category: literaturenote citekey: ziakopoulosexplainablemacroscopicmicroscopicinfluences2025 title: Explainable macroscopic and microscopic influences of COVID-19 on naturalistic driver aggressiveness derived from telematics through SHAP values of SVM and XGBoost algorithms authors: "Ziakopoulos, Apostolos; Sekadakis, Marios; Katrakazas, Christos; Kallidoni, Marianthi; Michelaraki, Eva; Yannis, George" year: 2025 date: 2025-00-00 2025 doi: 10.1016/j.jsr.2024.12.010 publication: Journal of Safety Research url: "https://doi.org/10.1016/j.jsr.2024.12.010" zotero_key: GNIPGZGJ zotero_storage: SAN9TCIH collections: transpordi_planeerimine folder: 001_artiklid firstAuthor: "Ziakopoulos, Apostolos"

status: converted

Contents lists available at ScienceDirect

Journal of Safety Research

journal homepage: www.elsevier.com/locate/jsr

Explainable macroscopic and microscopic influences of COVID-19 on naturalistic driver aggressiveness derived from telematics through SHAP values of SVM and XGBoost algorithms

Apostolos Ziakopoulos * , Marios Sekadakis , Christos Katrakazas, Marianthi Kallidoni , Eva Michelaraki , George Yannis

National Technical University of Athens, Department of Transportation Planning and Engineering, 5 Iroon Polytechniou St, Athens GR-15773, Greece

ARTICLE INFO

Keywords: COVID-19 Driver behavior Naturalistic driving Smartphone data Explainable machine learning SHAP values

ABSTRACT

Introduction: This study aims to quantify the impacts of the COVID-19 pandemic on driver behavior as expressed by harsh accelerations (HA) measured by smartphone telematics data. Method: Over 35,5000 naturalistic driving trips were analyzed, fused with additional data sources such as: (a) Apple driving requests; (b) Oxford government response metrics; and (c) Our World in Data metrics for the COVID-19 pandemic. Machine learning algorithms were implemented on two scales: (a) a macroscopic scale involving daily analysis of aggregate driver behavior across the network with an SVM algorithm; and (b) a microscopic scale, involving trip-based analysis of driver trips with an XGBoost algorithm. SHAP values interpret the outputs of both algorithms, quantifying the influence of pandemic indicators with driver behavior and aggressiveness. Results: Macroscopic results (i.e., daily analysis) indicated that high total average speed values reduce HA rates, while this trend reverses with high driving speed. High values of Reproduction Rate, Total Cases per million people were found to reduce HA rates, while Total Fatalities per million people have little contribution on HA rates. Microscopic results (i.e., trip-based analysis) indicated that high speeding, total trip distance, and trip duration are associated with increased HA counts. Drivers perform more HAs on speeds between 30–50 km/h, while after 50 km/h, the contributions of speed lead to fewer HAs. A mild HA reduction was observed as Apple driving requests increase. Mild HA reductions also manifest when COVID-19 new daily cases and total cases per million increase as well. Drivers performed more HAs when daily deaths from COVID-19 were either relatively low (around 0–20 fatalities) or relatively high (around 110–120 fatalities), while the Stringency Index has an unclear contribution, indicating that pandemic measurements were more influential on HA counts compared to policy measures taken by the state.

1. Introduction

The outbreak of the Coronavirus Disease of 2019 (COVID-19) pandemic, also known as the SARS-CoV-2 virus, has influenced human norms since the very beginning of the spread in December 2019 (Zhu et al., 2020) and continued to do so in constant waves for more than two years. In order to mitigate the contagion and the virus spread, the majority of countries around the world imposed strict measures, such as constraint of all non-essential movements or even stay at home requirements (i.e., lockdowns), with the total aim of reducing human activity and contact, which contributes to the spread of COVID-19.

Intending to halt the high transmissibility of the virus, strict

restrictions on non-essential activities were implemented by the majority of governments across the globe. Various confinement measures, such as lockdown orders, school and non-essential business closures, and restrictions on travels were imposed in order to slow down the spread of the disease. As the case and death number diminished, restrictive policies were gradually lifted, altering the mobility patterns anew. Much research has been conducted in this area to investigate the dynamics of the pandemic in various countries all over the world and better understand the impact of COVID-19 on the transportation sector (Sharifi & Khavarian-Garmsir, 2020). While the medical community has sprung into action to address the public health impacts of the pandemic, researchers in other disciplines are dedicating efforts to understand the

E-mail address: apziak@central.ntua.gr (A. Ziakopoulos).

^* Corresponding author.

many ways in which societies have been impacted. Regarding road transport and road safety, it is highly fruitful to explore the manner in which driving behavior has been affected by the pandemic.

Building upon recent studies, the aim of the present research is to detect and quantify the impacts of the COVID-19 pandemic on driver behavior and aggressiveness as expressed by harsh accelerations (HA) measured by smartphone telematics data. The core hypothesis of this study is that the COVID-19 pandemic, along with the corresponding government responses, has measurable impacts on driver behavior. These impacts can be detected through telematics and mobility data both at an aggregated daily level and at the disaggregated trip level. It should be mentioned here that HAs are not the only indicator for driver aggressiveness, as speeding, harsh braking and swerving can also be also aggressiveness factors.

To achieve this objective, machine learning algorithms are implemented on two scales: (a) a macroscopic scale, which involves a daily analysis of aggregate driving behavior across the network with a Support Vector Machine (SVM) algorithm; and (b) a microscopic scale, which involves a trip-based analysis of driver trips with an Extreme Gradient Boosting (XGBoost) algorithm. The selection of two potent but different machine learning algorithms was driven by the desire to showcase the flexible applicability of SHapley Additive exPlanation (SHAP) values and the informative power explainable machine learning (ML) approaches can provide, rather than the absolute superior performance of a specific algorithm. Microscopic examinations provided insights on the correlations of increased HA rates with pandemic indicators on a lower trip-level basis, while macroscopic modeling revealed the impact of the pandemic on HAs for the entirety of the driving sample during the COVID-19 period. After the development of the ML models, SHAP values are used to interpret the outputs of both algorithms, providing a quantification on how pandemic indicators affect driver behavior and aggressiveness.

The contributions of the present study can be expressed as follows: This is the first study attempting to demystify complex non-linear relationships regarding quantified driver aggressiveness during the COVID-19 pandemic by utilizing explainable ML tools, to the authors' best knowledge. These investigations actively draw from driver behavior telematics data, which are innovatively augmented by the aforementioned multi-source data pillars regarding COVID-19 spread but also government response, the exploitation of which is highly novel, and, more importantly, informative. This is further evidenced by the complex relationships that were discovered, underlying within the data, which would not have been easily identified by more traditional modeling techniques. Overall, utilizing naturalistic telematics measurements of HA provides a valuable approach, in the capacity that HAs serve as a quantified objective measure of aggressiveness, as opposed to self-reported questionnaires that might be misinterpreted while also suffering from response bias (Rosenman et al., 2011).

The paper structure is the following: initially, the literature with regards to driving behavior, road safety and traffic volumes during the COVID-19 pandemic is reviewed. Subsequently, the following sections present the methodological background for the deployed methods and an overview of the data needed for the analysis. Finally, both the microscopic and macroscopic results are provided and the main findings are discussed, before closing the paper with conclusions drawn for researchers and practitioners.

2. Literature review

2.1. Direct road safety impacts of the pandemic

Within the context of the present study, direct road safety impacts comprise indicators regarding direct road safety parameters such as crashes and injuries. Many researchers are endeavoring to explore the dynamics of the pandemic in several countries across the globe in order to understand the impact that COVID-19 had on road traffic safety and shared mobility. To begin with, Jaekel and Muley (2022) examined the road traffic safety impact of the pandemic in Germany and Qatar and performed an assessment during the first wave of COVID-19. It was revealed that mobility patterns were reduced apart from trips to residential locations and parks in Germany. Moreover, the total number of fatalities, serious, and slight injury crashes was reduced during the first wave with Qatar showing a higher decrease than Germany. Similarly, an assessment of traffic driving and walking volumes during the initial months of COVID-19 indicated that the total number of crashes as well as traffic volumes in Qatar were reduced by 37% and 30%, respectively (Muley et al., 2021).

