iterations, Lrep-
resents the number of leaf nodes of the kth tree, Vl
krepresents
the feature related to the node l, I(·) represents the indicator
function.
C. GRID-BASED ANALYSIS IN GIS
A geographical information system (GIS) is a framework for
gathering, integrating, managing, and analyzing data. It can
analyze the spatial locations and organize different layers of
information into visualizations using maps and 3D scenes.
With this capability, GIS reveals deeper and broader spatial
insights in data.
In this article, GIS is applied to further detect the spatial
relationships between the fatality rate and the influential fac-
tors. However, since we cannot directly calculate the fatality
rate using the traffic data points, the gridding technique in
GIS is then applied to tackle the problem. It divides the
studied region into many fishnets (grids), and then maps the
VOLUME 4, 2016 3
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2946401, IEEE Access
MA et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
traffic data points into these grids. The fatality rate Ri
fof the
ith grid can be calculated using Equation 7.
Ri
f=Ni
f
Ni
a
(7)
where Ni
fmeans the total number of fatal accidents, and Ni
a
means the total number of accidents.
III. CASE STUDY
A. DATA COLLECTION
To validate the proposed framework, we conducted a case
study in Los Angeles County, USA. The dataset was col-
lected from California Statewide Integrated Traffic Records
System (SWITRS). It records every traffic accident reported
in Los Angeles County during 2010-2012, totaling 307,971
accidents. Each record contains the information of the acci-
dent severity, the number of parties involved, the surrounding
environment, and the location. Table 1 presents an overall
summary of the data. The data contains 28 collision features,
18 time and location features, and 7 environment features,
so the data dimension is 53. Besides, since the study focuses
on analyzing the leading causes of traffic fatalities, we set
whether the case involves fatality as the classification target.
Therefore, 3,146 fatal accidents are marked as negative cases
while 304,825 non-fatal accidents are marked as positive
cases.
TABLE 1: Data Summary
Item Description
Negative cases (Fatal accidents) 3,146 cases
Positive cases (Non-fatal accidents) 304,825 cases
Total number of accidents 307,971 cases
Number of collision description variables 28 features
Number of time and location variables 18 features
Number of environment variables 7 features
Total variable dimension 53 dimensions
Time span 2010-2012
Geographic area Los Angeles County
Table 2 shows the detailed description of the data features.
The second column in Table 2 shows the feature abbreviation
and description. The third column presents the data type.
Here C represents categorical features, N represents numeric
features, B represents binary features, and S represents string
features. For example, VIOLCAT in Collision Description
means the Violation Category, and this categorical feature
recorded what kind of traffic violation has the vehicle com-
mitted, including speeding, impeding traffic, traffic signals
etc. POP in the Environment represents the surrounding
population density of the accident spot. For more details,
please refer to the official SWITRS website.
B. DATA PREPROCESSING
Since the raw data always has some flaws, preprocessing is
usually necessary before applying the data into the math-
ematical model. The preprocessing process in this study
TABLE 2: Feature Description
Feature abbreviation
and description
Data
type3
Collision
Description
CHPTYPE:CHP1Beat Type C
VIOLCODE:PCF2Violation Code C
VIOLCAT: PCF Violation Category C
CRASHTYP:Type of Collision C
INVOLVE:Motor Vehicle Involved With C
PED:Pedestrian Action C
CHPFAULT:CHP Vehicle Type is at fault C
SHIFT:CHP effect of new 12-hour shifts C
VIOLSUB:PCF Violation Subsection C
BEATTYPE:Beat Type C
PARTIES:Party Count N
PEDCOL:whether involved a pedestrian B
BICCOL:whether involved a bicycle B
MCCOL:whether involved a motorcycle B
TRUCKCOL:whether involved a big truck B
ETOH:whether involved drinking party B
STFAULT:indicates who is at fault C
HITRUN:Hit And Run C
PCF:Primary Collision Factor C
RIGHTWAY:Control Device C
NOTPRIV:whether on private property B
DIRECT:Direction of the offset distance C
INTERSECT_:whether at an intersection B
KILLED:counts of victims with 1-degree injury N
INJURED: counts of victims with 2,3,4 of injury N
CRASHSEV: the severity level of the collision N
BEATNUMB: Beat Number S
VIOL: PCF Violation S
Time and
Location
YEAR_:Collision Year: C
MONTH_:The month of the year C
DAYWEEK:The Day of Week C
TIMECAT:3-hour categories time C
DATE_: the date when the collision occurred C
TIME_: the time (24 hour time) N
LAPDDIV:City Division LAPD S
STATEHW:whether on a state highway B
POINT_X: The longitude of the geocoded loca-
tion
N
POINT_Y: The latitude of the geocoded location N
JURDIST: Reporting District S
DISTANCE: Offset distance from the secondary
road
N
LOCATION: the location code PRIMARYRD S
JURIS: Jurisdiction S
POSTMILE: markers that indicate the mileage
in California
N
SECONDRD: A secondary reference road S
SPECIAL:Special Condition C
RAMP:Ramp Intersection C
Environment
WEATHER:the weather condition C
WEATHER2:the additional weather condition C
LIGHTING:lighting condition C
POP:Population level C
ROADSURF:Road Surface C
CHPRDTYP:CHP Road Type C
RDCOND1:Road Condition 1 C
1: California Highway Patrol
2: Primary Collision Factor
3: C-categorical, B-binary, N-numeric, S-string
4VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2946401, IEEE Access
MA et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
includes three parts: data formatting, data cleaning, and data
balancing.
1) Data Formatting
The first part is data formatting. The features collected in
this study are mostly categorical but not ordinal. However,
some machine learning algorithms like logistic regression
cannot operate on categorical values directly. They require
the input variables and the output variables to be numeric.
Therefore, these categorical data were converted into dummy
variables in this study. Note that a dummy variable means to
use numeric value 0 or 1 to represent a binary variable. For
example, in the original dataset, categorical data DIRECT
has four independent labels, including north, east, south,
and west. After one hot encoding, four dummy variables,
DIRECT_N, DIRECT_E, DIRECT_S, and DIRECT_W, are
given to indicate the driving direction of the vehicle. In this
way, 29 categorical features were formatted into 359 dummy
variables, and the data dimension is expanded to 383.
2) Data Cleaning
Data cleaning usually refers to the process of excluding
useless data. Two types of data, including noisy data and
high correlational data [37], [38], need to be deleted from the
dataset. Noisy data means the data is irrelevant to our prob-
lem or has limited contribution to the model but increases the
computation complexity. For instance, POINT_X, POINT_Y,
LAPDDIV, and JURIS recorded the location and jurisdiction
information. In this study, they are not the potential causes of
accident fatality. Therefore, they are excluded from the data
mining process.
