Correlation-based feature selection (CFS).

CFS is an attribute selection method developed by Hall and Smith [17], and is a type of filtering model. It ranks attributes based on a correlation-based heuristic evaluation function [18], which evaluates a subset of attributes that are independent of each other but related to the class label. It calculates attribute-class and attribute-attribute correlations based on a good feature subset containing features that are highly related to the class. Irrelevant attributes with a low correlation to the class are ignored; hence, redundant attributes are removed because usually they are closely related to one or more other attributes.

Pearson correlation (PC).

PC [19] is a method used for extensive relationships between attributes, and it evaluates the strength of the relationship between two vectors based on the covariance matrix of the data. It evaluates all attributes related to a class to rank them in an order from high to low and handles the relevant attributes involved in the data exploration process. Attribute selection must first generate and rank attribute subsets. Generating attribute subsets is a search process that is used to compare candidate attribute subsets with the determined ones. If the evaluation result of the new candidate attribute subset is better than that of the previous subset, the new attribute subset is designated the important attribute set. This process is repeated until the termination condition is reached.

Information gain (IG).

IG [20] is widely used in high-dimensional data to measure the effectiveness of attributes in classification. It is the expected amount of information required to reduce entropy and is a filtering method that evaluates the importance of attributes by measuring the IG of related classes [21]. It calculates the information mutual to different attributes for the class, and is a very commonly used univariate method for evaluating attributes. It evaluates the importance of attributes based on the IG of a single attribute at a time and provides the entire information based on this value. In ranking attributes, a higher IG indicates a better discriminative class. It is an effective way of determining the correlations between attributes and classes, and its equation is expressed as follows: (1) where c represents class, a represents attribute, and p(c | a) represents the probability of c when a is known.

Gain ratio (GR).

The GR algorithm is an extension of the IG proposed by Quinlan [20], and it overcomes the shortcomings of IG wherein the attribute selection biases have several different values. GR standardizes IG through the split information (SI) equation, which is expressed as (2) where D denotes the training dataset and D_v is the subset of partitioning D for attribute a. The GR can be obtained using the following formula: (3)

When SI_a(D) approaches zero, the ratio becomes unstable. To avoid this, IG of the selected test must be at least equal to the average gain of all inspection tests. Finally, the attribute with the largest GR is selected as the segmentation attribute.

Decision tree (DT).

DT is a machine learning method developed by Quinlan [20] that exhibits excellent performance in several application fields. It is a nonlinear and non-parametric data-mining tool that can be used for supervised classification and in regression problems. It is easy to use, explains mathematics simply without complicated equations, and visually presents all analysis results in an easy-to-understand hierarchical tree diagram (or if-then) with only a brief description for quick comparison. In machine learning, DT is widely used in traffic safety studies, and it can quickly identify and easily interpret complex patterns related to road crashes [22]. To overcome the shortcomings of statistical methods, DT can be selected as the first method to identify road accidents and allow for an easy understanding of the results. Unlike statistical methods, DT does not require any prior assumptions about or restrictions on the relationship between attributes and it can overcome the multicollinearity problem.

Repeated incremental pruning to produce error reduction (RIPPER).

RIPPER is a rule-based classifier proposed by Cohen [23]. It is a widely used rule induction algorithm that derives a set of rules from the training set. The RIPPER algorithm can be divided into three steps: growth, pruning, and optimization. In the first step, the divide and conquer method is used to add conditions to the rule until it is perfectly classified as a subset of the data. Similar to a DT, the IG criterion is used to identify the next split attribute. In the second step, when the specificity of the rule is increased, entropy is no longer reduced, and the rule is immediately pruned. In the third step, the first and second steps are repeated until the stopping criterion is reached. Various heuristics are used to optimize all the rules. The advantage of RIPPER is that the rules are relatively easy to interpret and suitable for unbalanced problems [24].

Random forest (RF).