In terms of macroscopic road safety, the number of road crashes, injuries, and fatalities had fallen dramatically during the COVID-19 lockdown measures, particularly during the first lockdown phase. In Greece and Spain, a sharp decrease in traffic crashes, fatalities, and injuries was revealed (Saladi´e et al., 2020; Sekadakis et al., 2021). Similarly, Carter (2020) indicated that during the COVID-19 lockdown period (i.e., from March 15, 2020 to May 16, 2020), road crashes and fatalities in North Carolina decreased by 50% and 10%, respectively, while serious injuries increased by 6%, compared to pre-lockdown baseline. However, on the contrary to the reduction of road crashes, another study indicated that the fatality and injury rates in Greece increased during the same period (Sekadakis et al., 2021).

Another study was conducted in order to examine the impact of lockdown measures taken to control the spread of COVID-19 on road fatalities (Gupta et al., 2021). To achieve this goal, data were collected for a 6-month timeframe for 15 countries based on severity of COVID-19. Lockdown stringency and other daily mobility measures were selected to investigate the change in road mobility patterns due to lockdown measures. Results revealed that the high residential mobility as well as the lockdown measures reduced the crash fatality ratio. Thus, stringent law enforcement and appropriate compliance from road users can effectively lessen road traffic fatalities in unprecedented times.

Roe et al. (2021) investigated the effect of the pandemic on daily driving behavior of older adults before and after the advent of very rapid increase in the total number of U.S. COVID-19 cases. Results indicated that drivers reduced the proportion of days driven as well as the length of trips they took during the pandemic compared with the same time period during the previous year. Participants managed to drive slower with fewer speeding incidents and had different trip destinations. Thus, it was revealed that older adults moderated the quantity and quality of their trips when faced with a pandemic.

An interesting study on travel behavior before and during the pandemic revealed that COVID-19 altered the transport mode preferences considerably but a small variation in the frequency of trips was identified (Anwari et al., 2021). It was found that during the first wave of the pandemic, the total number of fatalities, major and minor crashes, and traffic volumes reduced and was coupled with restrictive measures rather than the COVID-19 incidences. Interestingly, driving volumes showed a statistically significant inverse linear relationship with the stringency index not only on weekdays but also on weekends. The transportation mode choice was affected, especially during the first pandemic wave (Bucsky, 2020); nevertheless, the decrease in use was not identical for all transport modes. In particular, public transport saw the greatest decline.

In addition, road safety exposure metrics have been observed to fluctuate, such as distance/miles driven. A remarkable decrease in vehicle miles driven as well as in days driven per week by 35% and 37% respectively, was identified among adolescents (Stavrinos et al., 2020). Among other findings, it was reported that driving time and distance decreases were lower among ethnic minorities, older and employed adolescents, but higher among people with greater dire prosocial tendencies. A similar study was conducted in order to investigate the effect of mobility restrictions for COVID-19 containment on urban vehicular traffic on the road network. Results showed that private vehicle trips were reduced by 65% during the lockdown period (Aletta et al., 2020). Lastly, Anke et al. (2021) demonstrated the impact of the pandemic on driving and walking volumes. The data confirmed a profound effect of COVID-19 on mobility patterns with a shift away from the modal share of public transport and increases in car usage as well as high proportion of cycling and walking.

Michelaraki et al. (2021) explored the impact of the COVID-19 pandemic on driving behavior and mobility patterns during 2020 in four countries, namely Greece, Saudi Arabia, Cyprus, and Brazil. A descriptive exploration of a 12-month timeframe of data was conducted and driving or walking traffic volumes, driving behavior indicators (i.e., speeding percentage, average driving speed, harsh accelerations/100 km, harsh brakings/100 km, total duration, driving duration) as well as road crashes were collected and analyzed. The results indicated that speeding percentage, average driving speed, and harsh events (except for Cyprus) were increased during the lockdown period. At the same time, a reduction in traffic volumes and the total number of people walking in each country was also observed.

Nonetheless, from the examined studies, it is evident that most of the road safety impacts of the pandemic are measured through traditional, hard indicators such as crashes, fatalities and injuries. There is much untapped knowledge to be gained, however, by examining surrogate safety measures (SSMs) such as harsh driving events. Harsh events are much more frequent and prevalent throughout the network, providing a more complete and seamless picture of the influences of the pandemic on specific aspects of road safety, such as driving behavior as expressed through aggressiveness measured by harsh events from telematics. Since there is limited evidence regarding the road safety impacts of the pandemic at the microscopic and macroscopic analysis scales, exacerbated by the inherent scarceness of crashes, surrogate safety measures are a promising vehicle to fill this gap. Therefore, the motivation for the current paper is obtained, as the paper aims to quantify the impacts of the COVID-19 pandemic on driver behavior and aggressiveness on both at macroscopic and microscopic scales, namely per day throughout the sample and per driver trip respectively.

2.2. Driver behavior evaluations during the pandemic

Many studies were undertaken in order to examine aspects of driving behavior during the COVID-19 pandemic. Katrakazas et al. (2020) investigated the impact of COVID-19 on driving behavior using road safety and naturalistic data indicators before and during the first two months of the pandemic, focusing on Greece and the Kingdom of Saudi Arabia. The study found that reduced traffic volumes due to lockdowns led to a slight increase in driving speeds by 6–11%, a 12% rise in harsh driving events, and a 42% increase in mobile phone use while driving. Similarly, Katrakazas et al. (2021) reported that during the pandemic, average speed, speeding, and harsh braking incidents per 100 km increased, while the total number of road crashes in Greece decreased by 49% compared to the pre-pandemic period.

An interesting study aiming to examine the effects of COVID-19 on road safety during and after the pandemic was conducted by Vingilis et al. (2020). Preliminary information revealed that COVID-19 had a negative impact on road safety and there were several factors brought about by the COVID-19 pandemic, such as speeding and stunt driving, increased consumption of alcohol and drugs, as well as anxiety and stress. Moreover, Vanlaar et al. (2021) investigated self-reported risky driving behaviors, such as speeding, distraction while driving, and consumption of alcohol or drugs during the COVID-19 pandemic. The majority of respondents reported that their driving behavior did not change, and most positively, a small proportion revealed that they were less likely to engage in risky driving behaviors. However, it should be mentioned that notable driver proportions answered that they were more likely to engage in risky driving behaviors during the pandemic, as compared to the period before the appearance of COVID-19. For instance, the driving behaviors most frequently reported by the subset of drivers who admitted to being more likely to engage in risky actions during the pandemic, included speeding, drinking and driving, and distraction, with frequency increases of up to 7%.

Wagner et al. (2020) demonstrated that during the beginning of the pandemic, decreased traffic volumes and, therefore, congestion led to an environment with more opportunities for speeding. Furthermore, in Alberta, police recorded a 200% increase in the number of drivers exceeding the speed limit by more than 50 km/h (Heidenreich, 2020). In British Columbia, a 700% increase in the total number of cars impounded for excessive speeding during March 2020 was identified (Chan, 2020). Similarly, in California, from March to May 2020, average speeding during peak hours increased by 10 to 15 miles per hour (Hughes et al., 2021). Finally, in Ontario, Toronto Police reported a 35% spike in speeding tickets and an almost 200% rise in excessive speeding (i.e., driving at least 50 km/h over the speed limit) compared to the same period the previous year (City of Toronto, 2020).

2.3. Interpreting road safety machine learning models through SHAP values

In recent studies, machine learning interpretability has gained momentum to understand and explain how several contributing factors affect crash probability and severity. Initially, Parsa et al. (2020) utilized SHapley Additive exPlanation (SHAP) values to shed light into the results of the XGBoost algorithm they used to predict the probability of real-time crash occurrence in expressways in Chicago. They concluded that the difference in speed before and after crashes at the upstream direction have the greatest impact on crash probability, whereas low values of speed and volume at upstream locations after crashes lead to a higher probability for crashes.

Similar methodologies (i.e., combining XGBoost with SHAP) were also used by later studies (Chang et al., 2022; Wen et al., 2021; Yang et al., 2021) on real-time crash prediction and driver engagement with mobile phones (Ziakopoulos et al., 2023) with high degrees for both accuracy and interpretability of results. SHAP has also been found the most efficient way for model interpretability in a recent review on explainable machine learning and road safety studies by Wen et al. (2022). Further to the previous, Yuan et al. (2022) utilized SHAP to detect the heterogeneity of traffic parameter influence on conflict occurrence in a real-time framework. Specifically, the authors identified traffic volume, variation of speed and truck percentages as parameters influencing conflicts and, that, in addition, there appear to be critical pairing patterns of these factors affecting real-time safety levels. The application of SHAP enhanced random parameters modeling in revealing heterogeneity.