High correlational data includes two kinds of data. The
first kind is the data that are high correlational to the target.
For example, CRASHSEV has the similar meaning with
our target “fatality”. It will influence the accuracy of our
model. Therefore, it should be deleted. The second kind
of data is those closely related to each other. For example,
TIME means the record of accident time. TIMECAT has
the similar meaning but divides the time into nine intervals.
As TIMECAT is more convenient for further treatment and
analysis, TIME is deleted. Pearson correlation coefficient is
deployed to remove high correlational data. It is a method
based on co-variance and can give information about the
magnitude of the association or correlation. In each pair that
has a correlation coefficient higher than 0.9, we would delete
one feature in order to mitigate the multicollinearity problem.
In summary, 15 noisy features and 18 high correlational
features are excluded from this study. The remained number
of features reduces to 350.
3) Data Balancing
As shown in Table 1, there are 3,146 negative cases and
304,825 positive cases. This means the dataset is very im-
balanced and could make our analysis biased. Therefore,
the dataset needs to be balanced. Two sampling schemes
namely oversampling and undersampling are commonly used
to address this problem [39]. However, both of them have
disadvantages. Oversampling will induce more massive com-
putation and may cause overfitting, while undersampling
may discard potentially useful information. In this study,
we combine both undersampling and parts of oversampling
concept to balance their disadvantages. This is achieved by
conducting the under-sampling ten times and calculate the
average performance. Each time, all the 3,146 negative cases
were selected, and the same amount of positive cases was
sampled and selected without replacement. Combine them to
form a new dataset of 6,292 cases, and then cross-validate
the mathematical model and get the result. Averaging the ten
results gives the final result.
IV. RESULTS
A. MODELING AND OPTIMIZATION
The experiment was run on a computer with 16 GB ram,
Intel (R) Xeon CPUL5640, Windows 7 operating system.
The coding environment is Python 3.6. Parameter optimiza-
tion of XGBoost is firstly conducted. As stated in Section
II, parameter optimization refers to the adjustment of the
algorithm parameters so that the algorithm can better model
the problem. After referring some literature [30], [40]–[42],
the following parameters of XGBoost are optimized in this
section:
•lr: Learning rate. This parameter adjusts the size of the
learning steps. Too small will lead to local optimum and
slow down the calculation, while too large may miss the
optimal value and not converge.
•n_estimators: Number of boosting rounds(or trees). This
parameter represents the number of training iterations
on the data. Too small will lead to under fitting, while
too large will cause overfitting.
•max_depth: Maximum depth of a tree. Increasing this
value will make the model more complex and more
likely to overfit.
•min_child_weight: Minimum sum of instance weight
(hessian) needed in a child.
Note that all the optimization process was conducted using
5-fold cross-validation to obtain stable results. To begin with,
we set lr to {0.01, 0.05, 0.1, 0.15, 0.2} and n_estimators to
{10, 20, 30, ..., 1000}. The optimization result is shown in
Figure 2. It can be seen that the pair {lr, n_estimators}={1.0,
920} offered an optimal value. After this, max_depth and
min_child_weight were then processed to be tuned. As
shown in the Figure 3, when these two parameters were set as
max_depth=10 and min_child_weight=1, the model was able
to obtain an optimal result with 0.8673 in Accuracy. Note that
the optimization criteria used here is Accuracy. Its calculation
is shown in Figure 4.
In addition, to further test the performance of XGBoost,
we compared its performance with five other commonly
seen models: Multiple Linear Regression (MLR), Logistic
Regression (LR), Multi-Layer Perceptron (MLP, one typical
Artificial Neural Network model) , Support Vector Machine
VOLUME 4, 2016 5
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2946401, IEEE Access
MA et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
FIGURE 2: Optimization of lr and n_estimators
FIGURE 3: Optimization of max_depth and
min_child_weight
FIGURE 4: Calculation of Precision, Recall and Accuracy
in the confusion matrix.
(SVM) and Random Forest (RF). 5-fold cross validation is
applied to stabilize the results. Also, two more commonly
used criteria for binary classifications are tested [42]. Their
calculations are shown in Figure 4. Precision reflects the
percentage of corrected predictions among the positively
predicted cases, while recall represents the percentage of
corrected predictions among the real positive cases. All Ac-
curacy, Precision, and Recall are within [0,1], and the higher
values mean better predictions.
The performances of the models are shown in Table 3. As
it can be seen, the indicators of the non-linear algorithms,
MLP, SVM, RF, and XGBoost are significantly higher than
the linear-based algorithms, MLR and LR. XGBoost outper-
forms MLP, SVM and RF with the highest numbers in all the
three indicators. This proves that selecting XGBoost to model
our problem and further analyze the variable importance is a
reasonable choice.
TABLE 3: Comparisons of models
Models MLR LR MLP SVM RF XGBoost
Accuracy 0.7739 0.7854 0.8278 0.8280 0.8463 0.8673
Precision 0.7249 0.7901 0.8335 0.8080 0.8674 0.8758
Recall 0.8490 0.7790 0.8205 0.8510 0.8178 0.8562
B. VARIABLE IMPORTANCE
As introduced in Section II, XGBoost can help calculate
the variable importance. Figure 5 presents the 15 features
that obtained the highest importance. The most influential
factor is ETHO, followed by PARTIES, CRASHTYPE_C,
INVOLVE_C, LIGHTING_A, etc. Explanations of these fea-
tures can be seen in Table 2.
FIGURE 5: Top 15 influential factors
It can be observed from Figure 5 that some features
have similar meanings or belong to the same category of
data. For example, MCCOL and CHPFAULT_02 both mean
whether the accidents involve motorcycles. DAYWEEK_6,
DAYWEEK_5, DAYWEEK_4, and DAYWEEK_7 all refer
to the date category. LIGHTING_A and LIGHTING_C both
refer to the lighting condition. PED_A and PEDCOL both
relate to pedestrian accidents. INVOLVE_C refers to motor-
motor collision, and it could be discussed when analyzing
motor-pedestrian collisions. Therefore, we grouped the 15
features from Figure 5 into eight representative ones for a
more explicit analysis and discussion. Table 4 lists these
6VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2946401, IEEE Access
MA et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
eight features and presents the fatality rates with and without
the relevant feature. More details will be discussed in the
following section.
