ANALYSIS OF ROADWAY FATAL ACCIDENTS USING ENSEMBLE-BASED META-CLASSIFIERS

International Journal of Artificial Intelligence and Applications (IJAIA), Vol.11, No.4, July 2020
DOI: 10.5121/ijaia.2020.11408 101
ANALYSIS OF ROADWAY FATAL ACCIDENTS USING
ENSEMBLE-BASED META-CLASSIFIERS
Waheeda Almayyan
Computer Information Department, Collage of Business Studies, PAAET, Kuwait
ABSTRACT
In the past decades, a lot of effort has been put into roadway traffic safety. With the help of data mining,
the analysis of roadway traffic data is much needed to understand the factors related to fatal accidents.
This paper analyses Fatality Analysis Reporting System (FARS) dataset using several data mining
algorithms. Here, we compare the performance of four meta-classifiers and four data-oriented techniques
known for their ability to handle imbalanced datasets, entirely based on Random Forest classifier. Also,
we study the effect of applying several feature selection algorithms including PSO, Cuckoo, Bat and Tabu
on improving the accuracy and efficiency of classification. The empirical results show that the Threshold
selector meta-classifier combined with over-sampling techniques results were very satisfactory. In this
regard, the proposed technique has gained a mean overall Accuracy of 91% and a Balanced Accuracy that
varies between 96% to 99% using 7-15 features instead of 50 original features.
KEYWORDS
Data mining; roadway traffic safety; Imbalanced Data; Meta-classifiers; Ensemble classifier; Data
Sampling.
1. INTRODUCTION
Nowadays, traffic road accidents are considered a major problem that confronts people health all
around the world. The figures from World Health Organization about global road accident
fatalities showed that fatalities were approximately 1.35 million people annually worldwide [1].
Road traffic injuries are now the leading killer of people aged 5-29 years. The cost of traffic
accidents is estimated to be 3% of Gross Domestic Product worldwide and more than 90% of
road traffic deaths that occurs in low-and middle-income countries. Analysing the severity of
accidents is the key of improving the road safety [2]. Recent roadway safety studies have focused
on identifying the contributing factors that impact accident severity. Nevertheless, several risk
causes are waiting to be discovered or analysed [3].
A number of publications have examined the use of data mining methods in many aspects [4].
Data mining can be defined as a process that uses a variety of data analysis tools to discover
patterns and relationships in data that may be used to make valid predictions. Data mining tools
are based on highly automated search procedures. Basically, it overcomes the weaknesses of
traditional techniques that operate under the assumption that data are distributed normally or
according to another distribution, which can be incorrect and may be difficult to validate. Data
mining methods intense rely on use of computing power. One of its strength points, it is able to
handle categorical variables with a large number of categories, incomplete and noisy datasets [4].
Considerably less researchers considered feature selection compared to the growth achieved in
resampling techniques [5]. Under imbalanced scenarios, minority class samples can easily be

102
discarded as noise. However, such risk can be reduced if the irrelevant features in the feature
space are removed.
The literature review indicates a great interest in adapting data mining algorithms in analysing
road accidents data [8-18]. Our main task is to develop machine learning based intelligent model
that could classify the severity of injuries in FARS dataset more accurately. The Fatal Accidents
Dataset contains all serious accidents that happened on public roads in 2007 reported to the
National Highway Transportation Safety Administration (NHTSA) [6]. The dataset can be
downloaded from California Polytechnic State University and all data originally came from
FARS. The dataset contains 37,248 records and 55 attributes. The data description can be found
in the document FARS Analytic Reference Guide during the years 1975-2007 [7]. The objective
of this study is twofold: firstly, to propose an algorithm for extracting factors that are
significantly related to the car accidents according to their injury severity. Secondly, to
investigate the impact of data-oriented/re-sampling techniques methods on enhancing the
classification performance the imbalanced dataset.
The remaining parts of the paper are organized as follows. Section 2 summarizes the recent
research related work that addresses the problem of roadway traffic safety, followed by sections
3 which include the methodology, the applied meta-classifiers and the applied feature selection
techniques. The evaluation environment, and performance metrics along with the evaluation of
the different models and experimental results are discussed in Section 4. Finally, conclusions are
presented in Section 5.
2. RELATED WORKS
In 1987, individual states in the USA were allowed to raise speed limits on rural freeways from
55 to 65 mph. Ossiander and Cummings analysed the effect of the increased speed limits through
designing an ecological study of crashes and vehicle speeds on Washington State freeways from
1974 through 1994 [8]. They concluded that the incidence of fatal crashes more than doubled
after 1987 and the death rate go up by 27% compared to increase in 10% in the states that did not
increase the speed limit. Solaiman et. al. developed an Internet-based prototype GIS and Road
Accident View System for automated road accident analysis and visualization [9]. The suggested
system had the capabilities to perform query over accident information, trend analysis, statistical
analysis, color-coded mapping and other accident information displayed within the web-based
environment. In addition, it could predict the possible dangerous accident sites from the data
gathered using the map API the system.
Researchers in [10] used partition-based and density-based clustering to analyse the road
accidents data. They first cluster the accident data using K-modes algorithm and then
association-rule mining technique is applied to identify the correlation among various sets of
attributes in which an accident may occur for each cluster. Chang et.al in [11] applied the
classification and regression tree model (CART) to analyse the Taiwan traffic accidents data in
2001. It studied the relationship between fatal injuries and driver/vehicle, highway/environment
and accident variables. The results indicated that the most important variable related to crash
severity is the vehicle type. They identified that pedestrians, motorcycle and bicycle riders have
higher risks of being injured than other types of vehicle drivers in traffic accidents. Krishnaveni
and Hemalatha applied several classification models to predict the injury severities in traffic
accidents that occurred in Hong Kong during 2008 [12]. Naive Bayes, AdaBoostM1, PART
Rule, J48 and Random Forest classifiers were selected for classifying the type of injury severity.
Meanwhile, Genetic Algorithm was applied to reduce the dimensionality of the dataset. The final
results showed that Random Forest outperformed the other four algorithms.