The RF algorithm combines the concept of the random subspace method [25] and bootstrap aggregation, which was developed by Breiman [26]. RF is a process of creating several different DTs with different samples at each node, and using the score of each DT as the average of its final scores to obtain more accurate results. For outliers and imbalanced datasets, RF is more robust and scalable than DTs and can overcome nonlinear trends. Furthermore, it reduces bias and overfitting by using multiple trees to reorganize the training data. It is worth noting that RF’s success in providing high-precision prediction is primarily achieved by solving the trade-off between overfitting and prediction accuracy [26]. It can provide high and stable classification performance, the workload of adjusting its parameters is small, and it can be applied to classification and regression cases. Compared with multiple regression and neural networks, RF is highly interpretable and does not require any specific data distribution and normalization of variables in different ranges because it does not require rescaling, transformation, or modification [27].

Extra tree (ET).

ET is a machine learning algorithm that is an extension of the RF algorithm proposed by Geurts et al. [28]. It increases the randomness of the RF algorithm, and its operation is very similar to that of RF. However, the method of building DT in forests is different. There are two main differences between ET and RF. First, RF employs random sampling (bootstrapping) to select the sampling dataset as the training dataset of each DT, whereas ET generally does not employ random sampling; therefore, each DT uses the original training dataset. Second, RF chooses the best split point to partition the DT, whereas ET selects the split point randomly. Once the split point is selected, these two algorithms select the best subset among all the attribute subsets. Therefore, ET adds randomization and still retains optimization. These two differences prevent ET from overfitting and enables it to provide a better performance [28].

Logistic model tree (LMT).

LMT is a classifier proposed by Landwehr et al. [29] that combines DT and linear logistic regression. It uses a LogitBoost algorithm [30] to gradually improve the logistic regression model along the corresponding path from the root to the leaf. In the LMT operation, each intermediate node is split based on the gain ratio, and the LogitBoost algorithm is used to update the node model inherited from the parent node. Furthermore, the LogitBoost algorithm generates a linear regression model for each node in the tree, and the LMT uses the CART algorithm [31] to prune the tree. This learning method can overcome the problems associated with local learning. Furthermore, LMT uses simple logistic variable selection to reduce the number of parameters in the logistic regression and improve the classification performance. It uses cross-validation to find multiple LogitBoost iterations to prevent overfitting of the training data, and the LogitBoost algorithm uses additive logistic regression with least-squares fitting for each class.

Step 1: Data collection.

This data used in this study was obtained in Taoyuan City, which is one of the six largest cities in Taiwan. It has an area of 1,220 square kilometers and population of 2,271,969. Compared with other counties and cities in Taiwan, wherein the average population growth rate is decreasing, Taoyuan City continues to show a growth trend. Owing to the city’s large population, traffic flow is significant. During peak hours or holidays, numerous people and vehicles flow in urban areas or between cities, which causes frequent traffic jams on several important roads in urban areas, highways, and expressways. Compared with other metropolitan cities in Taiwan, Taoyuan City has the highest growth rate in the number of motor vehicle registrations because its population has continued to increase in recent years.

This study obtained public data on road traffic accidents from the Taoyuan City government. The data contained the road environment characteristics and related factors of the accidents that occurred between 2017–2018, and comprised 123,365 records with 109 attributes. The collected road accident cases were carefully recorded by the policies in semi-structured texts and numbers. These attributes included demographics related to drivers (age, gender, driver license, drinking status, degree of injury), road (light, road surface, sight distance, traffic signs), environment (accident location, time, date, weather), and vehicle dimensions (vehicle type, purpose of vehicle use, areas of the vehicle involved in the collision).

Step 2: Data preprocessing.

Data mining is a step in the knowledge discovery in databases process, and its primary objective is to extract high-level knowledge from low-level information. That is, handling a large amount of raw data automatically is necessary to identify the most important and meaningful patterns. Each data-mining process requires a data preprocessing step [35], and its main purposes are to remove irrelevant, inconsistent, and missing values in a dataset, convert the dataset into a uniform format, and sort and induce the dataset for further use. The data preprocessing used in this study is described as follows:

