Analysis of Road Traffic Accidents Severity Using a Pruned Tree-Based Model

People are killed in car accidents every day. As a result, there has been a great deal of research into the factors that determine accident severity in road traffic accidents. Despite the United Nations' goal of reducing the number of road traffic deaths and injuries by 2030, between 20 and 50 million people are injured each year, with 1.3 million dying. Despite having approximately 60% of the world's vehicles, low-and middle-income countries account for 93% of these fatalities. Road traffic accidents have a detrimental impact on most countries, costing them up to 3% of their GDP [1]. Furthermore, more males between the ages of twenty and forty die in traffic accidents each year than females, resulting in financial issues in families and increasing poverty [2].

A cross-disciplinary study is required to discover and analyze the factors affecting and contributing to the severity of road traffic accidents and driver injuries, which are global concerns. Literature has identified driver inattention induced by distraction, overloaded attention, and boring driving as human factors contributing to road traffic accidents [3]. Road traffic accidents are exacerbated by distraction, weather conditions, sleep deprivation, improper lane changes, and nighttime driving [2]. While some studies look at the elements that influence road traffic accidents and their severity, others dig deeper to find the most likely causes of road traffic accidents and what determines their severity.

Numerous research has been undertaken in various countries to offer road accident prediction and analysis methods [4]. Cao et al. [5] present a systematic strategy for identifying severe driving episodes concerning time and place using batch clustering and real-time clustering techniques. Fawcett et al. [6] suggested a Bayesian hierarchical model for projecting accident numbers in future years at places within a pool of probable road safety hotspots. The model assists road safety practitioners in determining the location of potential future hotspots, allowing them to take a proactive rather than a reactive strategy for road safety plan implementation.

Models for forecasting road traffic events were also investigated in the literature. By mining seven months of accident data and 1.6 million users' G.P.S. records, Chen et al. [7] created a deep stack denoise autoencoder model to learn human mobility's hierarchical features to predict traffic accident risks. Lu et al. [8] created a road traffic prediction model called TAP-CNN that takes into account traffic accident-impacting factors such as traffic flow, weather, and light. The samples utilized to test the model's accuracy revealed that the model outperformed the classic neural model in forecasting road traffic. You et al. [9] employed the Matched case-control approach and support vector machines (SVMs) methodologies to assess risk status and estimate the likelihood of an accident using discrete loop detector traffic data and web-crawl weather data. The SVMs classifier with web-crawl weather data improves crash prediction accuracy while reducing false alarm rate when compared to the SVMs classifier without data.

Authors in the studies [10-12] investigated the prediction of the severity of the repercussions incurred by road traffic accident victims. This includes employing latent class clustering, an unsupervised probabilistic clustering approach, to investigate the trends, prevalence, and severity of bike motorist collisions in Denmark [10]. In the study [11], a deep learning-based convolutional neural network was used to extract weights of traffic accident features to improve the accuracy of prediction. Rezaie Moghaddam et al. [12] used artificial neural networks and a deep learning approach to predict road traffic accident severity using the unique traffic accident severity prediction conventional neural network (TASP-CNN) model. Although machine learning techniques and models have been used to predict traffic accidents and their severity, little emphasis has been made on the precision and accuracy of these models' performance. Furthermore, the implication of unknown data in the dataset utilized by these models warrants further investigation because it is commonly assumed that a model performs well based on the output.

This study builds on earlier research to assess the effectiveness of multiple classical machine learning models, addressing previously neglected areas of forecasting factors that influence accident severity. The J48 pruned tree approach is also investigated in the paper as a machine learning model that efficiently deals with unknown data in the dataset. The J48 pruned tree-based classifier model is then used to compare and analyze a dataset of road traffic accidents with seven other cutting-edge classical machine learning models. Nave Bayes, Bagging, KNN, Logistic Model Tree, Decision Tree, Random Tree, and Logistic Regression are among these models. The findings will add to the body of knowledge in this domain by emphasizing the impact of unknown variables in the dataset on the performance of classical machine learning models and how this affects the precision of predicting road traffic accidents and their severity.