103
Kwon et al. applied several classification algorithms for ranking main factors that cause
accidents [13]. With a binary logistic regression model used as the basis for the comparisons,
they applied the decision tree classifier and Naive Bayes algorithms. Bahiru et al. built a decision
tree based on J48, ID3, CART and Naïve Bayes classifiers to model the severity of injury [14].
The experimental result showed that the accuracy of J48 classifier is higher than others models.
The author believed that closely studying the road traffic accident data with the accident severity
and time components of the accident can help to identify any hidden temporal patterns. Silva and
Saraee proposed novel approach of combining Decision Tree algorithm and Time-Series
Calendar Heatmaps technique to extract the knowledge with time factors from the accident
datasets that happened on a region in North of England [15]. Based on the decision tree models
and evaluation measures, they noticed that there is a correlation between hour and month of the
accident and the severity of the accident.
Researchers in [16], noticed that traffic accidents datasets are usually imbalanced. Therefore,
they investigated under-sampling, oversampling and a mix technique that combines both
techniques. Different Bayes classifiers were used to analyse imbalanced and balanced traffic
crashes datasets in Jordan for three years. The results indicated that the most influencing
parameters were the number of vehicles involved, accident pattern, number of directions,
accident type, lighting, surface condition, and speed limit. Moreover, they noticed that using the
balanced data sets using oversampling technique with Bayesian networks improved classification
results. Li et al. [17] applied Apriori algorithm, Naive Bayes classifier and k-means clustering
algorithms on the FARS Fatal Accident dataset to study the relationship between injury severity
and other attributes such as collision manner, light condition, drunk driver weather conditions.
The analysis result suggested that the human factors like drunk or not and the collision type have
a stronger effect on the fatality rate more than environmental factors like roadway surface,
weather, and light conditions. Pakgohar et al. [18] explored the role of human factors on
incidence and severity of road crashes in Iran by employing descriptive analysis; Logistic
Regression, Classification and Regression Tree. The study indicated the human responsibility on
the occurrence of fatal accidents. The result of the study recommends the important role of
issuing ‘Driving License’ and using ‘Safety Belt’ safety policies which might lower the severity
of injuries in traffic accidents in Iran.
3. METHODOLOGY
First of all, as a pre-processing step we calculated injury_severity variable according to number
of casualties and the number of persons involved in the accident and binned to two categories,
high and low severity. This variable is adapted from [17] and it represents the percentage of
severity in accidents and calculated as
Injury_severity = FATALS / PERSONS (1)
Where FATALS is the number of fatalities and PERSONS is the number of persons involved in
the accident. Injury_severity is referred as “class” in the analysis. And according to this, the
number of samples in the FARS dataset are 25545 instances of minor car accidents and 11703
instances for serious accidents. So, the current imbalance ratio in FARS dataset is 2:1 for
majority and minority classes, respectively. Therefore, we decided to tackle this issue with more
investigation.
Several studies give equal importance to all classes, assuming that datasets are balanced and this
often results in poor classification results [19-23]. The imbalance problem happens when the
numbers of instances of one class outnumber the others. This case is commonly known in the
field of real datasets which need to be usually handled [19,20]. Schierz et al. [21] compared four

104
Cost-sensitive classifiers, namely Naïve Bayes, SVM, Random Forest and C4.5 decision tree to
classify pharmaceutical data. They noticed that SVM and C4.5 decision tree classifiers have
performed relatively well. Other studies discussed adapting data-oriented/resampling techniques
as a solution for the imbalance problem. Data level resampling is one of the many ways to handle
class imbalance problem. In a recent study, Singh proposed a data level resampling method to
improve learning from class imbalanced datasets in health applications [23]. The learning
process on different datasets has improved by incorporating the distribution structure of minority
class samples to generate new data samples using deep learning neural networks.
The primary objective of this research work is aimed at enhancing FARS dataset classification
with exploring minimum number of features which are significantly related to the car accidents.
The primary focus is on investigating the impact of data-oriented/re-sampling techniques
methods on enhancing the classification performance over the imbalanced dataset. In this study
we will evaluate the performance of four meta-classifiers and four data-oriented techniques to
handle the imbalance ratio.
3.1. Meta Classifiers
Numerous classification solutions have been suggested in literature to handle imbalanced
datasets. Several studies investigated the performance of classifiers which should help in
choosing the appropriate classification method. Ensemble learning algorithms which utilizes
ensembles of classifiers such as neural networks, Random Forest, bagging and boosting, have
received an increasing interest for of their ability to deliver an accurate prediction and robust to
noise and outliners than single classifiers [19,24]. The basic idea behind ensembled classifiers is
based upon the premise that a group of classifiers can perform better than an individual classifier.
In this research, we compared the performance of four different meta-classifiers known for their
ability to handle imbalanced datasets, explicitly, Bagging, Cost-sensitive, MetaCost and
Threshold Selection classifiers with Random Forest as base-classifier. Random Forest consists of
a combination of individual base classifiers where each tree is generated using a random vector
sampled independently from the classification input vector to enable a much faster construction
of trees. In 2001, Breiman proposed a promising tree-based ensemble classifier based on a
combination tree of predictors such that each tree depends on the values of a random vector
sampled independently and with the same distribution for all trees and named it Random Forest
[25]. Random Forest is as one of the best learning methods as it is more robust and can achieve
better performances than single decision trees.
3.1.1. Bagging Classifier
Leo Breiman proposed Bagging or Bootstrap aggregating as a meta-algorithm based on ensemble
of decision trees to improve classification and regression models in terms of stability and
classification accuracy in machine learning [25]. Additionally, it decreases the variance and
reduces overfitting. This technique can be used for any type of model such as NN, although it is
most applied to decision tree models [26].
In Bagging diversity is obtained by using bootstrapped replicas of the original training set, where
different training datasets are randomly drawn with replacement. Consequently, with each
training data replica a decision tree is built based on the standard approach. So, each tree can be
defined by a different set of variables, nodes and leaves. Finally, their predictions are combined
to obtain the final result. The final results can be obtained in regression cases by averaging votes
and by combining the outputs of models during classification [27].

