Analysis of Spatial Concentrations of Large-Truck Crashes Using Data Mining Methods

2.1 The contributory factors of crashes involving large-trucks

The previous studies have identified a wide variety of factors that influence the frequency, the severity of injury and the occurrence of crashes involving large-trucks. The categories of the contributory factors include the crash characteristics (e.g., collision type), vehicle characteristics (e.g., size and weight), driver demographics and actions (e.g., age, moving straight), roadway attributes (e.g., alignment, grade), land use (e.g., rural & urban), traffic conditions (e.g., traffic control & speed limit), and environmental factors (e.g., weather, lighting). Dong et al. [3]⁠ reported that driver age, speed limit, and location type have significant influence only on the frequency of crashes involving large trucks. Another study developed a driver-focused truck crash prediction model and found that age, weight, height, gender, employment stability, and previous driving and vehicle history are significantly related to the likelihood of truck crash occurrence.

Different collision type affects the injury outcome of crashes involving large-trucks differently under disparate conditions. Uddin & Huynh [4]⁠, and Zhu & Srinivasan [5]⁠ both reported that the rear-end collisions decrease the probability of major injury. Zhu & Srinivasan [5]⁠ also indicated that head-on collisions involving large-truck lead to more severe injuries. Several studies suggested that the number of vehicles and the characteristics of large-truck also significantly affect the injury severity of crashes. Zheng et al. [17]⁠ and Chen F. & Chen S. [18]⁠ reported that large-trucks with weights over 20,000 lbs and hazardous material cargo increase the likelihood of severe injury, respectively. Moreover, Chen F. & Chen S. [18]⁠ and Islam et al. [8]⁠ discovered that there are significant differences between the single-vehicle and multiple-vehicle crashes involving large-trucks. Roads with a right curve also increase the odds of injury in large-truck crashes [6]⁠. With regards to traffic conditions, Uddin & Huynh [4]⁠ reported that speed limits between 45 to 60 mph increase the odds of no injury under normal weather conditions, and speed limits over 65 mph increases the odds of major injury under rainy weather.

The effects of driver’s age on the injury severity of crashes involving large-trucks have been inconsistent in the past studies. For example, Pahukula et al. [19]⁠ reported that young drivers are more likely to decrease the chances of no injury large truck crashes. On the other hand, Chen et al. [20]⁠ found that young drivers are more likely to be associated with incapable injuries and fatalities. A few studies have explored the impacts of different environmental factors on the injury severity of large-truck crashes. Naik et al. [21]⁠ have explored the weather impacts on the injury severity of large-truck-only crashes. The results of the study indicated that wind speed, rain, humidity, and air temperature have an association with the injury of single-vehicle large-truck crashes. Another study analyzed the injury severity of large-truck crashes under different lighting conditions on the rural and urban roadways of Ohio, USA [9]⁠. The study concluded that the age and gender of occupants, types of trucks, speed, annual average daily traffic, curve roadways, and adverse weather have different effects on the injury severity under different lighting conditions. Concerning the effects of land use, Islam & Hernandez [11]⁠ indicated that crashes involving large-trucks in rural areas are likely to result in fatal injuries. In contrast, crashes in urban areas reduced the chances of incapacitating injuries. However, the categorization of crash locations into rural or urban put them in a broad definition of land use. The current study explored the more granular land use (e.g., population, housing, & employment density) and urban design (e.g., road network density) attributes, which rarely has been explored in previous studies.

2.2 The methodologies used for analysis of large-truck crashes

Until now, the majority of the studies have used statistical models to analyze crashes involving large-trucks. Dong et al. [3]⁠ used the negative binomial model to identify the factors that influence the frequency of large-truck crashes. Most of the studies that analyzed the injury severity of large-truck crashes have used the discrete-outcome models because typically the injury severity levels in crash data are reported as discrete values [22]⁠. The commonly used discrete-outcome models for injury severity analysis are multinomial logit, ordered logit/probit, random parameters logit/probit model, and Bayesian binary logit model [4, 10, 19, 23]⁠. However, the non-linear relationship between the contributory factors and injury severity makes the application of statistical models questionable. Moreover, several studies have reported that there are substantial correlations among the contributory factors. And, the effects of the contributory factors are difficult to estimate due to the correlations among the contributory factors [24]⁠. Considering these limitations, a few studies have used the machine learning models such as classification and regression tree (CART) [25, 26]⁠, and gradient boosting decision tree [17]⁠. However, a crash event occurs due to simultaneous or subsequent interactions of several contributory factors. The dependent and independent model structure only captures the effects of an individual contributory factor on the frequency or injury severity of crashes. The ARM technique not only can discover the set of crash contributory factors that often occur together in certain types of crashes but also identify the direction of associations among them. However, the application of the ARM technique for the analysis of road traffic crashes has been limited. Hong et al. [27]⁠ have used the ARM technique to analyze the risk factor of truck-involved crashes that occurred on the expressways of South Korea. Another study used the association rules to analyze the characteristics and contributory factors of work-zone crash casualties [28]⁠. The current study used the ARM technique to discover the contributory factors that are likely to facilitate different levels of injury severity at the spatial concentration of crashes involving large-trucks. In addition to the commonly explored crash contributory factors, this study explored the granular level land use and urban design attributes. The following part of the study includes a description of the crash contributory factors, the methodologies, the results, the discussion, and the conclusion.