V. DISCUSSION
Before diving into the eight features, identifying the overall
distribution of the traffic accidents in Los Angeles County
can help us better understand the spatial relationship between
the fatality rate and the eight features. Figure 6 plots the
distribution of the accident density, the fatal accident density,
and the fatality rate. Figure 6 (1) and (2) were plotted using
the kernel density in GIS. It can be seen that most accidents
and fatal accidents happened in urban areas. Figure 6 (4)
was calculated using the grid-based analysis introduced in
Section II-C. It is drawn by firstly cutting the Los Angeles
County map in Figure 6 (3) into 60*60 grids. Then the Rf
of each grid was summarized and calculated. The grids with
Rf= 0 were presented using the green color, while others
are plotted based on their Rfvalues from low to high. It can
be observed that the fatality rates in eight areas are relatively
higher. They are:
•Area A: the city of Lancaster, the city of Palmdale and
their surrounding areas.
•Area B: Interstates 5 Highway between Santa Clarita
Valley and the Pyramid Lake.
•Area C: the southern portion of the California State
Route 14 (SR 14).
•Area D: the middle portion of the California State Route
2 (SR 2).
•Area E: the city of Malibu, the Malibu Creek State Park
and their. surrounding areas.
•Area F: the east of Los Angeles city and its surrounding
areas.
•Area G: city of Pomona, the city of West Covina and
their surrounding areas.
•Area H: interchange of Interstates 405, 5 and 210 High-
ways and its surrounding areas.
It can be observed from Figure 6 that the distribution of
the fatality rate is not highly correlated with the accident
density. This is because the density of accidents mostly relies
on traffic volumes. Higher traffic volumes in urban areas
will naturally lead to more traffic accidents [9]. However,
compared with other places, the higher fatality rates in area
A-H mean that these places are more dangerous, and the
traffic control and safety supervision are not doing well
there. Therefore, one object of this study is to look into the
reasons behind these dangerous areas and discuss possible
measurements to improve the safety performance. According
to the results shown in Table 4, the eight features may reveal
some clues.
A. DRINK-DRIVING
Drink-driving is the most important feature on the fatality
of the accidents according to our results. It can be seen
from Table 4 that when under the influence of alcohol, the
fatality rate is 4.645%, while the number drops to 0.714% for
those without alcohol. Previous studies have demonstrated
that driving under the influence of alcohol is considered an
elevated risk of traffic accidents [18], [43]. The reason behind
is that the response time and the awareness of the drivers
will be impaired under the influence of alcohol. When the
accidents are imminent, intoxicated drivers cannot judge the
dangers or take evasive actions in time. This highly increases
the potential fatality of traffic accidents.
Figure 7 (1) shows the distribution of the percentage of
traffic accidents associated with drink-driving. It indicates
that the numbers are higher in area A, C, G, H and F. This
might be caused by the relatively looser drink-driving control
there. Compared with Figure 6 (4), it can be inferred that
drink-driving is one of the crucial reasons why areas A, C,
G, H, and F have higher fatality rates. Hence, in these areas,
drink-driving enforcement should be strengthened.
B. PARTIES INVOLVEMENT
The second feature is the number of parties involved in
the accidents. It can be seen from Table 4 that when this
number is lower than or equal to 4, the fatality rate is
0.986%. While the number is higher than 4, the fatality rate
increases to 2.708%. More details about the number and the
fatality rate are shown in Figure 8. It can be seen that when
the number is larger than one, the fatality rate goes up as
the number of parties increases. This is reasonable because
multi-vehicle accidents are more likely to be serious traffic
accidents, which have a higher probability to cause fatality.
Previous studies also agreed that the more parties involved in
an accident, the higher the number of people to be involved,
which in turn increases the potential fatality rate [44]. Also,
in a multi-vehicle accident, those who are not wounded at
first and decide to escape at once from their vehicles are still
at risk of being hit by other upcoming vehicles. However, the
cause of multi-vehicle collisions is often hard to determine.
Increasing the road speed limit can be one choice to control
the rate of multi-vehicle accidents.
Another thing can be seen from Figure 8 is that when the
number of parties equals to one, the fatality rate is also high.
Common factors contributing to single-vehicle collisions in-
clude excessive speed, driver fatigue, driving under influence
and age of vehicle [45], [46]. Environmental and roadway
factors can also contribute to single-vehicle accidents. These
include inclement weather, falling rocks, poor lighting condi-
tion, narrow lanes and shoulders, insufficient curve banking
and sharp curves [47], [48]. These can be the reasons why
single-vehicle collisions have a high fatality rate.
C. REAR-END COLLISION
Whether the accident is a rear-end collision is the third factor
that differentiates the fatal and the non-fatal accidents. Rear-
end collisions are considered the most frequently occurring
type of traffic accidents [11]. In our data, there are more
than 108 thousand rear-end crashes, accounting for around
35.3% of the total accidents. It can be observed from Table 4
that if the type of accidents is rear-ended, the fatality rate
VOLUME 4, 2016 7
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2946401, IEEE Access
MA et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
TABLE 4: Fatality rates of the selected eight features.
Feature Description The fatality rate of the accidents
that not involve the feature (%)
The fatality rate of the accidents
that involve the feature (%)
ETOH Drink-driving 0.714 4.645
PARTIES Number of parties in the collision 0.986 (when the number <=4) 2.708 (when the number >4)
CRASHTYP_C Rear-end collision 1.408 0.385
LIGHTING_A Daylight 2.038 0.604
PEDCOL Pedestrian accidents 0.673 4.547
MCCOL Motorcycle accidents 0.912 3.24
DAYWEEK Day of the week 0.904 (weekday) 1.414 (weekend)
TIMECAT Time of the day 0.817 (6:00-24:00) 3.724 (00:00-6:00)
FIGURE 6: The distributions of (1) accident density, (2) fatal accident density, (3) main cities in Los Angeles County, and (4)
fatality rates (grid-based).
is 0.385%. Otherwise, the fatality rate is 1.408%. Fatality
rates of different crash types are shown in Figure 8. It can
also be observed that the rear-end crashes have the lowest
fatality rate while overturned and hit object crashes are the
highest. Overturned means the vehicles flip over its side or
roof and usually happens when the vehicle makes a high-
speed sharp turn. This apparently will lead to a higher fatality
rate. Hit object means leaving the roadway and hitting a fixed
object alongside the road. Such consequences always result
from behaviors like excessive speeds, drowsy driving, drink
driving, and reveal a higher fatality rate [11], [49].
D. LIGHTING CONDITION
Lighting condition has been concluded to be one of the most
critical factors affecting traffic accidents in previous literature
[50], [51]. Poor lighting condition will significantly increase
the likelihood of fatal accidents. In can be seen in Table 4
that the fatality rate during the daytime is only 0.604%, but it
increases to 2.038% under other lighting conditions. Detailed
fatality rates under different lighting conditions are presented
in Figure 8. It can be seen that when there is no street light,
or the street lights are not functioning, the fatality rates are
higher. This is reasonable since the drivers may not see the
road condition clearly and make wrong judgments. Besides,
when the surrounding is dark, drivers are more likely to be
influenced or dazzled by the strong headlights from other
vehicles. This is even more dangerous as the drivers may not
see the road at all.