105
3.1.2. Cost-Sensitive Classifier
Cost-sensitive classification learning is a type of learning in data mining that considers the
penalty of misclassification cost into consideration [28]. In Cost-sensitive learning process, the
objective is to develop a hypothesis that seek to minimize the high cost errors and the total
misclassification cost. Therefore, a Cost-sensitive classification technique takes the cost matrix
into consideration during model building and generates a model that generate the minimum
expected cost [29].
For two-class problem, let C(Min,Maj) denote the cost of misclassifying a majority class instance
as minority instance and C(Maj,Min) as the cost of misclassifying a minority class instance as
majority instance. When dealing with the class imbalance problem, the cost of misclassifying
minority examples is higher than the cost of misclassifying majority, examples (C(Maj,Min) >
C(Min,Maj)) and there is no penalty for correct classification (i.e., C(Maj,Maj)=C(Min, Min)=
0).
3.1.3. Threshold-Based Selector Classifier
Threshold Selector is a meta-classifier that sets a threshold on the probability output of a base-
classifier. Threshold adjustment for the classifier’s decision is one of the methods used for
dealing with imbalanced datasets [30]. A meta-classifier selects a mid-point threshold on the
probability output by a classifier. The midpoint threshold is set so that a given performance
measure is optimized [31]. By default, the probability threshold is assigned to 0.5, i.e. if an
instance is attributed with a probability of equal or less than 0.5, it is classified as negative for the
respective class, while if it is greater than 0.5, the instance is classified as positive.
Performance is measured either on the training data, a hold-out set or using cross-validation. In
addition, the probabilities returned by the base learner can have their range expanded so that the
output probabilities will reside between 0 and 1, which is useful if the scheme normally produces
probabilities in a very narrow range. For our experiments, the optimal threshold was selected
automatically by the meta-classifier by applying internal five-fold cross validation to optimize
the threshold according to 𝑓-measure (Eq. 13), as measure of a model’s accuracy [32].
3.1.4. MetaCost Classifier
MetaCost procedure is based on relabelling the classes of the training examples, and then
employs a modified training set to produce the final model. MetaCost depends on an internal
Cost-sensitive classifier in order to relabel classes of training examples. Nevertheless the study
by Domingos made no comparison between MetaCost’s final model and the internal Cost-
sensitive classifier on which MetaCost depends [33]. This comparison is worth making as it is
credible that the internal Cost-sensitive classifier may outperform the final model without the
additional computation required to derive the final model. Boosting [34,35] can be effective and
can be better than bagging [36] in minimizing errors as it uses bagging internally. Using a
boosting procedure in MetaCost may improve MetaCost’s performance. This is the reason why
we choose to use boosting procedures in MetaCost in this paper. This meta-classifier makes its
base-classifier Cost-sensitive using the method specified in [33]. This implementation uses all
bagging iterations when reclassifying training data.
3.2. Feature Selection Algorithms
Feature Selection is a challenging machine learning-related task that aims at reducing the number
of features by removing irrelevant, redundant and noisy data while maintaining an acceptable

106
level of classification accuracy. Essentially, the feature space is explored to reduce the feature
space and prepare the conditions for the classification step. The performance of the selected
dimension reduction techniques is examined to find the most effective one.
3.2.1. Tabu Search
Tabu Search is a memory-based metaheuristic algorithm proposed by Glover in 1986 to solve
combinatorial optimization problems [37,38]. Since then, Tabu Search has been successfully
applied in other feature selection problems [39]. Tabu Search is a local neighbourhood search
algorithm that simulates the optimal characteristics of human memory functions. Tabu Search
involves a local search combined with a tabu mechanism.
It starts with an initial feasible solution X’ among the neighbourhood solutions, where  is
the set of feasible solutions, and at each iteration, the algorithm searches the neighbourhood of
the best solution N(X)  to obtain a new one with an improved functional value. A solution X’
 N(X) can be reached from X in two cases, X’ is not included in the Tabu list; and X’ is
included in the Tabu list, but it satisfies the aspiration criterion [40]. Surely, if the new solution
𝑋’ is superior to 𝑋best, the value of 𝑋best is overridden. To avoid cycling, solutions that were
previously visited are declared forbidden or tabu for a certain number of iterations and this surely
improve the performance of the local search. Then, the neighbourhood search is resumed based on
the new feasible solution 𝑋’. This procedure is iteratively executed until the stopping criteria is
met. After the iterative process has terminated, the current best solution so far 𝑋best is considered
the final optimal solution provided by the Tabu Search method [41].
3.2.2. Particle Swarm Optimization Search
The particle swarm optimization (PSO) is a population-based stochastic optimization technique
introduced in 1995 by Kennedy and Eberhart [42]. In PSO, a possible candidate solution is
encoded as a finite-length string called a particle pi in the search space. All of the particles make
use of its own memory and knowledge gained by the swarm as a whole to find the best solution.
With the purpose of discovering the optimal solution, each particle adjusts its searching direction
according to two features, its own best previous experience (pbest) and the best experience of its
companions flying experience (gbest).
Each particle is moving around the n-dimensional search space S with objective function f : S 
n  . Each particle has a position xi ,t fitness function f(xi,t) and ‘‘flies’’ through the problem
space with a velocity vi,t . A new position z1 S is called better than z2 S iff f (z1 )  f (z2 ).
Particles evolve simultaneously based on knowledge shared with neighbouring particles; they
make use of their own memory and knowledge gained by the swarm as a whole to find the best
solution. The best search space position particle i has visited until iteration t is its previous
experience pbest. To each particle, a subset of all particles is assigned as its neighbourhood. The
best previous experience of all neighbours of particle i is called gbest. Each particle additionally
keeps a fraction of its old velocity. The particle updates its velocity and position with the
following equations in continuous PSO [43]:
(2)
(3)