2.4. Overall literature findings

The existing literature on road safety during the COVID-19 pandemic primarily focuses on traditional indicators such as crashes, fatalities, and injuries, often overlooking more nuanced yet voluminous measures, for example surrogate safety measures (SSMs) such as harsh driving events. As per the aforementioned, SSMs offer valuable, previously unexplored insights into shifts in driving behavior during the pandemic by providing a more continuous and comprehensive view of the road network. By examining the literature, it is apparent that the impacts of the pandemic on driver behavior are different in each country and region, as they are influenced by different pandemic spread timeframes, cultures, government responses, and overall varying regulatory adherence.

This information has the potential to be found encoded in several different datasets measuring specific aspects of the transport network. There is therefore inherent value in combining the different rich but otherwise fragmented data sources that were created regarding the pandemic in an effort to uncover more of the underlying related effects. Moreover, by examining these data across different aggregation levels, manifold insights can be gained from the same case study. Thus, a research gap is presented for subsequent research efforts, such as the present study, to fill. Specifically, the examination of driver behavior through data fusion of multifaceted data and across different microscopic and macroscopic angles of analysis was considered. In addition, the aforementioned variety of results provides strong evidence of the existence of non-linear relationships, therefore applying non-linear methodologies was considered as a fruitful analysis direction.

3. Methodological background

This section briefly outlines: (a) the methodologies utilized in this research, mentioning the SVM and XGBoost algorithms; (b) the metrics employed for their evaluation; as well as (c) the SHAP values that are implemented for higher model explainability and interpretability.

3.1. SVM & XGBoost algorithms

Concerning the macroscopic analysis of driver behavior and aggressiveness during the pandemic, Support Vector Machines (SVMs) are utilized to perform multi-output regression. SVMs are supervised ML algorithms, mostly used for classification problems. Their key aim is to estimate the optimal hyperplane by maximizing the margin between the support vectors of each class, implementing a kernel trick in case of nonlinear mapping (Yu & Kim, 2012). SVMs have been extended to Support Vector Regression (SVR) by introducing a ε-insensitive loss function (i. e., a penalty-free error tube around the hyperplane).

For the microscopic analysis of HA counts per trip during the COVID-19 pandemic, Extreme Gradient Boosting (XGBoost) algorithms (Chen & Guestrin (2016) are employed. XGBoost algorithms are optimized supervised ML algorithms encompassing multiple Classification And Regression Trees (CART). Due to its adaptability and effectiveness, XGBoost has been consistently outperforming competitor methodologies in numerous ML competitions (Nielsen, 2016).

3.2. Model evaluation metrics

In order to compare model performance of the several configurations (hyperparameters and mix of considered inputs), and also to obtain performance estimates across macroscopic and microscopic levels, wellestablished machine learning error metrics were calculated, such as the Mean Absolute Error (MAE), the Mean Squared Error (MSE) and the Root Mean Squared Error (RMSE).

In lieu of percentage metrics, and to have common metrics for continuous and count data, a simple custom accuracy (CA) metric is introduced. The reasoning behind custom accuracy is to denote as acceptable each prediction that falls within an acceptable margin from the truth. With CA, an intuitive percentage of correct predictions is provided. In the context of a continuous variable, such as HA rates, predictions within 10% of the true value can be accepted as accurate. CA is expressed as follows; firstly the total number of predictions is classified as accurate or not accurate:

$$N_{\hat{y_i},10\%} = \begin{cases} accurate, & if \frac{|y_i - \widehat{y_i}|}{y_i} \le 0, 1 \ & not & accurate, & otherwise \end{cases}$$ (1)

Subsequently, CA is obtained as the percentage of accurate predictions from the total:

$$CA{5\%} = \frac{N{\hat{y}i, 10\%}[accurate]}{N{total}}$$ (2)

As CA is a percentage of accuracy, it features values between 0–100%; higher values indicate models with better predictive ability. The 10% threshold is adopted for the purposes of the present research. The same process is followed for count data, such as HA numbers, with all count predictions within a ± 1 tolerance from the true value being accepted as accurate.

3.3. SHAP values

A significant contribution of the paper is the exploitation of the SHapley Additive exPlanation (SHAP) values for interpreting the ML algorithm results of the present multisource data. SHAP is an explainable machine learning method assigning feature importance scores to each input variable for each prediction, instead of collectively. Shapley values were developed by Shapley (1953), and distribute total gains to the contributing players under the assumption that they all cooperate, defined as:

$$\phii = \sum{S \subseteq N \setminus {i}} \frac{|S|!(n-|S|-1)!}{n!} \left[ \nu(S \cup {i}) - \nu(S) \right]$$ (3)

Where N is a group with n features, S is a coalition within that group, yielding a worth ν(S) of coalition, which is the total expected sum of payoffs that can be obtained by cooperation of coalition members. SHAP values decompose the model prediction based on coalitional game theory (Molnar, 2020). Following Lundberg and Lee (2017), SHAP can be defined as:

$$g(z') = \phi0 + \sum{i=1}^{M} \phi_i z_i$$ (4)

Where g is the explanation model, *z*ʹis the coalition vector, assuming the value of 1 when features are observed and 0 otherwise, ϕⁱ ∈ R is the contribution of feature i and M is the number of simplified input features. As SHAP is model-agnostic, it can be utilized to interpret the outputs of both the macroscopic SVMs and the microscopic XGBoost models of the present research. SHAP values are applicable to both classification and regression problems.

4. Data sources

In the present study, datasets are superimposed temporally for the entire year of 2020 to obtain insights from before, during, and after the emergence of COVID-19. The start of the COVID-19 pandemic in Greece is recognized as the date of the first known diagnosed case, on February 26, 2020. Greece followed the WHO declaration of end to COVID-19 as a global health emergency, on May 5, 2023, which is more than three years into the pandemic, and after the present study period. In 2020, two strict lockdown restriction periods limiting non-essential movement were declared (dates can be seen in Table 1). Some restrictions to specific activities followed in subsequent years but were less strict and out of the present study period and scope.

4.1. Oseven database

OSeven provided a random dataset with naturalistic driving trips from its database in order to further investigate the macroscopic and microscopic influence of COVID-19 on driver aggressiveness. The database covered the period from January 1, 2020 to December 31, 2020, which included thousands of travels across Greece. The aforementioned one-year dataset includes data from before and during, the pandemic and after the imposition and lifting of two non-essential

Table 1 Lockdown Measures – Key Dates in Greece.

Greece – Lockdown Measures
1st Lockdown restrictions on non-essential movements	23 March 2020 → 04 May 2020
2nd Lockdown restrictions on non-essential	07 November 2020 → 31 December
movements	2020
	(Continued into 2021)

movement lockdown restrictions.

It should be noted that OSeven Telematics (oseven.io) uses a smartphone application and a platform in order to exploit data from smartphone sensors (e.g. GPS, accelerometer, and gyroscope data). For each trip, a vast quantity of data were collected, communicated through Wi-Fi or cellular network, and valuable critical information such as features, highlights, and driving scores was generated in order to assess the driver's profile and performance. Data were then transferred to the OSeven backend infrastructure, where it was analyzed with filtering, signal processing, machine learning algorithms, and safety/eco rating models. From the OSeven dataset, seven measurements were used (i.e., average speed, HA/100 km (or HA rate), HA count, average driving speed (namely the average speed excluding stopping periods), distance traveled, driving duration, and distance during risky nighttime hours). In accordance with the General Data Protection Regulation (GDPR) and related EU legislation, the OSeven platform complies with explicit privacy policy and solid information security protocols. As a result, OSeven has given the data in an entirely anonymized manner, with no geolocation information for the journeys included in the dataset. Trips were selected within the general vicinity of Greece for the duration of the sampling period without examining any precise geolocation information due to GDPR compliance. A similar dataset was utilized in previous analysis by the authors (Katrakazas et al., 2020; 2021).

Regarding the classification of accelerations as HAs, there is not a specific threshold of the accelerometer value or similar rule-based approach for the determination of the harsh events, but rather intellectually protected Machine Learning algorithms and rules serving to detect spikes at any acceleration profile by the accelerometers regardless of values. Therefore, if the vehicle kinematics match the predefined ML patterns, the current acceleration is flagged as a harsh acceleration event and considered for the subsequent analyses, and this approach generates HAs across all ranges of speed values.