Figure 7 (2) shows the distribution of the percentage of
accidents happened under poor lighting conditions (no street
lights or the lights are not functioning). It can be observed
that the numbers in area B, area C, area H and the main roads
8VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2946401, IEEE Access
MA et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
FIGURE 7: The grid-based distribution of the percentage of (1) the drink-driving accidents, (2) the accidents under poor
lighting condition (no street lights or the lights are not functioning), (3) the accidents involving pedestrians, (4) the accidents
involving motorcycles, (5) the accidents happened on weekends and (6) the accidents happened during 0am-6am.
around area A are higher than other places. Compared with
Figure 6 (4), it can be inferred that poor lighting conditions
may be one of the reasons of the higher fatality rate in area
B, area C, area H and the surrounding areas of A. Therefore,
it is suggested that the maintenance of the street lights should
be strengthened in these areas.
E. PEDESTRIAN INVOLVEMENT
The fifth factor is the pedestrian involvement. Pedestrians
are referred to as one typical kind of vulnerable road users
(VRUs), which means they are at higher risk in the traf-
fic compared with other road users [52]. This is because
pedestrians do not have sufficient protections when traffic
accidents happen. Also, drinking and traffic-rule violation is
a severe risk not only to motor drivers, but also to pedestrians,
especially when they crossing the roads [55].
It can be observed from Table 4 that for accidents involving
pedestrians, the fatality rate reaches 4.547%, while for acci-
dents without the involvement of pedestrians, the rate is only
0.673%. More kinds of fatality rates of accidents between
motor vehicles and other parties are shown in Figure 8. It
can be seen that the fatality rate of pedestrian accidents is
almost eight times higher than motor-motor accidents (Other
Motor Vehicle). This could be the reason why INVOLVE_C
(motor-motor collisions) stands out in Figure 5 since its
relatively low fatality rate differentiates the fatal and non-
fatal accidents.
Distribution of the percentage of pedestrian accidents is
presented in Figure 7 (3). It indicates that area F, H and
G have a higher percentage of pedestrian accidents. This is
because that in these areas, the population density are higher.
Compared with Figure 6 (4), it can be inferred that pedestrian
accidents might be one of the reasons why area F, H, and G
have higher fatality rates. Therefore, it is suggested that, in
these three areas, traffic operation on pedestrians should be
strengthened, for example, reducing the vehicle speed limit,
installing more raised crosswalks, adding pedestrian lights,
etc. [53]–[55].
F. MOTORCYCLE INVOLVEMENT
Motorcycle accidents are the sixth factor. It can be observed
from Table 4 that for motorcycle accidents, the fatality rate
is 3.240%, but for accidents without the involvement of
motorcycles, the fatality rate decreases to 0.912%. Reasons
VOLUME 4, 2016 9
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2946401, IEEE Access
MA et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
FIGURE 8: Fatality rate under different conditions.
behind the high fatality rate of motorcycle accidents can
be divided into two aspects. On one hand, compared with
automobile drivers, motorcycle drivers are not protected by
seatbelts or airbags. Therefore, when the accidents happen,
they are more likely to be ejected from the seats, and their
heads might be directly bumped without protection form any
cushions. These will increase the likelihood of fatality. On
the other hand, the stability of motorcycles is poor. They are
easily influenced by elements like debris, uneven road, wet
road surface and other small objects [56].
Figure 7 (4) shows the distribution of the percentage of
motorcycle accidents. It can be seen that area C, D and
10 VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2946401, IEEE Access
MA et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
E have higher percentages of motorcycle accidents. These
three areas are located in the mountainous areas of Los
Angeles County. Steep grades and curves of highways built
on mountainous areas pose serious safety threats on passing
vehicles, especially motorcycles [57]. Compared with Figure
6 (4), it can be inferred that motorcycle accidents can be
one of the causes of the high fatality rates in area C, D
and E. Thus, in these areas, traffic control on motorcycles
should be strengthened, speed limits should be lowered,
and more warning signs should be installed to remind the
motorcyclists.
G. DAY OF THE WEEK
Day of the week appears to be an essential factor differ-
entiating the fatal and non-fatal accidents according to our
calculation results. It can be observed from Table 4 that the
fatality rate is 0.904% and 1.414% for accidents in weekdays
and weekends respectively. Fatality rates from Monday to
Sunday within the whole county are presented in Figure 8.
It also indicates that the fatality rates in weekends are higher
than those on weekdays. This might be because that accidents
involving alcohol use, speeding, not using seatbelts, etc.,
are more likely to happen when people are having fun on
weekends [10].
The distribution of the percentage of accidents happened
on weekends is presented in Figure 7 (5). It can be seen that
the percentage of weekend accidents are higher in area B,
D, E, F and H. Compared with Figure 6 (4), it can be inferred
that the problems in weekends may be one of the reasons why
area B, D, E, F, and H have higher rates of fatality. Therefore,
in these areas, traffic control during the weekend should be
strengthened.
H. TIME OF THE DAY
Different times within a day have different fatality rates. It
can be seen from Table 4 that compared with other time
periods, the fatality rate is significantly higher from 0 am to
6 am, reaching 3.724%. Detailed fatality rates of different
time periods within a day are shown in Figure 8. It shows
that the fatality rate reaches the highest during 3 am and 6
am. Although the traffic volume during this period is not as
large as other periods, factors like fatigue driving, speeding,
poor lighting conditions, and loose traffic control all could be
possible reasons for exacerbating the severity of the accidents
after mid-night [58].
Figure 7 (6) presets the distribution of the percentage of
accidents happened from 0 am to 6 am. It can be seen that the
rate is higher in area F and H. Compared with Figure 6 (4),
it can be inferred that one of the reasons of the high fatality
rate in area F and H might be the problems during 0 am to 6
am. Therefore, in these two areas, the traffic control should
be strengthened during this period.
I. SUMMARY
Based on Table 4, Figure 6, Figure 7 and Figure 8, possible
reasons behind the high fatality rate in area A-H are analyzed
and discussed. According to the results revealed by our
methodology, relevant countermeasures to those areas can be
summarized as follows:
•Area A: In this area, drink-driving and poor lighting
condition are two crucial reasons for the high fatality
rate. Therefore, in area A, drink-driving control should
be emphasized, more street lights should be installed,
and regular maintenance should be strengthened.