107
3.2.3. Cuckoo Search
Cuckoo search is a recently developed population-based metaheuristic algorithm developed by
Xin-She Yang and Suash Deb in 2009 [44]. Since then it has been used as a successful adaptive
search strategy for solving optimization problems. Recent studies showed that Cuckoo search
algorithm is computationally efficient and easy to implement with less parameters [45].
The basic idea behind the Cuckoo search algorithm is derived from the brood parasitism of some
cuckoo species. These species use the nests of other host birds to lay their eggs in that look like
the pattern and color of the native eggs to reduce the probability of discovering them and rely on
these birds for accommodating their eggs. Sometimes, some of host birds discover and throw the
alien eggs away or simply abandon their nests and build a new one in another place. For the
cuckoo search algorithm, each egg in a nest represents a solution, and a cuckoo egg represents a
new solution. The goal is to employ the new and potentially better solutions (cuckoos) to replace
a not-so-good solution in the nests [45]. Hence, the Cuckoo search algorithm is more efficient in
exploring the search space as it will make sure the algorithm will not fall into a local optimum.
3.2.4. Bat Search
The Bat Algorithm is a meta heuristic Swarm Intelligence algorithm proposed in 2010 by Yang
[46], who was inspired by the abilities of bats in searching for their prey and discriminating
different types of insects and obstacles even at complete darkness. Bats emit loud sound pulses
that help them detect target and avoid obstacles. In order to transpose this behaviour into an
intelligent algorithm, the author states three hypothesis. First all bats will use echolocation to
identify its prey. Secondly, all bats fly randomly and their trajectory is characterized by their
internal encoded frequency (freq), velocity (v) and position in space (x). At each iteration of the
algorithm these three variables are updated as:
freqi = freqmin + (freqmax − freqmin) · β (4)
vit
= vit−1
+ (xit−1
− x_bestj) · freqi (5)
xit
= xit−1
+ vit
(6)
where β ∈ [0,1] is a random vector drawn from a uniform distribution. Moreover, as the bat
attains a position closer to its target then, it will decrease its loudness (Ai) and increase its rate of
the pulse emission (ri) as follows:
Ait+1
= α · Ait
(7)
rit+1
= ri0
· [1 − e−γ·t
] (8)
where α (0 < α < 1) and γ (γ > 0) are constants. Finally, the author assumes that the
loudness will vary from a large value to a minimum one.
4. EXPERIMENTS AND RESULTS
To estimate the generalized error of our method, we have applied 10-fold cross-validation to
avoid overfitting the learned models. For a highly imbalanced datasets, accuracy may be
confusing. Therefore, we considered more appropriate performance measures to compare
different classifiers for their ability to handle imbalanced datasets such as balanced accuracy

108
which considers both sensitivity and specificity. Hence, if two models delivered the same
sensitivity value, the model that demonstrated higher balanced accuracy will be prioritized for
selection. Specificity, Sensitivity, Accuracy, Balanced Accuracy, Precision, 𝑓-measure and MCC
are the quality metrics for assessing the performance of the learning performance of each
classifier before and after data resampling [47]. Their formulae are shown in Equations 9-15.
Where TP: true positives; TN: true negatives; FP: false positives; FN: false negatives.
One of the interesting aspects in data mining is to build computational models with abilities to
extract hidden knowledge using data mining schemes. Regarding model settings, not all
classifiers require parameter optimization. For example, ClassBalancer automatically reassigns
weights to the instances in the dataset such that each class has the same total weight [48,49],
therefore, it does not require adjustment. For Bagging, the only parameter to optimize would be
the number of bags. In our case, the number of bags was adjusted to 100. For Cost-sensitive
Classifier and MetaCost, the cost for misclassification was initially applied in accordance with
the imbalance ratio, which, in case it did not provide a sensitivity of at least 0.5, was further
increased to arrive at the final model. All the meta-learning methods evaluated in this study were
implemented in the WEKA software suite, which is a Java based open source data-mining tool
[48]. The number of trees for Random Forest was arbitrarily set to 100, since it has been shown
that the optimal number of trees is usually between 64 and 128, and increasing the number of
trees does not necessarily improve the model’s performance [27,49].
To determine which data-oriented technique is the most suitable for our FARS dataset and before
executing the feature selection, we start with performing experiments over the class-imbalanced
dataset. The performance values of the meta-classifiers and the suggested sampled dataset are
also illustrated in Table 1 to facilitate the comparison with the investigated methods. Observing
the values of Specificity, Balanced Accuracy and Precision it is clear that the Threshold selector
algorithm has the highest values where it accomplished 94%, 97% and 97%. Whereas, the values
of Sensitivity, Accuracy, 𝑓-measure and MCC showed that Bagging, Cost-sensitive and
MetaCost outperformed the Threshold selector algorithm. We have noticed that Bagging, Cost-
sensitive and MetaCost classifiers recorded a comparable performance in most cases. In general,
all average Balanced Accuracy were compared and provided improvement in the results with
average Accuracies of 90%.