1. Remove irrelevant, inconsistent, and missing values: To delete useless data, irrelevant and inconsistent attributes were removed, and then the attributes with more than half of the missing values were deleted. Furthermore, observations with missing values were also deleted. After removing the attributes and data, the collected data retained 39 attributes (including classes) with 83,875 records.
2. Convert the dataset into a uniform format: In the original data, some attributes were found to be similar; hence, we merged these attributes into a single attribute. Furthermore, we renamed the attributes with inconsistent names and reorganized all columns and rows in a uniform format.
3. Code semi-structured data into a structured format: In this step, we coded some attributes into binary zero-one (no or yes) representations, such as sight distance quality, pavement edge line, drunkenness, and escaping the accident. To simplify attribute values, if a nominal/ordinal attribute had too many linguistic values, we merged some linguistic values into a single linguistic value. For example, traffic roads included national, provincial, county, country, urban, village, dedicated, and agricultural production roads; we merged agricultural production roads with few occurrences into a reasonably similar village road. In the class attribute, there were four accident severity levels: (1) dead within 24 h, (2) dead within 2–30 days, (3) injured, and (4) uninjured. Given that the samples of some accident severity levels were small, a class imbalance could occur. Therefore, we merged accident severity levels (1), (2), and (3) into class Y (injury or death) and represented uninjured as class N (uninjured). The attribute definitions and values of the complete processed data are listed in Table 2.
4. Generate different CF datasets: After the collected data were processed through steps (1), (2), and (3), we wrote a Python program to search for the number of accident occurrences in the same location from the district, road section (street), lane, number, and road intersection of the accident location. Eventually, we generated six datasets with accident frequencies greater than 2, 3, 4, 5, 8, and 10, and they were named CF ≥ 2, CF ≥ 3, CF ≥ 4, CF ≥ 5, CF ≥ 8, and CF ≥ 10, respectively.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Attribute definitions and collected data values.

https://doi.org/10.1371/journal.pone.0272956.t002

Step 3: Three-stage dimension reduction.

Attribute selection is a dimensionality reduction technique that solves the problem of determining the most useful attribute set for a given problem and improves the efficiency of data storage and processing [36]. The following section details the proposed three-stage dimension-reduction method.

• Stage 1: Removing multicollinear attributes. This stage applies the variance inflation factor (VIF) to diagnose the collinearity problem through multiple linear regression [37]. The VIF is the ratio of the variances of the cases when there is multicollinearity between the interpretation variables and when there is no multicollinearity. According to Hair et al. [37], when VIF = 1 is not correlated, 1 < VIF < 5 is moderately correlated and VIF ≥ 5 is highly correlated. Furthermore, if VIF ≥ 4, we need to examine whether there is a multicollinearity problem. Therefore, this stage deletes the independent variable when its VIF value is greater than or equal to 4 (VIF ≥ 4).
• Stage 2: Machine learning algorithm to select attributes. This stage first proposes a simple integrated attribute selection method wherein the key attributes must appear more than three times in the four attribute selection methods; it is called COM_3. This stage uses the CFS, PC, IG, and GR attribute selection methods to identify the important attributes that affect accident severity. We then apply the proposed COM_3 to integrate the important attributes of the four attribute selection methods to obtain the key attributes.
• Stage 3: Factor analysis (FA) to reduce data dimension. FA is a generalization of principal component analysis that reduces numerous attributes to fewer dimensions (factors). The primary objective of FA is to transform the coordinate system in order to minimize the correlation between the system variables [38]. There are several methods to determine the number of dimensions extracted through FA, and the most widely used method is the principal component with eigenvalues greater than one as the dimensions [39]. Furthermore, we used the FA results to delete attributes with low loadings. As a rule of thumb, we recommend interpreting only factor loadings with an absolute value greater than 0.5, which explains approximately 25% of the variance [37].
• Step 4: Classification and rule generation. To explore the relationship between the attributes and injury degree, this step uses five RBML classifiers to classify six different accident frequency datasets and generate their rule sets. The five RBML classifiers are DT, RIPPER, RF, ET, and LMT, and they were selected for the following reasons: (1) DT is a commonly used baseline method, and DT and RIPPER are easy to implement. (2) RF and ET are ensemble classifiers, and ET is an extension of the RF. (3) LMT combines DT and linear logistic regression. In the experiment, each dataset was implemented using 10-fold cross-validation, and the average of 100 repeated implementations was calculated to present their results for comparing their performance. Finally, the rule sets were generated in an IF-THEN form, and a DT diagram was used to easily understand the key attributes and study results.
• Step 5: Evaluation. Evaluation is a standard method for measuring the effectiveness of an RBML classifier, which uses accuracy, sensitivity, specificity, AUC, and F1 metrics to evaluate the performance of the five RBML classifiers. This study used AUC because the receiver operating characteristic curve (ROC) has diagnostic power in imbalanced classes, and it can handle both the positive detection and false alarm rates. Furthermore, AUC is a measure of the discriminative strength between these two rates without considering misclassification costs or class prior probabilities [40]. F1 (F-score or F-measure) is a widely used measurement standard in information retrieval and class imbalance problems, and it is the harmonic mean between precision and sensitivity [41]. The classification results were calculated using the confusion matrix [42], which has two dimensions. One is the actual class of the object and the other is the class predicted by the classifier. The accuracy, sensitivity, specificity, and F1 metrics are based on the confusion matrix, and their equations are expressed as follows [41]: (4) (5) (6) (7)
• where TP, TN, FP, and FN represent true positive, true negative, false positive, and false negative, respectively.