4.2. Apple mobility trend reports

Moreover, the analysis leveraged mobility data reports sourced from Apple (Apple, 2020), with a specific emphasis on driving navigation requests. These driving navigation requests were adopted as a proxy measurement to gauge the trend of traffic volume or overall mobility due to its availability at the time of the study. This proxy measurement of the mobility trends concentrating explicitly on driving requests within the Greek dataset from Apple Maps. The data were compiled by combining the number of daily driving requests made by Apple Maps users who asked for navigation directions. It is crucial to note that these Apple requests were measured as a percentage deviation from an established baseline of 100%, set on January 13th, 2020, prior to the advent of the COVID-19 pandemic within Greece.

4.3. Oxford database

An open-access database containing COVID-19 response measures, originally created by Oxford University, was exploited in the present study (Hale et al., 2020; Hale et al., 2021). Specifically, this database contains daily data from the COVID-19 government response tracker across the world. The main data categories of this database are containment and closure, economic policies, health system policies and miscellaneous policies related to COVID-19. Special focus was given on a metric titled "Stringency Index," which was used to quantify the Greek government response actions. This index is derived from the Oxford database as well. The stringency index, which ranges from 0 to 100, represents the government response stringency index and composite measure, which is based on nine response indicators, such as school closures, workplace closures, and travel bans, rescaled to a value between 0 and 100 (i.e., 100 = strictest response).

4.4. COVID-19 metrics and important dates

The dataset that was exploited and included all the COVID-19 metrics was the database of Our World in Data, 2020 (OWD, 2020). The OWD database captured the daily evolution of COVID-19 metrics in 2020 (i.e., COVID-19 Total Cases, Total Fatalities, New Cases, Total Cases per million people, Reproduction Rate, New Fatalities, and Total Fatalities per million people during the pandemic). Table 1 overviews the key dates related to Greek lockdown restriction measurements. Two lockdown periods of non-essential movements due to the COVID-19 pandemic were put in place by the Greek government within 2020.

4.5. Matching process and brief descriptive statistics

As a first step in the matching process, the OSeven database was used as basis. In that database, data from driver trips were filtered to remove outliers from the dataset and to ensure that the remaining entries were within realistic limits. This entailed removing all trips with more than 60 HA counts, with rates higher than 200 HA/km or with driving speeds higher than 180 km/h, as such numbers are typically unattainable in environments of Greece and not representative of driving conditions. Thus, 842 trips were discarded, and the remaining sample consisted of 355,320 trips. An overall workflow chart to facilitate the different aspects of the study is shown in Fig. 1.

Moreover, Table 2 includes a description for each measurement that is analyzed in this study and the corresponding data source of each measurement. The COVID-19 reproduction rate is based on calculations by Arroyo et al. (2022).

For the considered clean subset of 355,320 trips, descriptive statistics for the examined variables, such as mean, standard deviation, maximum value, and minimum value, are shown in Table 3. COVID-19 metrics, measurements, and mobility had a different sample size (N = 366) than driving data because they were daily observations for the year 2020. The datasets were merged for analysis purposes, with the daily value of the remaining datasets (i.e., OWD, Oxford and Apple datasets) being allocated to each journey (provided by OSeven). As a note, the Apple database was missing records for two days of the year 2020 (May 11 & 12), therefore, the respective sample size is slightly reduced (N = 364).

5. Results

5.1. Macroscopic analysis with SVM

For the macroscopic analysis, daily naturalistic driving data were analyzed for a timeframe spanning from 01/01/2020 to 31/12/2020. The response variable was harsh accelerations (HA) per 100 km, therefore the variable was expressed in a rate format. The SVM algorithm and analysis were implemented following the Scikit-learn package in Python environment (Pedregosa et al., 2011). In order to evaluate the model predictions, a train-test split of the collected data was implemented. The data are split into 80% for training (spanning January 2020 to mid-October 2020) and 20% for testing (covering mid-October 2020 to December 2020).

The GridSearchCV, which is an approach that exhaustively examines all parameter combinations (Pedregosa et al., 2011), was first utilized for the hyperparameter tuning, aiming to minimize the accuracy metric RMSE. The best-performing combination hyperparameters C, epsilon and gamma were also manually tuned and the optimized combination is presented in Table 4.

The obtained performance metrics for the test subset were: MAE = 1.012, MSE = 1.622, RMSE = 1.274, CA% = 85.14%. The value of CA indicates an accurate prediction with a 10% tolerance for more than 85% of the daily data. Given the accuracy metrics, the SVM algorithm performs well in predicting the evolution of HAs during the pandemic. In Fig. 2, observed and predicted data are depicted for the test set timeframe. The SVM model performance verifies the feasibility of conducting

Fig. 1. Overall workflow of the study.

predictions of aggregate driving behavior by integrating naturalistic driving and COVID-19 factors. It is also evident, from a visual inspection of the results, that the applied multi-output regression follows the trend of the observations and therefore can be considered realistic.

SHAP values were also calculated to interpret the contribution of each variable (feature) to model output. In Fig. 3, the SHAP summary plot is presented for every feature across all datapoints, as an enhancement of traditional feature importance, following Molnar (2020). Specifically, the position of each dot within the figure provides SHAP values (notably, not feature values) for each observation of a feature. The value of the feature in the graph is represented by the coloration from low (blue) to high (pink). The y-axis demonstrates the features sorted in descending order of importance and the x-axis shows the distribution of SHAP values from lowest to highest score. Within each feature, each data point is represented by a point indicator (dot), and points with identical x-axis values are stacked, increasing the width of each bar proportionately to the frequency of the data points at that value. Based on the figure, Average Speed is found to be the most influential feature followed by Average Driving Speed and Distance Traveled.

All variables have distinct SHAP values distributions for low and high feature values, without many overlaps in these values. The high variable values of Average Speed are associated with low SHAP values (i. e., they reduce the predicted HA rates which is the output of the SVM model); low feature values of Average Speed have high SHAP values (i. e., they increase the predicted HA rate output of the SVM model). Apple Driving Requests and Driving Duration follow the same SHAP value trend. On the contrary, Average Driving Speed presents a reverse relationship with the estimated SHAP values. Regarding COVID-19 factors, Reproduction Rate, Total Cases and Total Cases per million people are associated with lower HA rate predictions as they have negative SHAP values in higher feature values. On the other hand, Total Fatalities and Total Fatalities per million people present little to no contribution to the SVM model output.

The SHAP feature importance graph reveals a reverse mechanism of Average Speed and Average Driving Speed values on the model outcome. High Average Speed values contribute negatively to HA rate, while high Average Driving Speed values contribute positively. This difference may be explained by the duration of stops, which is excluded from the trip duration in case of Average Driving Speed. The removal of physical stops from one speed parameter allows the model to adjust and correct the contributions of Average Speed with those of Average Driving Speed, using patterns in their fluctuations to generate HA predictions. Moreover, SHAP values indicate that harsher driving behavior is observed in short-distance trips and less aggressive driving on middle and long-distance trips. Drivers also become more aggressive, as expressed by HA rates, in trips with a shorter driving duration.

Through the visual inspection of COVID-19 parameters SHAP values, fruitful insights are gained for aggregate driver behavior during the pandemic. Fewer recorded Apple Driving Requests are associated with a positive contribution to the HA rate, denoting more driver aggressiveness when fewer vehicles circulate due to imposed restrictive measures. Cases outbursts both in absolute and relative terms contribute to less performed HA/100 km. Similarly, in periods of high transmissibility (i. e., high Reproduction Rate values), there is a decrease in HA rate. Probably, the intensified police enforcement had a synergic role for diminishing HAs. It is also noteworthy that for small variable values of Total Cases, Total Cases per million people and Reproduction Rate, a positive boundary is formed, implying the limited effect of low disease prevalence on driving behavior. Total Fatalities and Total Fatalities per million people SHAP values are unclear and do not contribute significantly to the SVM model.

5.2. Microscopic analysis with XGBoost

The microscopic analysis aimed to determine if the pandemic-related characteristics had quantifiable impacts on the number of HAs per trip performed by drivers. For that purpose, the merged trip-based dataset was examined, comprising 355,239 trips. All microscopic analyses were conducted in R-studio (R Core Team, 2019). The full dataset was split to training/test subsets randomly, with an 80% to 20% ratio, while maintaining comparable distributions of the response variable (i.e., HA counts) in the examined subsets. Care was taken so that the datasets covered the same time periods. XGBoost algorithms were implemented on the training set, and their predictive capabilities were evaluated on the test set.