109
In our experiments, three over-sampling techniques, explicitly Synthetic Minority Oversampling
(SMOTE), oversampling minority and Class Balancer together with the Under-sampling method,
were applied. SMOTE technique resamples the dataset by applying the over-sampled the
minority class. While, Over-sampling minority technique achieves oversampling of the minority
class, rather than under-sampling of the majority class, so that both classes have the same
number of instances. Under-sampling technique produces a random subsample of a dataset to
implement under-sampling of the majority class. These over-sampling techniques work through
increasing the number of examples in the minority class to balance the distribution of the data
sets and improve the detection rate of the minority class. Whereas Class Balancer reweights the
instances in the data so that each class has the same total weight. Regarding the data-oriented
techniques results, obviously oversampling the minority class is the best algorithm in all
measures. It accomplished 93%, 99%, 96% 100%, 99% 96% and 89% in terms of Sensitivity,
Specificity, Accuracy, Balanced Accuracy, Precision, 𝑓-measure and MCC. Here we can notice
that data- oriented techniques have helped in improvement of prediction performance.
Table 1. Classification results of FARS Dataset
Model Sensitivity Specificity Accuracy Balanced
Accuracy
Precision 𝑓-
measure
MCC
Bagging 0.891 0.829 0.872 0.914 0.919 0.905 0.698
Cost-sensitive 0.886 0.828 0.868 0.914 0.918 0.902 0.689
MetaCost 0.882 0.848 0.871 0.924 0.926 0.904 0.694
Threshold
selector
0.809 0.943 0.851 0.971 0.968 0.882 0.645
SMOTE 0.868 0.954 0.911 0.977 0.949 0.907 0.790
Oversampling
minority
0.928 0.988 0.958 0.994 0.987 0.957 0.890
Class
Balancer
0.852 0.920 0.886 0.960 0.914 0.882 0.747
Under
sampling
0.894 0.842 0.878 0.921 0.925 0.910 0.711
Feature selection techniques are meant to identify a set of crucial features that have maximum
relevancy for target classes and minimum redundancy with other features in the dataset at the
same time. The next step is to investigate the potential of using feature selection techniques. At
the end of this step a subset of features is chosen for the next round. The optimal features by the
PSO, Cuckoo, Bat and Tabu techniques are listed in Table 2. It is noteworthy that the number of
features has remarkably reduced, compared with original dataset. In this phase we reduced the
size of FARS features from 50 to only 7-15 features.
Table 2. Feature selection results
Feature Selection Algorithm Features No. Features Details
PSO 15 MINUTE VE_TOTAL
PEDS HARM_EV
MAN_COLL
REL_ROAD SP_LIMIT
ALIGNMNT
C_M_ZONE NOT_MIN
HOSP_HR SCH_BUS
CF1 DRUNK_DR
VE_FORMS

110
Cuckoo 13 STATE MINUTE
VE_TOTAL PEDS
HARM_EV MAN_COLL
REL_ROAD
ALIGNMNT
C_M_ZONE HOSP_HR
SCH_BUS DRUNK_DR
VE_FORMS
Bat 14 VE_TOTAL PEDS
HARM_EV MAN_COLL
REL_ROAD
ALIGNMNT
TRA_CONT HOSP_HR
HOSP_MN SCH_BUS
CF1 DRUNK_DR
VE_FORMS WEATHER
Tabu 7 MINUTE HARM_EV
REL_ROAD HOSP_MN
SCH_BUS DRUNK_DR
VE_FORMS
Table 3 shows the classification results of the features selected by PSO. It can be observed that
the classification Balanced Accuracy using PSO technique varies between 91% and 96% with the
selected 15 feature set. In this step Threshold selector has achieved the results of 88%, 93%,
96.5% and 96%, in terms of Sensitivity, Specificity, Balanced Accuracy and Precision. While the
score of other classification models were similar with an average Accuracy of 86%, 89% for 𝑓-
measure and 67% for MCC. Observing the data-oriented techniques, obviously under sampling
the majority class is the best algorithm in most measures. It achieved the results of 92%, 90%,
94%, 93% and 77%, in terms of Sensitivity, Accuracy, Precision 𝑓-measure and MCC.
Oversampling minority technique scored the highest Specificity and Balanced Accuracy scores.
We noticed that the overall performance has not been affected by reducing the number of
features from 50 to 15.
Table 3. Performance comparison of the different learning paradigms after applying PSO algorithm
Model Sensitivit
y
Specificity Accuracy Balanced
Accuracy
Precision 𝑓-
measure
MCC
Bagging 0.880 0.835 0.866 0.917 0.921 0.900 0.683
Cost-sensitive 0.874 0.849 0.866 0.924 0.926 0.899 0.682
MetaCost 0.873 0.852 0.867 0.926 0.928 0.900 0.683
Threshold
selector
0.833 0.931 0.864 0.965 0.963 0.893 0.671
SMOTE 0.859 0.936 0.897 0.967 0.930 0.893 0.766
Oversampling
minority
0.842 0.940 0.891 0.970 0.935 0.886 0.746
Class Balancer 0.851 0.885 0.868 0.942 0.881 0.866 0.724
Under
sampling
0.920 0.869 0.904 0.935 0.939 0.929 0.772
Table 4 presents the classification results of the features selected by Cuckoo search technique. It
can be observed that the classification Balanced Accuracy using Cuckoo technique varies
between 91% and 98% with the selected 13 feature set. In this step Threshold selector model has
achieved the best results of 96%, 98% and 98% in terms of Specificity, Balanced Accuracy and

111
Precision. While the score of other classification models were similar with an average of 88% for
Sensitivity, an Accuracy of 87%, 𝑓-measure of 90% and MCC of 68%. Observing the data-
oriented techniques readings, obviously over sampling the minority class is the best algorithm in
most measures. It achieved the results of 93%, 96%, 94%, 98%, 96%, 94% and 88%, in terms of
Sensitivity, Specificity, Accuracy, Balanced Accuracy, Precision, 𝑓-measure and MCC.
Oversampling minority technique scored the highest Specificity and Balanced Accuracy scores.
We noticed that the overall performance has not been affected by reducing the number of
features too.
Table 4. Performance comparison of the different learning paradigms after applying Cuckoo algorithm
Accuracy
Precision 𝑓-
measure
MCC
Bagging 0.886 0.829 0.868 0.914 0.918 0.902 0.689
Cost-sensitive 0.887 0.830 0.869 0.915 0.919 0.903 0.691
MetaCost 0.882 0.843 0.870 0.921 0.925 0.903 0.691
Threshold
selector
0.790 0.964 0.844 0.982 0.979 0.874 0.631
SMOTE 0.874 0.885 0.879 0.942 0.883 0.878 0.755
Oversampling
minority
0.931 0.961 0.946 0.981 0.960 0.945 0.879
Class Balancer 0.857 0.874 0.866 0.937 0.872 0.865 0.726
Under
sampling
0.922 0.872 0.906 0.936 0.940 0.931 0.777
Next, Table 5 shows the classification results of the 14 features selected by Bat search technique.
We observed that the classification Balanced Accuracy using Bat algorithm varies between 90%
and 93%. Threshold selector, in this step, has gained the results of 86%, 93% and 93% in terms
of Specificity, Balanced Accuracy and Precision. While the score of other classification models
were similar with an average of 87% for Sensitivity, an Accuracy of 85%, 𝑓-measure of 89% and
MCC of 65%. Observing the data-oriented techniques, obviously over sampling the minority
class is the best algorithm in most measures.
As it achieved the results of 91%, 90%, 94%, 77% and 93%, in terms of Sensitivity, Accuracy,
Precision, MCC and 𝑓-measure. Oversampling minority scored the highest Specificity and
Balanced Accuracy scores. Worth mentioning that SMOTE algorithm scored the best scores in
Specificity and Balanced Accuracy. The results obtained suggest that over sampling the minority
class technique has the ability to enhance the predictive accuracy of the Random Forest.
Table 5. Performance comparison of the different learning paradigms after applying Bat algorithm
Accuracy
Precision 𝑓-
measure
MCC
Bagging 0.879 0.819 0.860 0.909 0.914 0.896 0.671
Cost-sensitive 0.878 0.818 0.859 0.909 0.913 0.895 0.668
MetaCost 0.874 0.807 0.853 0.903 0.908 0.891 0.654
Threshold
selector
0.848 0.858 0.851 0.929 0.929 0.887 0.647
SMOTE 0.855 0.934 0.895 0.967 0.928 0.890 0.760
Oversampling
minority
0.917 0.883 0.906 0.942 0.945 0.931 0.776
Class Balancer 0.837 0.900 0.869 0.950 0.893 0.864 0.715
Under sampling 0.903 0.847 0.886 0.924 0.928 0.916 0.730