4.2.1 Data visualization.

To visualize the data, we used the Geocoding API of Google Maps to obtain the latitude and longitude based on the location (address) of the road accidents and then applied the Choropleth and HeatMap map visualizations in the Python tool to generate visual maps. The dataset shows that the Taoyuan, Guishan, and Luzhu districts had more traffic accidents than other districts. The Fuxing district is a mountainous area of aboriginal people with few accident accidents. We also screened the top ten traffic accident locations in Table 5.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Top 10 traffic accident locations in Touyuan City.

https://doi.org/10.1371/journal.pone.0272956.t005

4.2.2 Finding key attributes through three-stage dimension reduction.

Through Step 3 of the proposed computational steps, the experiment used the proposed three-stage dimension reduction to sequentially determine the key attributes. Here, we illustrate only the CF ≥10 dataset to determine the key attributes because this dataset has important management implications for exploring the key attributes of road accidents.

• The VIF is used to delete collinear attributes. The dataset contained 38 key attributes and one class after COM_3 integrated attribute selection. Next, we used multiple linear regression [37] to generate the VIF for diagnosing the collinearity problem, and used a VIF value greater than or equal to four [37] to delete the collinear attributes, as illustrated in Table 6. We observed that VIF ≥ 4 for four attributes (Sign_ty, Sign_st, Pro_eq, and E_use). Thereafter, we deleted these four attributes before the next stage of the analysis.
• COM_3 was used to integrate the important attributes of the four attribute selection methods. After the FA deleted four attributes, the dataset retained 34 conditional attributes and one class attribute. This study applied CFS, PC, IG, and GR to identify the important attributes, and then used the proposed COM_3 to integrate these important attributes from the four attribute selection methods. We only list the results of attributes selected by at least three attribute selection methods in Table 7, and these 11 attributes are the key attributes.
• Factor analysis (FA) was applied to reduce the attributes. FA was used to obtain the five principal components with eigenvalues greater than or equal to one (λ ≥ 1, the total variance explained 72.48%), as listed in Table 8, and deleted the age attributes with an absolute value of loading less than 0.5 [37], as listed in Table 8. Therefore, we renamed the five principal components to vehicle-related (vehicle), experience and skill, work-related (work), avoid responsibility, and gender dimensions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Results of the collinearity test using VIF of multiple linear regression.

https://doi.org/10.1371/journal.pone.0272956.t006

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. Results of the four attribute selection methods and COM_3 in the CF ≥ 10 dataset.

https://doi.org/10.1371/journal.pone.0272956.t007

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 8. Results of factor analysis in CF ≥ 10 dataset.