For this analysis, the response variable, HA, is in count data format. Therefore, Poisson-based cost functions were investigated. After examination of several alternatives, Tweedie regression was selected (Tweedie, 1984). This distribution is also governed by the Tweedie variance power parameter: values towards 1 shift the distribution closer to a Poisson distribution, while values towards 2 shift it closer to a Gamma distribution (XGBoost developer team, 2021).

Initially, hyperparameter tuning with 5-fold Cross-Validation was implemented, seeking the internal model architecture that minimized RMSE. In total, 8,000 hyperparameter combinations were examined, selected at random from the range shown in Table 5. The combination with the best performance was also fine-tuned manually subsequently. The process resulted in the determination of the optimized hyperparameters shown on the rightmost column and lasted 74 h on a workstation computer (CPU: Intel Core i7 8700 @3.2 GHz; RAM: 32 GB: 2 × 16 GB DDR4 @ 2400 MHz).

A Tweedie-based cost function was employed to train a 5-fold XGBoost regression model with the best-performing hyperparameters,

Table 2 Measurement description and data source.

| Measurements |------------------------ | Average Speed | Average Speeding | HA/100 km (or HA rate) | HA count | Average Driving Speed | Distance Traveled | Driving Duration | Distance in risky hours | Apple Driving Requests | COVID-19 Reproduction Rate | COVID-19 Total Cases | COVID-19 Total Fatalities | COVID-19 New Cases | COVID-19 New Fatalities | COVID-19 Total | COVID-19 Total | Stringency Index | Government Response Index | Stringency Legacy Index | School Closure | Containment Health Index | Cancel Public Events | Fiscal Measures | Workplace Closure | International Travel Controls | Testing Policy | Gathering Restrictions | Internal Movement Restrictions | Emergency Investment | Stay At Home Requirements | Investment in Vaccines | Measurement Description | | | -------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------|--| | Average speed during the total duration of the trip (km/h) | OSeven | | | Average speed over the speed limit in a trip, i.e. driving over the "Speed Limit + Tolerance" | OSeven | | | Number of harsh accelerations per distance (100 km) | OSeven | | | Number of harsh accelerations (count) | OSeven | | | Average speed during driving with stops been excluded from the duration of the trip (km/h) | OSeven | | | Total trip distance (km) | OSeven | | | Total trip duration (sec) | OSeven | | | Distance driven during risky hours (00:00–––05:00) in a trip (km) | OSeven | | | Requests for driving in Apple Maps (%): expressed as a change percentage (100% − baseline on January 13th, 2020) | Apple | | | Real-time estimate of the effective reproduction rate (R) of COVID-19. | OWD | | | Total confirmed cases of COVID-19 | OWD | | | Total deaths attributed to COVID-19 | OWD | | | New confirmed cases of COVID-19 | OWD | | | New deaths attributed to COVID-19 | OWD | | Cases per million
people | Total confirmed cases of COVID-19 per 1,000,000 people | OWD | | Fatalities per
million people | Total deaths attributed to COVID-19 per 1,000,000 people | OWD | | | Government Response Stringency Index: composite measure based on 9 response indicators including school closures, workplace
closures, and travel bans, rescaled to a value from 0 to 100 (100 = strictest response) | Oxford | | | The index records how the response of governments has varied over all indicators in the database, becoming stronger or weaker
over the course of the outbreak. It is calculated using all ordinal indicators. | Oxford | | | Oxford also reports a legacy stringency index that approximates the logic of the first version of the Stringency Index, which only
had seven components (compared to nine). | Oxford | | | Record closings of schools and universities | Oxford | | | This is a composite measure based on thirteen policy response indicators including school closures, workplace closures, travel
bans, testing policy, contact tracing, face coverings, and vaccine policy rescaled to a value from 0 to 100 (100 = strictest). If
policies vary at the sub-national level, the index is shown as the response level of the strictest sub-region. | Oxford | | | Record cancelling public events | Oxford | | | Announced economic stimulus spending. Note: only record amount additional to previously announced spending | Oxford | | | Record closings of workplaces | Oxford | | | Record restrictions on international travel. Note: this records policy for foreign travelers, not citizens | Oxford | | | Record government policy on who has access to testing. Note: this records policies about testing for current infection (PCR tests)
not testing for immunity (antibody test) | Oxford | | | Record limits on private gatherings | Oxford | | | Record restrictions on internal movement between cities/regions | Oxford | | in
Healthcare | Announced short term spending on healthcare system, e.g. hospitals, masks, etc. Note: only record amount additional to
previously announced spending | Oxford | | | Record orders to "shelter-in-place" and otherwise confine to the home | Oxford | | | Announced public spending on COVID-19 vaccine development. Note: only record amount additional to previously announced
spending | Oxford | |

following Bischl et al. (2016). The final Tweedie variance power value denotes that the examined data are closer to a Poisson distribution shape. The resulting feature importance metrics are shown in Table 6.

For the training dataset, the obtained performance metrics were: MAE = 0.618, MSE = 1.114,

RMSE = 1.056, CA±¹ = 92.05%. These error metrics are reasonable, as they pertain to count data which assume considerably larger values compared to rate data. Subsequently, the predictive capabilities of the created XGBoost algorithm were tested on the test (holdout) subset. The obtained performance metrics were: MAE = 0.867, MSE = 2.737, RMSE = 1.654, CA±¹ = 85.64%. The value for CA denotes that the XGBoost algorithm predicts the number of HAs in a given trip correctly with a ± 1 tolerance for more than 85% of the cases in a new dataset. Based on these metrics, XGBoost yields very satisfactory results overall.

A closer examination of the feature importance metrics reveals interesting underlying relationships. Specifically, the most important feature in terms of gain (feature importance) appears to be average speeding in km/h. Average driving speed, in absolute terms, also ranks considerably high at the fourth spot. This is expected, as aggressive drivers likely perform a HA and subsequently maintain higher speeds (speeding) or are already speeding and engage in HA as they navigate through traffic and other vehicles. The following features are exposure measurements, in the form of the total duration and distance of the trip, which are critical for good predictive performance. Their high rankings are consistent with findings of past related research (Ziakopoulos et al., 2020; Katrakazas et al., 2021; Kontaxi et al., 2021a), which reported trip distance and trip duration as significant exposure variables for harsh event count predictions. Driving requests from Apple (termed GRdriving) have a fairly high importance score, as they can be considered a proxy quantity for the overall vehicles present in the network, although they cannot readily substitute traffic volume.

All following variables are parameters describing the spread of COVID-19 directly or the measures taken by the Greek state in response to the pandemic, with the exception of risky hour driving. In particular, COVID-19 cases, both as new cases and Total Cases per Million appear to be the most informative parameter measuring effects on driver behavior. It is possible that case-related parameters were the easiest to communicate to the public during 2020 and thus the ones affecting private car mobility in the most direct way, affecting traffic and headways for HA generation more in turn. Afterwards follow parameters describing the epidemiologic Reproduction Rate (also known as R0) and new reported deaths. The Stringency Index followed in importance, indicating that a composite measure is more informative for HA generation by drivers. After the 13th position of feature importance, individual measures and state policies continue to emerge, with gradually decreasing importance.

Overall, it appears that parameters related to the trip score highest when examining HA counts per trip, followed by features describing the progress of the COVID-19 pandemic. At an initial glance, this is expected, as it is reasonable for particular trip characteristics to affect HA generation more than the overarching pandemic profile; nonetheless,

Table 3 Descriptive statistics of the analysis variables.