112
Table 6 shows the classification results of the 7 features selected by Tabu search technique. We
observed that the classification Balanced Accuracy using Tabu technique varies between 91%
and 96%. In this step Threshold selector has obtained the results of 92%, 96% and 96% in terms
of Specificity, Balanced Accuracy and Precision. While the scores of other classification models
were similar with an average of 85% for Accuracy, 91% for Precision, 89% for 𝑓-measure and
66% for MCC. As for the data-oriented techniques, obviously Oversampling the minority class is
the best algorithm in most measures. It achieved the results of 93%, 91%, 94%, 94% and 80%, in
terms of Sensitivity, Accuracy, Precision, MCC and 𝑓-measure. Noteworthy that SMOTE
algorithm scored the best performance at Specificity and Balanced Accuracy with a scores of
91% and 95% respectively.
Table 6. Performance comparison of the different learning paradigms after applying Tabu
Accuracy
Precision 𝑓- measure MCC
Bagging 0.872 0.826 0.858 0.913 0.916 0.894 0.665
Cost-sensitive 0.870 0.844 0.861 0.922 0.924 0.896 0.671
MetaCost 0.869 0.835 0.858 0.918 0.920 0.894 0.664
Threshold
selector
0.817 0.922 0.850 0.961 0.958 0.882 0.643
SMOTE 0.865 0.913 0.890 0.957 0.908 0.886 0.761
Oversampling
minority
0.937 0.871 0.916 0.936 0.941 0.939 0.804
Class Balancer 0.846 0.861 0.853 0.930 0.858 0.852 0.701
Under sampling 0.916 0.829 0.889 0.915 0.921 0.919 0.740
Now, to test the prediction performance of the agreement between the feature selection
techniques a group of experiments were conducted. Figure 1 visualizes the Venn diagram of the
top five relevant features between PSO, Bat, Cuckoo and Tabu which are: MINUTE,
HARM_EV, REL_ROAD, SCH_BUS and DRUNK_DR. Table 7 shows the classification results
of the above mentioned five relevant features. we observed that the classification Balanced
Accuracy using Tabu technique varies between 87% and 92%. In this step Threshold selector has
obtained the results of 85%, 92% and 92% in terms of Specificity, Balanced Accuracy and
Precision. While the scores of other classification models were similar with an average of 81%
for Accuracy, 83% for Sensitivity, 85% for 𝑓-measure and 55% for MCC.
As for the data-oriented techniques, obviously under sampling the majority class is the best
algorithm in most measures. It achieved the results of 87%, 90% and 88%, in terms of
Sensitivity, Precision and 𝑓-measure. Noteworthy that SMOTE algorithm scored the best
performance at Specificity, Accuracy, Balanced Accuracy and MCC with a scores of 91%, 86%,
95% and 70% respectively. It was fascinating to see that in general all different sets of classifiers
have resulted in an Accuracy above 90%.
Table 7. Performance comparison of the different learning paradigms after choosing most relevant features
Accuracy
Precision 𝑓-measure MCC
Bagging 0.837 0.756 0.811 0.878 0.882 0.859 0.562
Cost-sensitive 0.827 0.778 0.812 0.889 0.891 0.857 0.563
MetaCost 0.839 0.739 0.807 0.869 0.875 0.856 0.553
Threshold
selector
0.795 0.849 0.812 0.924 0.920 0.853 0.566
SMOTE 0.818 0.910 0.864 0.955 0.900 0.857 0.697

113
Oversampling
minority
0.821 0.757 0.802 0.879 0.883 0.851 0.539
Class Balancer 0.800 0.864 0.832 0.932 0.855 0.827 0.644
Under sampling 0.870 0.794 0.846 0.897 0.902 0.886 0.639
Figure 1. Relationship between PSO, Cuckoo, Bat and Tabu search algorithms
The last set of experiments confirmed that the suggested technique can efficiently compete with
the best prediction meta-classifiers with least number of features. The suggested classification
model is considered adequate enough for selection as the 10-fold cross-validation provided a
sensitivity value of at least 0.5 and a specificity value not less than 0.5 in all the experiments. It
also indicates that with the assistance of meta-classifiers performance has improved particularly
when handling imbalanced datasets. Moreover, employing Random Forest classifier as a base-
classifier surely improves their prediction accuracy. Among the meta-classifier-based methods,
Threshold selector provided the best performance in many cases. Additionally, Over-sampling
technique results are very satisfactory than other resampling techniques. In this regard, the
proposed technique has gained a mean Overall Accuracy of 91% and a Balanced Accuracy that
varies between 96% to 99% using 7-15 features instead of 50 features (Figure2). Based on the
result analysis, it is suggested that the following factors related to the crash and might affect the
fatality rate are: the minute which the crash occurred, the event that resulted in the most severe
injury, the location of the crash as it relates to its position within or outside the traffic way based
on the “First Harmful the Event”, if a school bus, or motor vehicle functioning as a school bus is
involved, the number of drinking drivers involved in the crash. We agree with previous studies
[17,18] which indicated that human factors and the collision type strongly affect the fatal rate
more than the environmental factors. The results obtained, might help in considering and
evaluating the factors related to fatal accidents.
Figure 2. Performance comparison of the different learning paradigms