https://doi.org/10.1371/journal.pone.0272956.t008

4.2.3 Classification (generating rules) and comparison.

From the proposed computational Steps 4 and 5 in Section 3, we used the five RBML classifiers (DT, RIPPER, RF, ET, and LMT) to classify the injury severity and generate rules, and then compared their performance based on five metrics. The experiment applied tenfold cross-validation and the average of 100 repeat implementations to obtain their results. From Table 3, it is evident that the seven datasets have a slight class imbalance problem; hence, we used the accuracy, sensitivity, specificity, AUC, and F1 metrics to evaluate the performance. The following describes the performance of the five RBML classifiers in the datasets of different accident frequencies, performance of the three-stage dimension reduction in the most valuable complete datasets CR ≥ 1 and CR ≥ 10, and the generation of DT rules.

• Comparing different accident frequency datasets. The results of six datasets with different accident frequencies are listed in Table 9. CR ≥ 3 and CR ≥ 4 had the same records and metric results; therefore, we only list the results of CR ≥ 4. The results indicate that LMT is the best RBML classifier for the five metrics, except for the sensitivity of the CR ≥ 5 dataset. Furthermore, the best accuracy obtained was 91.96% in the CR ≥ 4 dataset, and the highest AUC of the six datasets was 0.93, except for the CR ≥ 10 dataset. The best sensitivity was 0.97 for the complete dataset (CR ≥ 1), and the highest specificity of the six datasets was 0.97, except for the CR ≥ 1 dataset. The highest F1 in the complete dataset (CR ≥ 1) was 0.93. In summary, LMT is the best RBML classifier for the six different accident frequency datasets.
• Comparing different stage dimension reductions. This study illustrates the performance of each stage dimension reduction in the most valuable complete datasets CR ≥ 1 and CR ≥ 10; the results are listed in Table 10. In the CR ≥ 1 dataset, the best performance was that of the LMT classifiers in the five metrics. Furthermore, we observed the following: (i) after removing collinear attributes, the accuracy of the C4.5, JRip, and RF classifiers improved; (ii) the most important task was to remove collinear attributes and use COM_3 to integrate the key attributes, which also improved the accuracy of the C4.5, Jrip, and LMT classifiers; (iii) we applied FA to delete the attributes with insufficient cohesion, and although all performances did not improve further, the results of FA renamed the principal components.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 9. Results of the six datasets (full attributes) based on different accident frequencies.

https://doi.org/10.1371/journal.pone.0272956.t009

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 10. Results of three-stage dimension reduction for CF ≥ 1 and CF ≥ 10 datasets.

https://doi.org/10.1371/journal.pone.0272956.t010

In the most valuable CF ≥ 10 dataset, the removal of collinear attributes resulted in several improvements, especially in the accuracy and specificity of the LMT classifier because the LMT is the best RBML classifier across the five metrics. Compared to the CF ≥ 10 (full attributes) dataset, the deleted multicollinear attributes followed by the COM_3 integrated attributes also improved performance. The third stage (deleted multicollinear attributes + COM_3 integrated attributes + FA reduced the attributes that had insufficient cohesion) retained only 10 key attributes, which also improved the two-stage performance. In summary, the proposed three-stage dimension reduction discovered fewer key attributes to enhance the performance of the full-attribute dataset.

• Generating DT rules. DT is a popular traffic safety method that can quickly interpret complex patterns related to the attributes of road crashes [22]. Further, DT was the first method chosen to understand the rules of road accidents; hence, we used the DT diagram to show the rules of accident severity. The rules of the CR ≥ 10 dataset are shown in Fig 2, which indicates that "hit vehicle" was the most important attribute, and "moto" and "others" were the most prominent causes of injury/death in traffic accidents. The order of the top three attributes was hit vehicle > certificate type > vehicle.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. DT diagram of accident severity in the CR ≥ 10 dataset.