Measurements	Mean	Standard Deviation	Minimum	Maximum	Sample Size
Average Speed	35.15	18.88	1.97	168.38	355,320 trips
Average Speeding	3.99	6.00	0.00	139.47	355,320 trips
HA/100 km (or HA rate)	11.75	27.10	0.00	597.01	355,320 trips
HA count	0.89	1.95	0.00	121.00	355,320 trips
Average Driving Speed	42.58	17.56	5.57	178.36	355,320 trips
Distance Traveled	11.62	22.33	0.50	648.69	355,320 trips
Driving Duration	771.37	967.65	61.00	23900.00	355,320 trips
Distance in risky hours	0.37	4.01	0.00	427.70	355,320 trips
Apple Driving Requests	113.76	52.68	18.59	241.14	364 days
COVID-19 Reproduction Rate	0.83	0.52	0.00	1.48	366 days
COVID-19 Total Cases	21683.25	38415.56	0.00	138850.00	366 days
COVID-19 Total Fatalities	577.56	1104.33	0.00	4838.00	366 days
COVID-19 New Cases	367.16	666.91	0.00	3316.00	366 days
COVID-19 New Fatalities	12.56	27.15	0.00	121.00	366 days
COVID-19 Total Cases per million people	2080.32	3685.63	0.00	13321.43	366 days
COVID-19 Total Fatalities per million people	55.41	105.95	0.00	464.16	366 days
Stringency Index	48.38	28.17	0.00	84.26	366 days
Government Response Index	47.47	27.48	0.00	82.67	366 days
Stringency Legacy Index	54.45	29.99	0.00	88.10	366 days
School Closure	1.37	1.07	0.00	3.00	366 days
Containment Health Index	46.47	26.64	0.00	81.92	366 days
Cancel Public Events	1.35	0.94	0.00	2.00	366 days
Fiscal Measures	102562354.02	1609518624.26	0.00	2.9E + 10	366 days
Workplace Closure	0.72	0.90	0.00	2.00	366 days
International Travel Controls	2.12	1.32	0.00	3.00	366 days
Testing Policy	1.88	1.28	0.00	3.00	366 days
Gathering Restrictions	2.53	1.58	0.00	4.00	366 days
Internal Movement Restrictions	1.08	0.78	0.00	2.00	366 days
Emergency Investment in Healthcare	6534.25	132195.69	0.00	2681004.00	366 days
Stay At Home Requirements	0.99	0.73	0.00	2.00	366 days
Investment in Vaccines	8121.78	162077.17	0.00	3242505.00	366 days

Table 4 Hyperparameter optimization results for SVR for HA per 100 km in the training set.

Hyperparameter	Examined range	Optimized Value
C	1 – 1000	100
Epsilon	0.01 – 1	0.5
Gamma	0.0001 – 1	0.001
Evaluation metric	RMSE MSE MAE	RMSE

COVID-19 parameters do show feature importance contributions that are non-negligible. It can be also gleaned that in Greece during 2020, it was the epidemiologic behavior and advancement of the pandemic that governed private car mobility instead of any measures for opening/ closing personal and public functions.

As a last step in the microscopic analysis, SHAP values were calculated for the features with the 10 highest gain values to provide an explanation of the contributions of the more highly informative variables. Results per individual feature are shown in Fig. 4 for the top 10 contributing features. In essence, this figure constitutes SHAP

Fig. 2. Results of multi-output SVM prediction.

Fig. 3. Feature importance based on SHAP values for HA/100 km.

Table 5 Hyperparameter optimization results for XGBoost for HAs in the training set.

| Hyperparameter | Examined range | Optimized
Value | | |---------------------------|-------------------------------------------------|--------------------|--| | Learning rate | 0.000 – 0.500 | 0.015 | | | Τweedie variance
power | 1 – 2 | 1.2 | | | Gamma | 0 – 10 | 2.7 | | | Alpha | 0 – 10 | 7.0 | | | Lambda | 0 – 10 | 2.2 | | | Maximum tree depth | 1 – 50 | 25 | | | Number of rounds | 1 – 300 | 180 | | | Min child weight | 1 – 10 | 3 | | | Max delta step | 0.001 – 10 | 2.5 | | | Evaluation metric | RMSE MSE MAE Logloss
Poisson-nloglik | RMSE | |

contribution dependency plots. The SHAP dependency plots are graphical representations depicting the relationship between the values of a specific feature and the corresponding SHAP values. From Fig. 4, insights can be derived regarding the direction and magnitude of each variable influence, as well as the amount of non-linearity of the specific relationships. The SHAP calculation process and the SHAP dependence plotting were conducted following Chen et al. (2021). It is critical to note that SHAP values have to be interpreted within the existing model, features, and dataset, and not as isolated contributions of each variable.

It is evident that the driving variables have explicitly non-linear SHAP values that are 'left-tailed,' contributing in a respectively nonlinear manner to HA generation. In particular, speeding is shown to contribute negatively to HA count generation for small values (i.e. after about 5 km/h), but is associated with more HAs in a gradually diminishing number as shown in the first curve, as drivers reach physical limits for accelerating harshly in high speeds.

The SHAP graph for total trip distance is interesting, as it denotes harsher overall behavior (i.e., higher SHAP values for HA counts) gradually after trips of more than 100 m total distance based on the inflection point observed there, though several individual short trip cases with spikes of higher values can be observed. A comparable result is presented for total trip duration, hinting that drivers become more aggressive for trips lasting more than 1000 s approximately. The HA contributions of average driving speed are highly non-linear as well: it appears that drivers navigating between 30–50 km/h perform more HA maneuvers. Between 50 and 120 km/h, the contributions of speed revert

Table 6 Feature importance for XGBoost for HAs.

Number	Independent variables – Features	Gain	Cover	Frequency
1	Average Speeding	0.44750	0.22084	0.12131
2	Distance Traveled	0.14763	0.18031	0.16211
3	Driving Duration	0.14005	0.17415	0.14457
4	Average Driving Speed	0.09996	0.19006	0.18009
5	Apple Driving Requests	0.05379	0.07637	0.12833
6	COVID-19 New Cases	0.02804	0.03404	0.06873
7	COVID-19 Total Cases per million people	0.02464	0.03956	0.05218
8	COVID-19 Reproduction Rate	0.01968	0.02184	0.04927
9	COVID-19 New Fatalities	0.01536	0.01887	0.03735
10	Stringency Index	0.00754	0.00693	0.01838
11	Government Response Index	0.00390	0.00494	0.00953
12	Distance in Risky Hours	0.00332	0.01896	0.00757
13	Stringency Legacy Index	0.00185	0.00236	0.00441
14	School Closure	0.00180	0.00165	0.00441
15	Containment Health Index	0.00173	0.00277	0.00389
16	Cancel Public Events	0.00083	0.00080	0.00196
17	Fiscal Measures	0.00076	0.00260	0.00210
18	Workplace Closure	0.00059	0.00068	0.00133
19	International Travel Controls	0.00023	0.00030	0.00053
20	Testing Policy	0.00019	0.00013	0.00046
21	Gathering Restrictions	0.00018	0.00019	0.00042
22	Internal Movement Restrictions	0.00017	0.00019	0.00034
23	Emergency Investment in Healthcare	0.00011	0.00103	0.00032
24	COVID-19 Total Fatalities per million people	0.00006	0.00008	0.00015
25	Stay At Home Requirements	0.00006	0.00005	0.00012
26	Investment in Vaccines	0.00005	0.00033	0.00014

to negative, hinting at less aggressive behavior once drivers have attained a considerable absolute speed. The SHAP curve assumes positive values for a few cases after that, denoting highly excessive behavior in terms of speed and speeding simultaneously, leading to more HAs as a consequence.

Consistently with Gain values, COVID-19 parameters have more limited but non-negligible SHAP values. Driving requests from Apple (GRdriving) display a mild downward tendency, especially after values compared to 150% of the baseline (January 13th), which is expected as more vehicles circulate on the network, and correspondingly allow fewer headways sufficient for HAs. A similar downward trend is observed for absolute numbers of COVID-19 daily cases, outlining the

Fig. 4. SHAP values for the microscopic XGBoost individual model features for HAs.

opposite mechanism: beyond 1000 new daily cases fewer HAs are generated per trip, as it is possible that drivers are more careful with increased police enforcement present, and feel less of a need to resort to abrupt behavior in emptier road networks. Regarding the total spread of COVID-19, the downward trend reducing HA counts is observed mostly after circa 2000 cases per million, and the reasons are likely similar to those for new cases.

The R0 reproduction rate seems to have an oscillating and very circumstantial effect on SHAP values for HAs when the pandemic was in a recession (i.e., before 1.0), but when it was on the rise with values higher than 1.5, it had a negative contribution to HA generation, following the previous two indicators. It is worth noting, however, that new deaths by COVID-19 had a slightly U-shaped effect on driver aggressiveness as expressed by HA counts. Specifically, drivers performed more HAs at low (0–20) and high (110–120) daily COVID-19 deaths, while they were more conservative towards the middle values. This might be indicative of a public that is less compliant to restrictions as time progresses and the pandemic does not subside, reverting to pre-pandemic driver behaviors.

Lastly, the SHAP values for the Stringency Index are unclear and do not convey a coherent effect, indicating again that the pandemic was much more influential on driver aggressiveness compared to the policy measures taken by the state.