114
5. CONCLUSION
In this paper, we presented a computational method that can accurately determine the key
features that influence in classification in FARS dataset. For that reason, we applied and
compared the prediction performances of several meta-classifiers. Moreover, several data-
oriented approaches were applied to handle the uneven class ratios problem. Results from the
study showed that the Threshold Selection Classifier and Over-sampling technique had the better
predictive ability among the other techniques with using Random Forest as a base-classifier.
Experiments on the FARS dataset empirically proved that our proposed method can reduce the
number of features with almost 90% and obtain satisfactory results.
REFERENCES
[1] World Health Organization- WHO, Global status report on road safety, Available from World Wide
Web: https://www.who.int/violence_injury_prevention/road_safety_status/2018/en/ (accessed
December 2019).
[2] Qiu, C., Wang, C., Fang, B.& Zuo, X., (2014). A multiobjective particle swarm optimization-based
partial classification for accident severity analysis. Appl. Artif.Intell.28,555–576.
[3] Kwon, O.H.; Rhee, W.; & Yoon, Y. (2015). Application of classification algorithms for analysis of
road safety risk factor dependencies. Accident Analysis and Prevention, 75, 1-15.
[4] Kolyshkina, I., & Brookes, R. (2002). Data mining approaches to modeling insurance risk. Retrieved
December 14, 2019, from http:// www.salford-systems.com/doc/insurance.pdf
[5] Li, Y., Guo, H., Xiao, L., Yanan, L., & Jinling, L, (2016) Adapted ensemble classification algorithm
based on multiple classifier system and feature selection for classifying multi-class imbalanced data,
Knowledge-Based Systems, 94, 88–104.
[6] Trac Integrated SCM & Project Management. Fatal Accidents Dataset,
https://wiki.csc.calpoly.edu/datasets/wiki/ HighwayAccidents.
[7] U.S. Department of Transportation, FARS analytic reference guide, 1975 to 2009.
https://crashstats.nhtsa.dot.gov/Api/Public/ ViewPublication/811352, 2010.
[8] Eric M Ossiander & Peter Cummings, (2002), Freeway speed limits and traffic fatalities in
washington state. Accident Analysis & Prevention, 34(1):13–18.
[9] KMA Solaiman, Md Mustafizur Rahman & Nashid Shahriar, (2013), Avra Bangladesh collection,
analysis & visualization of road accident data in Bangladesh, In Proceedings of International
Conference on Informatics, Electronics & Vision, 1–6.
[10] Sachin Kumar & Durga Toshniwal, (2015), Analysing road accident data using association rule
mining. In Proceedings of International Conference on Computing, Communication and Security, 1–
6.
[11] Chang L. & H. Wang, (2006), "Analysis of traffic injury severity: An application of non-parametric
classification tree techniques Accident analysis and prevention", Accident analysis and prevention,
Vol. 38(5), pp 1019- 1027.
[12] S. Krishnaveni & M. Hemalatha, (2011), A perspective analysis of traffic accident using data mining
techniques. International Journal of Computer Applications, 23(7):40–48.
[13] Kwon, O.H.; Rhee, W. & Yoon, Y. (2015). Application of classification algorithms for analysis of
road safety risk factor dependencies. Accident Analysis and Prevention, 75, 1-15.
[14] T. K. Bahiru, D. Kumar Singh & E. A. Tessfaw, (2018), "Comparative Study on Data Mining
Classification Algorithms for Predicting Road Traffic Accident Severity," Second International
Conference on Inventive Communication and Computational Technologies (ICICCT), Coimbatore,
2018, 1655-1660.
[15] Silva, HCE & Saraee, MH, (2019), Predicting road traffic accident severity using decision trees and
time-series calendar heatmaps , in: The 6th IEEE Conference on Sustainable Utilization and
Development in Engineering and Technology (2019 IEEE CSUDET), 7 - 9 November 2019, Penang,
Malaysia.
[16] Mujalli, R.O.; López, G. & Garach, L. (2016). Bayes classifiers for imbalanced traffic accidents
datasets. Accident Analysis and Prevention, 88, 37-51.