This section includes an explanation of the machine learning algorithms, the adapted framework for the study as shown in Figure 1, and evaluation metrics. The experiment was carried out on a machine running Windows 10 with an Intel(R) Core (TM) i7-8650U CPU running at 1.90GHz (8 CPUs), 2.1GHz, 8GB of RAM, and a 500GB hard drive.

1.png

Figure 1. The framework for the road traffic accident using machine learning models

3.1 Dataset

The data set was prepared from manual records of road traffic accidents for the year 2017-2020 https://www.kaggle.com/datasets/saurabhshahane/road-traffic-accidents [32]. Each record contains explanatory variables called features that provide information about the accident. Among other information, the manual records contain the driving experience of the driver, the age of the driver, the service year of the vehicle, road surface conditions, and weather conditions. In the current study, these variables were analyzed to provide information about accident severity (fatal injury, serious injury, and slight injury), which is the target variable. All the sensitive information, such as the driver’s name, has been excluded during data encoding and finally, the data set contains 32 features and 12316 instances of road accidents in an excel format, where each column contains a feature, and each row contains an instance of the road accidents.

3.2 Feature selection and data pre-processing

Table 2. Description of the selected features

A dataset may contain features that do not provide relevant information about the target variable hence, analyzing such features can produce misleading results Features selection is the process of deleting unimportant features from the data set and choosing the important ones. Only 23 of the 32 features in the road accident data set, which are shown in Table 2, were shown to be useful for predicting accident severity. The choice was supported by preparatory research and subject-matter expertise.

Further to feature selection, data pre-processing was applied to increase the quality of the selected features. Data pre-processing includes i) handling missing values. For example, replacing missing numeric values in a column with the mean of the values in the same column. ii) Feature scaling: For example, the mean of each numerical column is subtracted from each value in the column, and the result is divided by the standard deviation of that column, and iii) Removing the outliers: that is, removing values outside of a specific range when compared to other values in the same column. In the current study, 23 of the chosen features have missing values that have been replaced with unknown ones. The missing data were either not accessible at the time of data collection or there was no information available to fill it in.

3.3 Machine learning algorithms

This section provides a brief description of the selected well-known machine learning algorithms which will be compared based on their performance in classifying accident severity. These algorithms are Naïve Bayes, Bagging, K-Nearest Neighbours, Logistic Model Tree, Decision Tree, Random Tree, Logistic Regression, and J48 pruned tree.

Naïve Bayes

Naïve Bayes classifier is one of the well-known machine learning algorithms that perform well in vast areas of application [33]. Naïve Bayes classifier assumes that $k$ classes $C_1, C_2, \ldots C_k$ to be predicted are conditionally independent and each class is represented by an n-dimensional vector $X=x_1, x_2, \ldots, x_n$ of features. This assumption implies that there are no dependencies among features. Thus, the assumption simplifies the Naïve Bayes computations as shown in Eq. (1)

$P(X)=\frac{P(X) P\left(C_i\right)}{P(X)}$      (1)

Furthermore, Nave Bayes is known to demand fewer processing resources and a little amount of training data to estimate the required parameter in the classification process.

Bagging

Bagging is one of the ensembles approaches that combines the performance of several models, trained on randomly sampled data from a given data set. The sampling is done with replacement. The idea is to improve the accuracy of a final model by combining the vote (that is, counting the output) of different models trained on the subset of the data set. In the basic bagging approach, the models are trained in parallel [34]. The process of training each model on a subset of the data set and combining the voting result is called bootstrap aggregating popularly referred to as bagging [35].