Note: Y denotes injury or death, N denotes uninjured.

https://doi.org/10.1371/journal.pone.0272956.g002

4.3.1 Key attributes.

According to Rolison et al. [43], the causes of vehicle collisions broadly depend on the driver characteristics, including skill level, inexperience, excessive speed, loss of control, failure to detect another vehicle, reckless driving, traffic violations, and drugs and alcohol. Furthermore, road crashes with older drivers often involved making mistakes at intersections, failing to give way when turning, failing to follow signs and signals, not seeing objects, and making incorrect turns and lane changes. Dimension reduction can extract important attributes without reducing the performance of the data analysis, and its purpose is to reduce the number of initial attributes to select a subset that retains sufficient information to obtain satisfactory results. After implementing the experiments, we discovered that the top order of top ten key attributes was Hit_veh > Certificate type > Vehicle > Action_t > Drive_q > Escape > Accident_t > Gender > Job > Trip_p in the most accident frequencies CF ≥10 dataset.

4.3.2 Results of changing the different stages of ordering.

In the three-stage dimension reduction, the ordering of the deleted multicollinear attributes and COM_3 integrated attributes would impact their evaluation results, because the two methods considered the target class. Therefore, we experimented with a different ordering of the two combined methods, and the results of deleted multicollinear attributes and COM_3 integrated attributes are listed in Table 10. Furthermore, the results of the COM_3 integrated attributes and deleted multicollinear attributes are listed in Table 11. Comparing Tables 10 and 11 based on the best performance of the LMT classifier, we observe that: (1) the proposed ordering (as Table 10) outperformed the COM_3 integrated attributes, and then deleted multicollinear attributes in the accuracy and AUC of CR ≥ 1 dataset; (2) in the CR ≥ 10 dataset, the proposed ordering is better than the COM_3 integrated attributes and then deleted multi-collinear attributes in accuracy; (3) the FA had no impact on the results of the evaluation metrics of the two datasets; however, it can check the cohesion of those attributes and remove those with insufficient cohesion.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 11. Results of COM_3 integrated attributes and then removing collinear attributes.

https://doi.org/10.1371/journal.pone.0272956.t011

In summary, the proposed ordering presented the best result, and the FA could rename the factor dimension to obtain the clustering status and intensity of attributes. We used FA to remove the age attribute in the CR ≥ 10 dataset because its absolute loading value was less than 0.5, as shown in Table 8; therefore, the full CR ≥ 10 dataset was reduced from 38 to 10 attributes. That is, FA could find the most useful attributes without reducing the evaluation performance, thereby improving the efficiency of data storage and processing [16].

Important descriptive statistics. The following key facts were summarized from the descriptive statistics of the complete CF ≥ 1 dataset, which are listed in Table 12:

1. In the vehicle dimension, motorcycles had the largest proportion of road collisions with 61.62% and 56.29% of the vehicle type of victim. The highest percentage of collision types (91.79%) was that of vehicles colliding with vehicles.
2. Regarding the human dimension, most road crashes occur at an excessive speed (83.08%), and male drivers are more likely (66.09%) to be involved in collisions. Furthermore, the accident cause is the driver at 66.19%, which is the same as the cause of vehicle collisions broadly depending on the characteristics of the drivers [43].
3. The road class, separating fast and slow lanes, road type, accident location, and signal type had the highest percentage of collision types in the road dimension.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 12. Road accident factors and attribute value more than 50%.

https://doi.org/10.1371/journal.pone.0272956.t012

Some suggestions. Based on discussions (1)–(3) and the results of this study, we provide the following suggestions to the government and stakeholders for reference:

1. It is necessary for the government to formulate basic laws and regulations for extensive traffic management and inform road users of the rules, methods, and essentials of safe road use. The government must manage, publicize, supervise, and ban them from achieving their results.
2. Traffic-related units must implement safe driving training and education (laws and regulations, safe driving, and accident prevention).
3. A rigorous test must be required to obtain a driver’s license to determine the drivers who can safely use a road.
4. Law enforcement units must continue to implement rigorous management and supervision, and ban offenders to eliminate drivers’ mentality of luck and speculation.
5. Road management agencies must make proper corrections and improve the quality of road projects to maintain driving safety, and the relevant driving safety settings must be fully applied.
6. Regarding the regulations, methods, and matters that should focus on pedestrians and slow vehicles, law enforcement agencies should be responsible for publicizing and educating them to prevent them from being affected or implicated in the safety of vehicles owing to illegal road use.