115
[17] Liling Li, Sharad Shrestha & Gongzhu Hu, (2017), “Analysis of Road Traffic Fatal Accidents Using
Data Mining Techniques”, Software Engineering Research, Management and Applications (SERA),
2017 IEEE 15th International Conference, DOI: 10.1109/SERA.2017.7965753
[18] Pakgohar A., Tabrizi R. S., Khalili M. & Esmaeili A., (2011), The role of human factor in incidence
and severity of road crashes based on the cart and LR regression: a data mining approach. Procedia
Computer Science, World Conference on Information Technology, 3, pp. 764– 769.
[19] Kotsiantis S, Kanellopoulos D & Pintelas P, (2006) Handling imbalanced datasets: a review.
GESTS Int Trans Comput Sci Eng 30(1):25–36
[20] López V, Fernández A, Moreno-Torres JG & Herrera F, (2012), Analysis of preprocessing vs. cost-
sensitive learning for imbalanced classification: open problems on intrinsic data characteristics.
Expert Syst Appl 39:6585–6608. https://doi.org/10.1016/j. eswa.2011.12.043
[21] Schierz AC, (2009), Virtual screening of bioassay data. J Chemin- form 1:21.
https://doi.org/10.1186/1758-2946-1-21
[22] Lin W-J & Chen JJ, (2013), Class-imbalanced classifiers for high-dimensional data. Brief Bioinform
14:13–26. https://doi. org/10.1093/bib/bbs006.
[23] N. D. Singh, (2018), "Clustering and learning from imbalanced data", arXiv preprint
arXiv:1811.00972.
[24] Dietterich, T.G., (2000), An Experimental Comparison of Three Methods for Constructing
Ensembles of Decision Trees: Bagging, Boosting, and Randomization, Machine Learning, 40, 139-
157. http://dx.doi.org/10.1023/A:1007607513941
[25] Breiman L, (2001), Random Forests, Mach Learn 45:5–32
[26] L. Breiman, Bagging predictors, Machine Learning, 24:123–140, 1996. L. Breiman. Some Infinity
Theory for Predictor Ensembles, Technical Report 577, UC Berkeley, 2000. URL
http://www.stat.berkeley.edu/˜breiman.
[27] Oshiro TM, Perez PS & Baranauskas JA, (2012), How many trees in a Random Forest?, Machine
learning and data mining in pattern recognition, Springer, Berlin, pp 154–168.
[28] Sun Y, Kamel MS, Wong AKC & Wang Y., (2007), Cost-sensitive boosting for classification of
imbalanced data. Pattern Recognition, 40: 3358–3378.
[29] Guo Haixiang, Li Yijing, Jennifer Shang, Gu Mingyun, Huang Yuanyue & Gong Bing, (2016),
Learning from class-imbalanced data: Review of methods and applications, Expert Systems with
Applications.
[30] Kotsiantis S, Kanellopoulos D & Pintelas P, (2006), Handling imbalanced datasets: a review, GESTS
Int Trans Comput Sci Eng 30(1):25–36.
[31] Chawla NV, Japkowicz N & Kotcz A, (2004), Editorial: special issue on learning from imbalanced
data sets. SIGKDD Explor Newsl 6:1–6. https://doi.org/10.1145/1007730.1007733.
[32] Powers D, (2011), Evaluation: from precision, recall and f-measure to roc., informedness,
markedness & correlation. J Mach Learn Technol 2:37–63.
[33] Domingos, P., (1999), MetaCost: A general method for making classifiers cost-sensitive, In
Proceedings of the 15th international conference on knowledge discovery and data mining, 155–164.
[34] Quinlan, J.R. (1996), Bagging, boosting, and C4.5, in Proceedings of the 13th National Conference
on Artificial Intelligence, pp. 725-730.
[35] Schapire, R.E., Y. Freund, P. Bartlett, & W.S. Lee (1997), Boosting the margin: A new explanation
for the effectiveness of voting methods, in Proceedings of the Fourteenth International Conference on
Machine Learning, 322-330.
[36] Bauer, E. & Kohavi, R. (1999), An Empirical Comparison of Voting Classification Algorithms:
Bagging, Boosting, and Variants. Machine Learning, Kluwer Academic Publishers, 36, 105-139.
[37] F. Glover, (1989), “Tabu search—part I,” ORSA Journal on Computing, vol. 1, no. 3, 190–206.
[38] F. Glover, (1990), “Tabu search—part II,” ORSA Journal on Computing, vol. 2, no. 1, 4–32,.
[39] Tahir, M.A., Bouridane, A., Kurugollu, F. & Amira, A., (2004), Feature Selection using Tabu Search
for Improving the Classification Rate of Prostate Needle Biopsies, In: Proc. 17th International
Conference on Pattern Recognition (ICPR 2004), Cambridge, UK.
[40] Tahir, M.A., Bouridane, A., Kurugollu, F., 2004b. Simulataneous Feature Selection and Weigthing
for Nearest Neighbor Using Tabu Search. In: Lecture Notes in Computer Science (LNCS 3177), 5th
International Conference on Intelligent Data Engineering and Automated Learning (IDEAL 2004),
Exeter, UK.

116
[41] Korycinski, D., Crawford, M., Barnes, J.W & Ghosh, J., (2003), Adaptive feature selection for
hyperspectral data analysis using a binary hierarchical classifier and Tabu Search, In: Proceedings of
the IEEE International Geoscience and Remote Sensing Symposium, IGARSS.
[42] J. Kennedy &R.C. Eberhart, (2001), “Swarm intelligence”, Morgan Kaufmann.
[43] J. Kennedy & R. Eberhart, (1997), “A discrete binary version of the particle swarm algorithm”,
Proceedings of the IEEE International Conference on Systems, Man, and Cybernetics, Vol.5, 4104–
4108.
[44] Xin-She Yang & Suash Deb, (2009), Cuckoo search via Lévy flights. In Nature & Biologically
Inspired Computing, 2009. NaBIC 2009. World Congress on, 210–214.
[45] Xin-She Yang & Suash Deb, (2014), “Cuckoo search: recent advances and applications,” Neural
Computing and Applications, vol. 24, no. 1, 169–174.
[46] X.-S. Yang, (2010), A new metaheuristic bat-inspired algorithm. In Nature Inspired Cooperative
Strategies for Optimization (NICSO 2010), volume 284 of Studies in Computational Intelligence,
Springer Berlin Heidelberg, 65–74.
[47] Powers, D.M.W., (2011), Evaluation: From Precision, Recall and F-Measure to ROC. Inf. Mark.
Correl. J. Mach. Learn. Technol. 2, 37–63.
[48] Hall M, Frank E, Holmes G et al (2009) The WEKA data mining software: an update. SIGKDD
Explor Newsl 11:10–18. https:// doi.org/10.1145/1656274.1656278
[49] Oshiro TM, Perez PS, Baranauskas JA (2012) How many trees in a Random Forest? In: Machine
learning and data mining in pattern recognition. Springer, Berlin, pp 154–168.

ANALYSIS OF ROADWAY FATAL ACCIDENTS USING ENSEMBLE-BASED META-CLASSIFIERS

More Related Content

What's hot (10)

Similar to ANALYSIS OF ROADWAY FATAL ACCIDENTS USING ENSEMBLE-BASED META-CLASSIFIERS (20)

Recently uploaded (20)

ANALYSIS OF ROADWAY FATAL ACCIDENTS USING ENSEMBLE-BASED META-CLASSIFIERS