4.1 DBSCAN (Density-Based Spatial Clustering of Applications with Noise)

Ester et al. [33]⁠ introduced the DBSCAN method to discover clusters in large spatial databases with noise. Most of the crash data include the geographic coordinates of the crash location. The DBSCAN clustering method can use the geographic coordinates to discover the regions with a high concentration of crashes. Also, roads are not always straight. There are intersections and curves in the road networks. Since DBSCAN can identify clusters with different shapes, it is suitable for application on road crash data. Multiple studies have applied DBSCAN clustering on road crash data [34, 35].

The DBSCAN algorithm has two major input parameters, which determine the density of a region. One is the Eps or ε neighborhood, which specifies the radius from a point to form the dense region. And, another is the MinPts, which specifies the number of points required to label an area around a point as a dense region. The DBSCAN clustering method uses these two input parameters to estimate the density around a point. It should be noted that for this study, a point indicates the geographic latitude and longitude coordinates of a crash event involving large-truck. Below, the cluster formation process of the DBSCAN algorithm is described in the context of road crash data.

Step 1: The algorithm selects an arbitrary point that has not been visited.

Step 2: Counts the number of crash events within the ε neighborhood of the point that was selected at step 1.

Step 3: If the number of crash events is equal to or more than the MinPts then the initially selected point is considered a core point. From the ε neighborhood of this core point, the cluster formation starts. On the other hand, if the number of crash events within the ε neighborhood of the point selected at step 1 is less than the MinPts then the algorithm moves to the closest point and repeats step 2. The algorithm may label the point selected at step 1 as noise or will make it a part of a cluster later, whose core point is within its ε neighborhood.

Step 4: The ε neighborhood of crash events or points within the initially identified cluster also become part of the cluster. If two core points are within each other’s ε neighborhood, then the two clusters are joined.

Step 5: The process above continues for other point in the crash data, and ends when all the points in the crash data are visited.

1.png

Figure 1. DBSCAN clustering

An illustration of the process of cluster formation by DBSCAN clustering on a road section would more helpful. Figure 1 above illustrates a segment of road, which is indicated by the space between the parallel black lines. The rectangular light gray shapes are markers like in the roads. For the figure above, let’s assume the MinPts = 6. In the figure above, the red stars are core points (i.e. P,Q & R), since there are six crash events within the ε neighborhood of those red stars. The dashed line from the core points to the border of circle indicates the Eps. The green points are called border point because they are within the ε neighborhood of the core points. One of the features of DBSCAN clustering is that a point is also part of a cluster if it close to many points from that cluster. Therefore, the three circles in the figure together indicate a cluster. The yellow stars are noise because they are not within the ε neighborhood of core points (P,Q & R).

4.2 Association Rule Mining (ARM)

The ARM technique is based on the market basket data analysis. It was first proposed by Agrawal et al. [36]⁠ to identify the frequent sets of items bought together in a transaction data of a supermarket. The technique is simple, and the association rules discovered by the technique are easy to explain. Let, D = {T1, T2, T3, ……., Tn} be a set of transactions, and I = {I1, I2, I3, ……..., In} be a set of items. Assume A and B are two items. A rule discovered by the association rule analysis for the aforementioned items is shown as “A⟶B”, where A is the antecedent, and B is the consequent. A rule can have multiple items as antecedent and consequent. In the context of market basket data analysis, the rule means if a customer buys A in a transaction, then the customer is more likely to buy B as well in the same transaction. For this study, a crash record was treated as a transaction in the market basket data. The nominal values of the crash contributory factors were treated as items of a transaction like in the market basket data. In this study, we have employed the commonly used Apriori algorithm because it is simple, and easy to explain, unlike other parametric and non-parametric methods. Generally, the apriori algorithm requires two steps. First, it iteratively scans the whole database for the frequent itemsets. In the second step, it generates association rules from the frequent itemsets. Support, confidence, and lift are the three important indicators of strong association for the rules generated by the apriori algorithm.