The influence of all features of the microscopic XGBoost analysis can be also examined collectively, as visualized in Fig. 5, created following the work of Liu and Just (2021). The figure displays a 10,000 indicative randomly extracted subsample of the total examined sample for readability reasons; similarly, examination of SHAP values for each unseen sample observation is possible. Once again, the first 10 variables/features are shown, while the rest are merged. The respective influence and contribution of each feature to HA generation is evident as previously discussed.

Overall, Fig. 5 serves as a visualization of the contributions of each variable and their voting process to determine the final SHAP score. The reversal observed after roughly the first 5,800 observations can be attributed to the randomness of the sample. Despite this, this reversal is indicative of the underlying non-linear functions of the XGBoost tree ensemble, and serves to illustrate the shifting contributions some variables may have amongst observation predictions.

The larger contributions are offered by the variables with darker color, which rank highly on the legend. Contributions from COVID-19 related features are present and quite discernible towards the x-axis of the graph.

6. Discussion

6.1. Current findings

The COVID-19 pandemic upset normal operations in arguably all aspects of life for citizens worldwide. With regards to transportation safety, "empty" roads have led to less risk exposure due to limited traffic flow and fewer fatalities from crashes, however the reduction of fatalities is not proportionate with the reduction in traffic volumes (Wegman & Katrakazas, 2021). Recent studies and reports have succeeded in identifying the quantified impact of the pandemic in safety indicators such as crashes and fatalities as well as risk exposure in terms of modal split and traffic volumes. However, the individual interpretation of the impact of COVID-19 indicators such as the strictness of measures and the number of cases and related fatalities on macroscopic and macroscopic safety is yet to be realized. This fact formed the motivation for the current paper, which aimed to investigate the effects of COVID-19 and driving behavior indicators on both a macroscopic scale (i.e., aggregated driving behavior across a network) and a microscopic scale (i.e., tripbased analysis of individual driving trips). To accomplish this task, popular machine learning algorithms (i.e. SVMs and XGBoost) were utilized along with SHAP, which is a state-of-the-art methodology for explainable machine learning and has been successfully implemented in real-time safety assessment.

The data utilized in the study were a combination of three sources: (a) naturalistic driving telematics data provided by OSeven Telematics, (b) Mobility trends from Apple demonstrating driving and walking requests in Apple maps, and (c) COVID-19 indicators obtained from a COVID-19 observatory developed in Oxford University. Although the COVID-19 indicators dataset was completely open-source and based on

Fig. 5. SHAP values for the microscopic XGBoost individual model features for HAs.

publicly available data, both the OSeven and Apple data were collected using non-disclosed methodologies. Nevertheless, the data were utilized in this study due to established applications of such datasets in published studies in well-renowned scientific journals (e.g., Katrakazas et al., 2021; Kontaxi et al., 2021b; Papadimitriou et al., 2019; Petraki et al., 2020) and the scarcity of naturalistic driving and traffic flow data during the pandemic period.

The results of the regression models that were developed demonstrated that both macroscopically and microscopically, the frequency of harsh acceleration events can be successfully predicted. Macroscopically, it was demonstrated that 85% of the HA per 100 km rates could be predicted with a ± 10% tolerance, whereas microscopic analysis can predict HA counts with a ± 1 tolerance with a similar percentage. It is important to keep in mind that since the units used as basis of analysis are different, the achievement of good performance for different scales validates the appropriateness of approach: Microscopic analyses can examine data as counts with the appropriate exposure metrics to provide context to the ML algorithms, whereas macroscopic analyses require rate-based approaches, as significant details of harsh event precursors are hidden due to data aggregation. The corresponding RMSE errors were 1.274 and 1.654, respectively, for the macroscopic and microscopic analysis, which may imply that noise in individual trip data could result in higher regression errors for the XGBoost algorithm, as count units are larger than rates.

Delving into the core of the study (i.e., the identification of individual feature contributions in predicting the micro- and macro-scopic level of road safety during the pandemic), it was demonstrated that speed and speeding were the most significant driving behavior factors in identifying HAs, a fact consistent with the literature and the significance of speed for safety–critical driving events. In both microscopic and macroscopic models, higher driving speeds led to more HAs, demonstrating the negative effect of aggressive driving in everyday road safety.

Total cases per million were the COVID-19 factor with the most influence, whereas concerning exposure, it was shown that duration and distance were significantly influential to the frequency of HAs. For both COVID-19 indicators and exposure indicators, higher values corresponded to a reduction in HAs, which seems reasonable. COVID-19 cases per million showed the practical spread of the virus in the community, and therefore it is reasonable to have been heavily influencing driving behavior both directly as depicted in harsh event frequency, as well as indirectly with the reduction in the distance and duration of trips. For example, if COVID-19 cases reached a peak, this implied that stricter response measures were in place and therefore citizens drove only for significant purposes. Nevertheless, it is odd that neither the reproduction rate nor the strictness of the measures were found to be highly contributing features, and thus do not have a significant impact on the frequency of HAs. It is possible that the long period of imposing measures for COVID-19 led the community to a disappointment in the effectiveness of such measures, and therefore driving behavior demonstrated a tendency to pre-pandemic standards, which led to insufficient influence of those indicators in predicting future behavior patterns.

The magnitude of the COVID-19 effects deserves additional discussion. The COVID-19 effects are small in comparison, as the variable of investigation, HA, is ultimately a kinematic variable. Therefore, it is reasonable that more traditional trip-related variables, such as speed, driving duration and distance, appear to influence HA generation more than pandemic-related parameters (depicted with darker colors in Fig. 5). It was considered that removing trip-related variables would eliminate the direct physical foundations of the models and provide an incomplete picture regarding HA predictions. The approach in this study ultimately examines how driver behavior interacts with both triprelated and pandemic factors.

The relationship between high COVID-19 fatalities and HAs was initially expected to be one of the most interesting ones. From the present analysis, the results did not reveal very powerful correlations, but still present ramifications that are worth considering. When examined on a daily basis, COVID-19 fatalities had a negligible effect, which barely registered on the SHAP values. When examined on a trip basis, new COVID-19 fatalities moved upwards moderately, entering the ten variables with the highest influence. The interpretation is that high daily COVID-19 fatalities were sufficient to trigger increased driver aggressiveness, particularly at low fatalities (0–20 daily deaths) and high fatalities (110–120 daily deaths).

The implications are twofold: Firstly, the behavior of some drivers was affected by the COVID-19 fatalities during two times: The first time is when the pandemic first spread, and still had low fatality rates. This effect can be attributed largely to the fear of the unknown both regarding the pandemic in general and the management of the transport network under pandemic conditions in particular. The second time refers to the peak of the pandemic, which translates into the acknowledgement of COVID-19 as a direct threat by the drivers and into an increase of their aggressiveness caused by the related stress. Secondly, this influence is only observed on a fraction of the driver sample, and does not persist across the aggregate population when viewed macroscopically, which means that its manifestation depends on the person.

Nonetheless, the relatively small magnitude of the COVID-19 fatalities compared to other pandemic metrics indicates that driver behavior and aggressiveness is influenced more by cases and reproduction rate. This finding implies that drivers were affected more by the general advent of the pandemic rather than its direct toll on the population.

The 'pure' influences of the pandemic may be challenging to distinguish by mobility patterns for additional reasons. It can be considered that several reactions are taken by drivers in response to outside effects or secondary conditions that might be unobserved in the telematics data. For instance, a case can be considered when close acquaintances of a driver are infected, and the driver limits their overall trips in response, despite macroscopic improvements in the health/ infection rates of the overall population as expressed by pandemic metrics. It is also possible that some drivers have become more conservative as a result of the pandemic due to the fact that they would want to avoid risking hospitalization due to a potential injury crash. Some of these questions might be answered by more in-depth research with longitudinal data of specific drivers, by observing them through the various phases of the pandemic. However, it should be stressed that driver telematics, which constitute the basis and backbone of our analysis, are quite rigorously protected by European data protection laws (i. e., GDPR), which in turn hinder very detailed research.

6.2. Study limitations and further issues

To the best of the knowledge of the authors, this is the first study that provided explainable ML prediction for macroscopic and microscopic road safety evaluations during the COVID-19 pandemic period. Despite the strengths of the analyses, the study is not without drawbacks.