•Area B: Poor lighting condition is one of the reasons
why area B has a high fatality rate. Thus, the local
government should increase the budget and improve the
lighting condition there. High percentage of weekend
accidents is another possible reason for the high fatality
rate in area B. Therefore, traffic control such as speed
limit should be strengthened in weekends there.
•Area C: Similar to area A, drink-driving and poor
lighting conditions are two principal causes of the high
fatality rate in area C. Furthermore, the high frequency
of motorcycle accidents also worsens the road safety
there. Hence, in area C, besides the drink driving control
and the improvement of lighting infrastructures, proper
managements on motorcycles should also be empha-
sized.
•Area D and E: The high fatality rates in area D and E
may be caused by motorcycle accidents and weekend
accidents. Therefore, in these two areas, the control of
motorcycles should be stricter, and the traffic manage-
ment in weekends need to be enhanced.
•Area F and H: Reasons behind the high fatality rate
in area F and H include drink-driving, pedestrian acci-
dents, weekend accidents as well as the high accident
rate from 0 am to 6 am. In this case, it is suggested
that in these two areas, the management for drink-
driving control should be stricter. Road infrastructures
for pedestrians should be further improved. Traffic con-
trol during weekends and the time from 0 am to 6 am
should be given more attention. Also, area H also reveals
poor lighting conditions, and relevant enhancements
should be delivered.
•Area G: In area G, drink-driving and pedestrian acci-
dents are the main driving factors of the high fatality
rate there. Therefore, in this area, drink-driving control
should be improved, and more road infrastructures for
pedestrians should be installed.
Note that this study is analyzing the influential factors
on fatality rates rather than accident rates. Fatality rates
represent the rates of the fatal accidents over all the accidents,
while accident rates mean the proportion of accidents over
traffic volume. These are two different problems. The former
problem assumes that if an accident happens, what are the
factors that will likely make it fatal. While the latter one
refers to the study of the factors that may lead to an acci-
dent from normal driving. Both problems are important and
typical in accident analysis [21]. Due to data availability, this
study focuses on the grid-based fatality rate. Future work will
VOLUME 4, 2016 11
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2946401, IEEE Access
MA et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
try to integrate traffic volume data and analyze the influential
factors on accident rates.
In addition, driving speed is a potentially important factor
that had been mentioned by some literature [11], [55]. How-
ever, the speed related factor "whether the vehicle is driving
in an unsafe speed (speeding)" derived from the "VIOLCAT"
feature in Table 2 did not rank into top 15 in our experiment
in Los Angeles. This means it is not as important as the top
15 in leading an accident into fatal according to our method-
ology. This is understandable because, on one hand, due to
data availability, this speed related factor may not perfectly
represent the influence of speed on accident fatality. On the
other hand, in many fatal accidents, it is not because that the
drivers violated the traffic rules but the drunk pedestrians
or other vulnerable road users did. A deeper investigation
on the influence of driving speed on fatality rates should be
conducted in the future.
VI. CONCLUSIONS
To conclude, this paper proposed a GIS-based data mining
methodology framework to analyze the influential factors of
fatal traffic accidents. The idea of this study is firstly using
XGBoost to build a binary classification model between fatal
and non-fatal accidents. Based on the identified important
factors using XGBoost models, the grid-based analysis (or
the area analysis) is then used to conduct the spatial analysis
on fatality rates. A case study in Los Angeles County was
conducted to validate the proposed method. Five commonly
seen machine learning algorithms, including MLR, LR, MLP,
SVM, and RF, were applied to model the influential fea-
tures and the traffic fatality. Results showed that XGBoost
obtained the highest modeling accuracy. It was then applied
to investigate the variable importance of the studied features.
Eight factors were found to be the most influential. They were
drink-driving, the number of parties involved, rear-end crash,
lighting condition, pedestrian involvement, motorcycle in-
volvement, the day of the week and time of the day. Through
the grid-based analysis in GIS, their spatial relationships
between the fatality rate were analyzed.
A. CONTRIBUTIONS
The strengths and contributions of the study lie in two
aspects. First is that we proposed an effective methodology
framework in analyzing the influential factors on the fatality
of traffic accidents. The methodology not only provided
accurate modeling performance compared with traditional
linear methods, but also calculated the variable importance
more objectively, and then specifically located the relation-
ships between the factors and the target. Such a process
can systematically analyze the most influential features on
traffic fatalities in a numerical way. Of course, the current
experiments only supported the effectiveness of the method
in analyzing the accidents and fatalities in Los Angeles,
whether the method fits other places and scopes need further
studies to verify.
The second contribution is that the case study conducted
in Los Angeles County uncovered eight areas with higher
traffic fatality rate. According to our numerical analysis,
the higher fatality rate in these areas may result from the
top influential factors discovered in our model. Based on
the spatial analysis, specific practical suggestions on how to
improve road safety in these eight areas are provided. These
can be useful references for the governments during policy-
making.
B. LIMITATIONS AND FUTURE WORK
Still, there are limitations in this paper. Due to the availability
of the data, some possible influential factors, such as traffic
volume, education, and road width, are not considered for
different grids. Also, the authors did not obtain the accident
data in more recent years like 2017-2018 with sufficient
features and factors, this restricted us from analyzing and
comparing the difference and improvements on the traffic
fatalities in Los Angeles. Further studies can be expanded
to consider those factors and data for a more comprehensive
analysis.
REFERENCES
[1] “WHO | 10 facts on global road safety.” [Online]. Available: http:
//www.who.int/features/factfiles/roadsafety/en/
[2] vi.chilukuri.ctr@dot.gov, “USDOT Releases 2016 Fatal Traffic
Crash Data,” Oct. 2017. [Online]. Available: https://www.nhtsa.gov/
press-releases/usdot- releases-2016- fatal-traffic- crash-data
[3] Y. Shen, E. Hermans, Q. Bao, T. Brijs, G. Wets, and W. Wang, “Inter-
national benchmarking of road safety: State of the art,” Transportation
Research Part C: Emerging Technologies, vol. 50, pp. 37–50, Jan. 2015,
doi: 10.1016/j.trc.2014.07.006. [Online]. Available: https://linkinghub.
elsevier.com/retrieve/pii/S0968090X14002083
[4] P. St-Aubin, N. Saunier, and L. Miranda-Moreno, “Large-scale automated
proactive road safety analysis using video data,” Transportation Research
Part C: Emerging Technologies, vol. 58, pp. 363–379, Sep. 2015, doi:
10.1016/j.trc.2015.04.007. [Online]. Available: https://linkinghub.elsevier.