K-Nearest Neighbor

K-Nearest Neighbor (KNN) [36] relies on the similarity or distance measure to compute the class of a given instance of a data set. K-nearest neighbor makes use of decision rules that provide a nonparametric means of assigning data instances to a class label, based on the k-classes closely related or close to the given instance.

Logistic Model Tree

The Logistic Model Tree (LMT) is a popular classification model. LMT is a machine learning algorithm that combines the classification algorithms Decision Tree (DT) and Logistic Regression (LR). One such system that learns decision-tree classifiers is C4.5. In practice, LMT produces more accurate results than related algorithms like C4.5 and CART. Unlike decision trees, leaf nodes on LMT feature a logistic regression function of linked attributes in addition to just being class labels. The regression function considers a subset of all data attributes [37].

Decision Tree

The decision tree [38] is a commonly used machine learning algorithm that uses a rule-based tree structure to split data into predefined classes. The decision rules for splitting the data are based on the characteristics and classification of the data set. A decision tree learns these rules, deduced from the data set, and predicts the value or class of the target variable by applying the rules to a given instance of the data set belonging to the class. There are several algorithms for decision tree generation. These include Iterative Dichotomiser 3 (ID3) [39], C4.5 [35], and Classification and Regression Tree (CART) [40] which has been recently applied for classifying accident fatality.

Random Forest

Random forest uses the bagging approach by combining the performance of several decision trees. The idea is to improve the accuracy of the random forest model by combining the accuracy of several decision trees. The mean or median of the outputs of the decision trees is taken as the output of a random forest where the value of the target variable is continuous, while the mode is used where the target variable is a discrete class label [31].

Given that a random forest is made of $M$ decision trees, the output $f_{r f}^M\left(X^I\right)$ of the random forest that takes the $i$th instance of a data set $X=x_1, x_2, \ldots, x_n$ of n-dimensional features is given by:

$\begin{aligned} & F_{r f}^M\left(X^i\right)=y=\left\{\frac{1}{M} \sum_{m=1}^M T_{\text {ree }\left(x^i\right)}\right., \\ & y=\text { continuous value } \max \left(\operatorname{count}\left(T_{\text {ree }\left(x^i\right)}\,\,\right)\right), \\ & y=\text { a discrete class label }\end{aligned}$      (2)

where, $T_{\text {ree }\left(X^i\right)}$ is the output of a decision tree when evaluated on the ith instance of $X$. The function  returns the number of times each class label was predicted by the decision trees while  returns the maximum of a given set of values.

Logistic Regression

Logistic regression is one of the simplest classification approaches which is often used as a baseline model for comparing the performance of classification algorithms. It makes use of the logit function to map the features to the independent variables. Logistic regression is defined by the equation [41]:

$F(z)=\frac{1}{1+e^{-z}}=\frac{1}{1+e^{-\left(b_1\,\,x_1+b_2\,\,x_2+\cdots+b_n\,\,x_n\right)}}$              (3)

where, $z=\hat{y}=b_1 x_1+b_2 x_2+\cdots+b_n x_n$ is the linear regression of the output variable on the features.

J48 Pruned Tree

The J48 pruned tree is a variant of the C4.5 algorithm, it is based on tree pruning and supports two methods of pruning, namely, subtree replacement and subtree raising. In subtree replacements, the nodes in the decision tree are replaced with the leaf nodes [42], while the subtree raising involves selecting and moving a subtree upwards in the tree such that the subtree is closer to the root. J48 uses a greedy search strategy, like the C4.5 algorithm, to build decision trees, and it permits tweaking various parameters to improve classification accuracy [43]. A significant benefit of the chosen model is that missing values in the dataset have little to no effect on how a J48 pruned tree is built.

3.4 Performance evaluation

In the experimental comparison of classifiers, the standard performance measures Accuracy, Precision, Recall, F1-Measure, MCC, Roc Area, PRC Area, processing time, Kappa statistic, and mean absolute error were computed [44-46]. To determine how well a classification method performed, a confusion matrix was also plotted.