4.2.1 Support

The support indicates the proportion of the considered item in the data set. Support for a rule like A→B means the proportion of data records where A and B occur together. Generally, it is determined based on the characteristics of the data, and domain knowledge. The support is calculated using the following equation. Here, N is the total number of crash records. The support for a rule such as A⟶B is the same as B⟶A. Therefore, we need another measure, which considers the direction of the association.

$S=P(A \cap B) / N$                 (1)

4.2.2 Confidence

Confidence is a measure that takes into account the direction of a rule and helps to differentiate between the rules such as A⟶B and B⟶A. The confidence of rule A⟶B is defined by the ratio of occurrence of A and B together to the occurrence of A only. The confidence value of rule A⟶B indicates that the chance of occurrence of B increases with the occurrence of A. The confidence value is calculated using Eq. (2). However, the support and confidence alone cannot explain the significance of a rule.

$C=P(A \cap B) / P(A)$                        (2)

4.2.3 Lift

The lift (L) measures the strength of an association rule. The lift is the ratio between the confidence and the expected confidence of an association rule. The occurrence of A and B together with the occurrence of B is considered the expected confidence. The lift value ranges from 0 to ∞. The chances of A and B occurring together is more than expected when the rule A⟶B has a lift greater than 1. A lift lower than 1 indicates the opposite. There is no association between the items when the lift value is equal to 1. The lift is calculated using Eq. (3).

$L=P(A \cap B) / P(A) \times P(B)$                      (3)

Unfortunately, there is no standard rule to determine these parameters. The minimum support value ranged from 10 to 40 percent in some studies that used the ARM technique for road crash analysis [37-39]⁠. However, using such high minimum support will leave out interesting rules which may include rare but important crash items. In some studies, the minimum support ranged from 0 to 5 percent [40-42]⁠. It is fair to say that the criteria to set the minimum support value is subjective. It should be noted that the nominal values of the crash contributory factors are mentioned as the crash items in this study.

The current study employed data mining methods to discover the associations between the contributory factors and the different levels of injury severity at the spatial concentrations of crashes involving large-trucks. This study explored the crash characteristics, driver’s actions, road geometries, traffic conditions, and environmental factors. Additionally, the newly explored attributes included land use (e.g. housing, population, and employment density), and urban design (e.g. road network density). The application of the DBSCAN clustering method identified the spatial concentrations of large-truck crashes. The spatial concentrations that did not include crashes from the three periods were removed. Then, the remaining spatial concentrations were divided into three groups based on the proportion of severe injury crashes in each spatial concentration to obtain data sets that are less biased towards the no injury category. The apriori algorithm, which is a popular association rule mining technique was employed on each group of spatial concentrations to discover the conditions that lead to the no injury, non-severe injuries, and severe injuries in crashes involving large-trucks.

The proposed framework has significant practical implications. The crash data are generally recorded by police officers, who are not specialized in road crash analysis. Moreover, it is not possible to inspect all the locations of crashes due to limited resources. Investigation of the spatial concentrations of crashes involving large-trucks by road safety experts may reveal more relevant crash contributory factors.

Further, the application of the ARM technique discovered different sets of crash items that are responsible for the different levels of injury severity in crashes involving large trucks. For example, the combination of crash items that significantly increases the chances of severe injuries is the rear-end collision and drinking & driving. Checking the alcohol intoxication level of the drivers passing through those locations can reduce the number and severity of the aforementioned crashes. Additionally, the findings suggested that severe injuries in crashes involving large-trucks are more likely to occur on roads with a speed limit between 40 to 80 km/h, in areas with low employment density and medium road network density. Road safety authorities can improve road conditions, raise traffic control, and put up warnings such as “crash-prone locations” to curve out severe injuries crashes at those locations.

The rear-end collisions and medium or high population density both are frequent conditions of non-severe injuries. In medium or high population areas, traffic congestion is frequent. Such conditions may facilitate rear-end collisions since a lot of vehicles travel in close proximity during congestion. Traffic authorities can put up signs and warning for drivers to avoid driving too close to other vehicles.

Several sets of antecedents can lead to no injury every time they appear together. According to the results, collisions between large-trucks and fixed objects, sideswipe same direction collision, medium road network or employment density, snowy roads, and clear weather have a significant chance of leading to no injury crashes involving large-trucks. Reinforcing the current traffic controls and signals vigorously may reduce the number of no injury crashes involving large trucks.

In summary, the current work was an interesting effort to obtain important and hidden insights about the crashes involving large-trucks. Future studies may apply another clustering algorithm to segment the spatial concentrations into more homogeneous groups. Also, future studies may increase the maximum number of crash items in the rule, which can reveal more insights about the conditions that lead to different levels of injury in crashes involving large-trucks.