com/retrieve/pii/S0968090X15001485
[5] S. E. Shladover and C. Nowakowski, “Regulatory challenges for
road vehicle automation: Lessons from the California experience,”
Transportation Research Part A: Policy and Practice, Oct. 2017, doi:
10.1016/j.tra.2017.10.006. [Online]. Available: http://www.sciencedirect.
com/science/article/pii/S0965856417300952
[6] G. Ginzburg, S. Evtiukov, I. Brylev, and S. Volkov, “Reconstruction
of Road Accidents Based on Braking Parameters of Category L3
Vehicles,” Transportation Research Procedia, vol. 20, pp. 212–218,
Jan. 2017, doi: 10.1016/j.trpro.2017.01.054. [Online]. Available: http:
//www.sciencedirect.com/science/article/pii/S2352146517300546
[7] C. Castañeda P. and J. G. Villegas, “Analyzing the response
to traffic accidents in Medellín, Colombia, with facility
location models,” IATSS Research, vol. 41, no. 1, pp. 47–
56, Apr. 2017, doi: 10.1016/j.iatssr.2016.09.002. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0386111216300437
[8] Y. Liu, Z. Li, J. Liu, and H. Patel, “A double standard
model for allocating limited emergency medical service vehicle
resources ensuring service reliability,” Transportation Research
Part C: Emerging Technologies, vol. 69, pp. 120–133,
Aug. 2016, doi: 10.1016/j.trc.2016.05.023. [Online]. Available:
http://linkinghub.elsevier.com/retrieve/pii/S0968090X16300602
[9] M. A. Abdel-Aty and A. E. Radwan, “Modeling traffic accident
occurrence and involvement,” Accident Analysis & Prevention, vol. 32,
no. 5, pp. 633–642, Sep. 2000, doi: 10.1016/S0001-4575(99)00094-
9. [Online]. Available: http://www.sciencedirect.com/science/article/pii/
S0001457599000949
[10] E. K. Adanu, A. Hainen, and S. Jones, “Latent class analysis
of factors that influence weekday and weekend single-vehicle crash
12 VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2946401, IEEE Access
MA et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
severities,” Accident Analysis & Prevention, vol. 113, pp. 187–
192, Apr. 2018, doi: 10.1016/j.aap.2018.01.035. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0001457518300411
[11] S. A. Mohamed, K. Mohamed, and H. A. Al-Harthi., “Investigating
Factors Affecting the Occurrence and Severity of Rear-End Crashes,”
Transportation Research Procedia, vol. 25, pp. 2098–2107, Jan.
2017, doi: 10.1016/j.trpro.2017.05.403. [Online]. Available: http://www.
sciencedirect.com/science/article/pii/S235214651730710X
[12] F. Netjasov, D. Crnogorac, and G. Pavlovi´
c, “Potential safety
occurrences as indicators of air traffic management safety
performance: A network based simulation model,” Transportation
Research Part C: Emerging Technologies, vol. 102, pp. 490–
508, May 2019, doi: 10.1016/j.trc.2019.03.026. [Online]. Available:
https://linkinghub.elsevier.com/retrieve/pii/S0968090X1831060X
[13] J. Lee, J. Chae, T. Yoon, and H. Yang, “Traffic accident severity analysis
with rain-related factors using structural equation modeling – A case
study of Seoul City,” Accident Analysis & Prevention, vol. 112, pp.
1–10, Mar. 2018, doi: 10.1016/j.aap.2017.12.013. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0001457517304517
[14] W. H. Schneider and P. T. Savolainen, “Comparison of Severity of
Motorcyclist Injury by Crash Types,” Transportation Research Record,
vol. 2265, no. 1, pp. 70–80, Jan. 2011, doi: 10.3141/2265-08. [Online].
Available: http://journals.sagepub.com/doi/10.3141/2265-08
[15] S. Hashimoto, S. Yoshiki, R. Saeki, Y. Mimura, R. Ando, and
S. Nanba, “Development and application of traffic accident density
estimation models using kernel density estimation,” Journal of Traffic
and Transportation Engineering (English Edition), vol. 3, no. 3, pp.
262–270, Jun. 2016, doi: 10.1016/j.jtte.2016.01.005. [Online]. Available:
https://linkinghub.elsevier.com/retrieve/pii/S2095756415305808
[16] H. A. Aziz, S. V. Ukkusuri, and S. Hasan, “Exploring the
determinants of pedestrian–vehicle crash severity in New York
City,” Accident Analysis & Prevention, vol. 50, pp. 1298–
1309, Jan. 2013, doi: 10.1016/j.aap.2012.09.034. [Online]. Available:
https://linkinghub.elsevier.com/retrieve/pii/S0001457512003533
[17] B. J. Russo and P. T. Savolainen, “A comparison of freeway median
crash frequency, severity, and barrier strike outcomes by median
barrier type,” Accident Analysis & Prevention, vol. 117, pp. 216–
224, Aug. 2018, doi: 10.1016/j.aap.2018.04.023. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0001457518301763
[18] F. Dapilah, B. Y. Guba, and E. Owusu-Sekyere, “Motorcyclist charac-
teristics and traffic behaviour in urban Northern Ghana: Implications
for road traffic accidents,” Journal of Transport & Health, vol. 4, pp.
237–245, Mar. 2017, doi: 10.1016/j.jth.2016.03.001. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S2214140516000232
[19] A. Karimi and E. Kashi, “Investigating the effect of geometric parameters
influencing safety promotion and accident reduction (Case study:
Bojnurd-Golestan National Park road),” Cogent Engineering, vol. 5, no. 1,
p. 1525812, 2018. [Online]. Available: https://www.tandfonline.com/doi/
abs/10.1080/23311916.2018.1525812
[20] D. Onozuka, K. Nishimura, and A. Hagihara, “Full moon and traffic
accident-related emergency ambulance transport: A nationwide case-
crossover study,” Science of The Total Environment, vol. 644, pp. 801–
805, Dec. 2018, doi: 10.1016/j.scitotenv.2018.07.053. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0048969718325361
[21] L. Mussone, M. Bassani, and P. Masci, “Analysis of factors
affecting the severity of crashes in urban road intersections,”
Accident Analysis & Prevention, vol. 103, pp. 112–122,
Jun. 2017, doi: 10.1016/j.aap.2017.04.007. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0001457517301355
[22] Z. Li, P. Liu, W. Wang, and C. Xu, “Using support
vector machine models for crash injury severity analysis,”
Accident Analysis & Prevention, vol. 45, pp. 478–486,
Mar. 2012, doi: 10.1016/j.aap.2011.08.016. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0001457511002363
[23] G. Casalicchio, C. Molnar, and B. Bischl, “Visualizing the Feature
Importance for Black Box Models,” arXiv:1804.06620 [cs, stat], Apr.
2018, arXiv: 1804.06620. [Online]. Available: http://arxiv.org/abs/1804.
06620
[24] “Geographic information system,” Aug. 2018. [On-
line]. Available: https://en.wikipedia.org/w/index.php?title=Geographic_
information_system&oldid=853176113
[25] J. Ma and J. C. Cheng, “Estimation of the building energy use
intensity in the urban scale by integrating GIS and big data technology,”
Applied Energy, vol. 183, pp. 182–192, 2016. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0306261916311679
[26] S. Hong, J. Heo, and A. P. Vonderohe, “Simulation-based approach
for uncertainty assessment: Integrating GPS and GIS,” Transportation
Research Part C: Emerging Technologies, vol. 36, pp. 125–137,
Nov. 2013, doi: 10.1016/j.trc.2013.08.008. [Online]. Available: https:
//linkinghub.elsevier.com/retrieve/pii/S0968090X13001721
[27] J. H. Kazmi and S. Zubair, “Estimation of Vehicle Damage Cost
Involved in Road Traffic Accidents in Karachi, Pakistan: A Geospatial
Perspective,” Procedia Engineering, vol. 77, pp. 70–78, Jan. 2014,
doi: 10.1016/j.proeng.2014.07.008. [Online]. Available: http://www.
sciencedirect.com/science/article/pii/S1877705814009850
[28] M. A. M. Din, S. Paramasivam, N. M. Tarmizi, and A. M. Samad,
“The Use of Geographical Information System in the Assessment of
Level of Service of Transit Systems in Kuala Lumpur,” Procedia
- Social and Behavioral Sciences, vol. 222, pp. 816–826, Jun.
2016, doi: 10.1016/j.sbspro.2016.05.181. [Online]. Available: http:
//www.sciencedirect.com/science/article/pii/S1877042816302567
[29] A. Zangeneh, F. Najafi, S. Karimi, S. Saeidi, and N. Izadi, “Spatial-
temporal cluster analysis of mortality from road traffic injuries using
geographic information systems in West of Iran during 2009–2014,”
Journal of Forensic and Legal Medicine, vol. 55, pp. 15–22, Apr. 2018,
doi: 10.1016/j.jflm.2018.02.009. [Online]. Available: https://linkinghub.
elsevier.com/retrieve/pii/S1752928X18300258
[30] L. Zhang and C. Zhan, “Machine Learning in Rock Facies Classification:
An Application of XGBoost,” in International Geophysical Conference,
Qingdao, China, 17-20 April 2017. Qingdao, China: Society of
Exploration Geophysicists and Chinese Petroleum Society, May 2017,
pp. 1371–1374, doi: 10.1190/IGC2017-351. [Online]. Available: http:
//library.seg.org/doi/10.1190/IGC2017-351
[31] L. Torlay, M. Perrone-Bertolotti, E. Thomas, and M. Baciu, “Machine
learning–XGBoost analysis of language networks to classify patients
with epilepsy,” Brain Inf., vol. 4, no. 3, pp. 159–169, Sep. 2017,
doi: 10.1007/s40708-017-0065-7. [Online]. Available: http://link.springer.
com/10.1007/s40708-017- 0065-7
[32] J. Ma and J. C. P. Cheng, “Identification of the numerical patterns behind
the leading counties in the U.S. local green building markets using
data mining,” Journal of Cleaner Production, vol. 151, pp. 406 – 418,
2017. [Online]. Available: http://www.sciencedirect.com/science/article/
pii/S0959652617305176
[33] H. Zheng, J. Yuan, and L. Chen, “Short-Term Load Forecasting
Using EMD-LSTM Neural Networks with a Xgboost Algorithm
for Feature Importance Evaluation,” Energies, vol. 10, no. 8, p.
1168, Aug. 2017, doi: 10.3390/en10081168. [Online]. Available:
http://www.mdpi.com/1996-1073/10/8/1168
[34] J. H. Friedman, “Stochastic gradient boosting,” Computational Statistics
& Data Analysis, vol. 38, no. 4, pp. 367–378, Feb. 2002, doi:
10.1016/S0167-9473(01)00065-2. [Online]. Available: https://linkinghub.
elsevier.com/retrieve/pii/S0167947301000652
[35] T. Chen and C. Guestrin, “XGBoost: A Scalable Tree Boosting
System,” in Proceedings of the 22nd ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining - KDD ’16.
San Francisco, California, USA: ACM Press, 2016, pp. 785–794, doi:
10.1145/2939672.2939785. [Online]. Available: http://dl.acm.org/citation.
cfm?doid=2939672.2939785
[36] H. Zhang, D. Qiu, R. Wu, Y. Deng, D. Ji, and T. Li, “Novel framework
for image attribute annotation with gene selection XGBoost algorithm
and relative attribute model,” Applied Soft Computing, vol. 80, pp.
57–79, Jul. 2019, doi: 10.1016/j.asoc.2019.03.017. [Online]. Available:
https://linkinghub.elsevier.com/retrieve/pii/S1568494619301358
[37] S. B. Kotsiantis, D. Kanellopoulos, and P. E. Pintelas, “Data Preprocess-
ing for Supervised Leaning,” International Journal of Computer Science,
vol. 1, no. 1, pp. 111–117, Jan. 2006.
[38] J. Benesty, J. Chen, Y. Huang, and I. Cohen, “Pearson Correlation
Coefficient,” in Noise Reduction in Speech Processing, ser. Springer
Topics in Signal Processing, I. Cohen, Y. Huang, J. Chen, and
J. Benesty, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg,
2009, pp. 1–4, doi: 10.1007/978-3-642-00296-0_5. [Online]. Available:
https://doi.org/10.1007/978-3-642-00296- 0_5
[39] A. Y.-c. Liu, “The Effect of Oversampling and Undersampling on Classi-
fying Imbalanced Text Datasets,” 2004.
[40] J. C. Cheng and L. J. Ma, “A non-linear case-based reasoning
approach for retrieval of similar cases and selection of target credits
in LEED projects,” Building and Environment, vol. 93, pp. 349–361,
VOLUME 4, 2016 13
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2946401, IEEE Access
MA et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
2015. [Online]. Available: http://www.sciencedirect.com/science/article/
pii/S0360132315300676
[41] J. Ma and J. C. P. Cheng, “Data-driven study on the achievement
of LEED credits using percentage of average score and association
rule analysis,” vol. 98, pp. 121 – 132, 2016. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0360132316300051
[42] M. A. Jun and J. C. P. Cheng, “Selection of target LEED credits
based on project information and climatic factors using data mining
techniques,” Advanced Engineering Informatics, vol. 32, pp. 224 – 236,
2017. [Online]. Available: http://www.sciencedirect.com/science/article/
pii/S1474034616301124
[43] B. Assemi and M. Hickman, “Relationship between heavy vehicle
periodic inspections, crash contributing factors and crash severity,”
Transportation Research Part A: Policy and Practice, vol. 113, pp.
441–459, Jul. 2018, doi: 10.1016/j.tra.2018.04.018. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S096585641630502X
[44] K. K. Yau, H. Lo, and S. H. Fung, “Multiple-vehicle traffic accidents in
Hong Kong,” Accident Analysis & Prevention, vol. 38, no. 6, pp. 1157–
1161, Nov. 2006, doi: 10.1016/j.aap.2006.05.002. [Online]. Available:
http://linkinghub.elsevier.com/retrieve/pii/S0001457506000807
[45] W. Haddon and V. A. Bradess, “ALCOHOL IN THE SINGLE
VEHICLE FATAL ACCIDENT: EXPERIENCE OF WESTCHESTER
COUNTY, NEW YORK,” JAMA, vol. 169, no. 14, pp. 1587–1593,
Apr. 1959, doi: 10.1001/jama.1959.03000310039009. [Online]. Available:
https://jamanetwork.com/journals/jama/fullarticle/325699
[46] J.-K. Kim, G. F. Ulfarsson, S. Kim, and V. N. Shankar, “Driver-
injury severity in single-vehicle crashes in California: A mixed
logit analysis of heterogeneity due to age and gender,” Accident
Analysis & Prevention, vol. 50, pp. 1073–1081, Jan. 2013, doi:
10.1016/j.aap.2012.08.011. [Online]. Available: http://www.sciencedirect.
com/science/article/pii/S0001457512002990
[47] V. Shankar and F. Mannering, “An exploratory multinomial logit
analysis of single-vehicle motorcycle accident severity,” Journal of Safety
Research, vol. 27, no. 3, pp. 183–194, Sep. 1996, doi: 10.1016/0022-
4375(96)00010-2. [Online]. Available: http://www.sciencedirect.com/
science/article/pii/0022437596000102
[48] K. K. W. Yau, “Risk factors affecting the severity of single vehicle
traffic accidents in Hong Kong,” Accident Analysis & Prevention, vol. 36,
no. 3, pp. 333–340, May 2004, doi: 10.1016/S0001-4575(03)00012-
5. [Online]. Available: http://www.sciencedirect.com/science/article/pii/
S0001457503000125
[49] M. Hiramatsu, “Broadside collision scenarios for different types of
pre-crash driving patterns,” JSAE Review, vol. 24, no. 3, pp. 327–334,
Jul. 2003, doi: 10.1016/S0389-4304(03)00033-X. [Online]. Available:
https://linkinghub.elsevier.com/retrieve/pii/S038943040300033X
[50] M. Jalayer, R. Shabanpour, M. Pour-Rouholamin, N. Golshani, and
H. Zhou, “Wrong-way driving crashes: A random-parameters ordered pro-
bit analysis of injury severity,” Accident Analysis & Prevention, vol. 117,
pp. 128–135, Aug. 2018, doi: 10.1016/j.aap.2018.04.019. [Online]. Avail-
able: https://linkinghub.elsevier.com/retrieve/pii/S0001457518301635
[51] R. O. Mujalli, G. López, and L. Garach, “Bayes classifiers for imbalanced
traffic accidents datasets,” Accident Analysis & Prevention, vol. 88, pp.
37–51, Mar. 2016, doi: 10.1016/j.aap.2015.12.003. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0001457515301548
[52] M. Vilaça, N. Silva, and M. C. Coelho, “Statistical Analysis of
the Occurrence and Severity of Crashes Involving Vulnerable Road
Users,” Transportation Research Procedia, vol. 27, pp. 1113–1120,
Jan. 2017, doi: 10.1016/j.trpro.2017.12.113. [Online]. Available: http:
//www.sciencedirect.com/science/article/pii/S2352146517310104
[53] M. G. Mohamed, N. Saunier, L. F. Miranda-Moreno, and S. V.
Ukkusuri, “A clustering regression approach: A comprehensive injury
severity analysis of pedestrian–vehicle crashes in New York, US and
Montreal, Canada,” Safety Science, vol. 54, pp. 27–37, Apr. 2013,
doi: 10.1016/j.ssci.2012.11.001. [Online]. Available: https://linkinghub.
elsevier.com/retrieve/pii/S0925753512002664
[54] S. Jung, X. Qin, and C. Oh, “Improving strategic policies for
pedestrian safety enhancement using classification tree modeling,”
Transportation Research Part A: Policy and Practice, vol. 85, pp.
53–64, Mar. 2016, doi: 10.1016/j.tra.2016.01.002. [Online]. Available:
https://linkinghub.elsevier.com/retrieve/pii/S0965856416000021
[55] M. Soilán, B. Riveiro, A. Sánchez-Rodríguez, and P. Arias, “Safety
assessment on pedestrian crossing environments using MLS data,”
Accident Analysis & Prevention, vol. 111, pp. 328–337, Feb. 2018,
doi: 10.1016/j.aap.2017.12.009. [Online]. Available: https://linkinghub.
elsevier.com/retrieve/pii/S0001457517304475
[56] Z. Hui, Y. Guang-yu, L. Sheng-xiong, Y. Zhi-yong, W. Zheng-
guo, H. Wei, Y. Yong-min, C. Rong, and L. Gui-e, “Analysis
of 86 fatal motorcycle frontal crashes in Chongqing, China,”
Chinese Journal of Traumatology, vol. 15, no. 3, pp. 170 – 174,
Jun. 2012, doi: https://doi.org/10.3760/cma.j.issn.1008-1275.2012.03.009.
[Online]. Available: http://www.sciencedirect.com/science/article/pii/
S1008127515302984
[57] F. Chen and S. Chen, “Differences in Injury Severity of Accidents
on Mountainous Highways and Non-mountainous Highways,” Procedia
- Social and Behavioral Sciences, vol. 96, pp. 1868–1879, Nov.
2013, doi: 10.1016/j.sbspro.2013.08.212. [Online]. Available: https:
//linkinghub.elsevier.com/retrieve/pii/S1877042813023380
[58] G. Zhang, K. K. Yau, X. Zhang, and Y. Li, “Traffic
accidents involving fatigue driving and their extent of
casualties,” Accident Analysis & Prevention, vol. 87, pp. 34–
42, Feb. 2016, doi: 10.1016/j.aap.2015.10.033. [Online]. Available:
https://linkinghub.elsevier.com/retrieve/pii/S0001457515301159
14 VOLUME 4, 2016