It is very important to note that the different data pillars, collected originally within different scopes by their providers, can only be informative in the broader sense, as many of them do not directly affect vehicle kinematics in the physical sense, constituting a limitation of the present study design. Specifically, data such as HAs, which denote naturalistic event recording, are not organically fused with COVID spread and restriction data in terms of a relationship directly observable in the field. Rather, these data are used as proxy explanatory variables and provide a description of the overarching circumstances within the transport network. The transport network has more nuances as well, such as interrelationships of different road user groups (most notably pedestrians for Athens) with motorized traffic, as well as different levels of familiarity based on travel purpose (e.g., tourist flows). Unfortunately, their explicit consideration leads to an extremely high amount of complexity, which is impossible to include in the present study.

There is heteroscedasticity in the ML results of this study, which might have been inherited from similar heteroscedasticity in inputs. The authors endeavored to mitigate this with an array of test runs, conducted in the initial modeling phase with various transformations, including squared, exponential and logarithmic values of variables, but none of these led to observably better models. In fact, there were several cases where the error metrics such as RMSE yielded extremely high value or erroneous values. More importantly, such transformations are detrimental to the comparison of variable feature importance and variable interpretability and explainability with SHAP values. Therefore, the more straightforward ML models, containing untransformed variables, were retained. Nonetheless, it is possible that there are specific data transformations that would have been useful in mitigating heteroscedasticity, which were not explored in this study.

While using a continuous timespan for the formulation of a test set is not the most traditional approach, the authors believe it has merit. Prediction for future time blocks appears to be a more reasonable and fruitful task for practical applications, instead of randomly determined days that have temporal overlap with the training data. Given that the pandemic period was a period with many fluctuations in everyday life, measures and behavior, with little to no pattern repetition or cyclicity, no time sequence effects are drawn from the algorithms. However, since sequential effects might enhance model quality in datasets with more cyclicity, building such models will be examined in future studies.

Apart from the data/modeling limitations that were discussed previously, the study is limited only on the analyses for HAs for one year of the pandemic and did not consider other surrogate safety indicators such as harsh brakings, speeding, conflicts, or more years of driving data. Different definitions of driver aggressiveness, possibly as a more composite indicator comprising the aforementioned (or other parameters) can always be devised based on data availability as well. Additionally, it is crucial to recognize that telematics data can be categorized based on road environment, including distinctions like highways, rural, and urban trips, as demonstrated by Papadimitriou et al. (2019). Future investigations into the impact of COVID-19 could adopt an infrastructure-centric approach rather than solely focusing on individual trips. By incorporating supplementary data containing geolocation and road type details, the existing methodology could be strengthened. This enhancement could facilitate spatial analyses of the examined variables, aligning with methodologies utilized by Ziakopoulos et al. (2022).

Furthermore, although the developed models have proven to be powerful both in this paper as well as in previous literature, the utilization of deep learning approaches (e.g., Recurrent, Convolutional or Long-Short Term Memory Neural Networks) might have provided noteworthy predictions as well. The comparison of ML to more traditional function-based models, such as linear or generalized linear models, is always interesting, as long as the analysts are mindful of caveats such as multicollinearity effects. Finally, the consideration of other explainable ML approaches such as Local Interpretable Model-agnostic Explanations (LIME) or Global Sensitivity Analysis (GSA) in tandem with SHAP could lead to more insights.

7. Conclusions

The objective of the current study was to quantify the impact of the COVID-19 pandemic on driver behavior and aggressiveness as expressed by harsh accelerations (HAs) measured by smartphone telematics data. To achieve that aim, machine learning algorithms were implemented on two scales: (a) a macroscopic scale, which involves a daily analysis of aggregate driver behavior across the network with Support Vector Machine algorithms (SVMs), and (b) a microscopic scale, which involves a trip-based analysis of driver trips with an Extreme Gradient Boosting (XGBoost) algorithm. SHAP values are used to interpret the outputs of both algorithms, providing a quantification on how pandemic indicators affect driver behavior and aggressiveness.

Macroscopic results (i.e., daily analysis) indicated that high total average speed values reduce HA rates per 100 km, while this trend reverses with high driving speed. High values of factors related to the pandemic, such as Reproduction Rate, Total Cases and Total Cases per million people were found to reduce HA rates per 100 km, while Total Fatalities and Total Fatalities per million people have little to no contribution to the examined HA rates. Microscopic results (i.e. tripbased analysis) indicated that high speeding values are associated with increased HA counts, as is the total trip distance after 100 m and trip duration after 1000 s. Regarding driving speed, drivers perform more HAs between 30–50 km/h, while after 50 km/h, the contributions of speed lead to fewer HAs. The COVID-19 parameters indicate small but noteworthy findings. An increase in Apple mobility requests is correlated with a slight decrease in HAs. Slight decreases in HA are also observed as both new daily COVID-19 cases and total cases per million rise. Moreover, drivers performed more HAs at low (0–20) and high (110–120) daily COVID-19 fatalities, while they were more conservative towards the middle values, while the Stringency Index has an unclear microscopic contribution, indicating that the pandemic measurements were more influential on HA counts compared to the policy measures taken by the state.

The current findings present considerable implications. Regarding the direct core hypothesis of the study, it can be stated that the COVID-19 pandemic, along with the corresponding government responses, did have measurable impacts on driver behavior. The pandemic metrics regarding the reproduction and infection rates of the disease affected driver behavior much more than actual COVID-19 related fatalities. This finding suggests that the broader public health context had a greater effect on driving behavior than the direct toll of the pandemic. Additionally, when examining HAs, which are kinematic effects, the overall pandemic influence is secondary compared to other kinematic and triprelated variables.

Therefore, the study also highlights how variables from different domains can influence behavior at multiple scales within the transport network. While pandemic-related factors were not the most dominant in terms of feature importance, valuable insights were still obtained. As a related consequence, conceptually disentangling the influence and effects of the pandemic is a persistent challenge. This challenge is especially pronounced for empirical fields such as road safety, as the pandemic permeates all elements of the transport network, from the global administration to the individual decisions and actions.

The broader contributions of this research are multifaceted. In terms of study design, the study demonstrates how diverse datasets from different scales and sources can be integrated to explore complex aspects of the transport network,such as how driver behavior, reflected by harsh acceleration, is influenced. Methodologically, the study achieved accuracy exceeding 85% in the test subsets both macroscopically and microscopically, emphasizing the value of non-linear Machine Learning algorithms for analyzing intricate phenomena. Moreover, this result was achieved by utilizing a continuous time period as test set for the macroscopic analysis, indicating the utility of the adopted approach when developing predictive tools, as authorities would be able to obtain accurate predictions on behavior shifts in times when an event comparative to the COVID-19 pandemic is unfolding.

CRediT authorship contribution statement

Apostolos Ziakopoulos: Writing – review & editing, Writing – original draft, Visualization, Software, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Marios Sekadakis: Writing – review & editing, Writing – original draft, Validation, Software, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Christos Katrakazas: Writing – original draft, Investigation, Formal analysis, Data curation. Marianthi Kallidoni: Investigation, Formal analysis, Data curation. Eva Michelaraki: Writing – original draft, Investigation, Formal analysis, Data curation. George Yannis: Supervision, Funding acquisition.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

The authors would like to thank OSeven Telematics, London, UK, for providing all necessary data exploited to accomplish this study.

References

Aletta, F., Brinchi, S., Carrese, S., Gemma, A., Guattari, C., Mannini, L., & Patella, S. M. (2020). Analysing urban traffic volumes and mapping noise emissions in Rome (Italy) in the context of containment measures for the COVID-19 disease. Noise Mapping, 7(1), 114–122.
Anke, J., Francke, A., Schaefer, L. M., & Petzoldt, T. (2021). Impact of SARS-CoV-2 on the mobility behaviour in Germany. European Transport Research Review, 13(1), 1–13.
Anwari, N., Ahmed, M. T., Islam, M. R., Hadiuzzaman, M., & Amin, S. (2021). Exploring the travel behavior changes caused by the COVID-19 crisis: A case study for a developing country. Transportation Research Interdisciplinary Perspectives, 9, Article 100334.
Apple,. (2020). COVID‑19 Mobility Trends Reports Apple [WWW Document] Accessed 11/06/20. https://www.apple.com/covid19/mobility
Arroyo, F., Bullano, F., Kucinskas, S.-M., & C.. (2022). Tracking R of COVID-19 A New Real-Time Estimation Using the Kalman Filter. PLoS ONE, 16(1), Article E0244474. https://doi.org/10.1371/journal.pone.0244474.
Bischl, B., Lang, M., Kotthoff, L., Schiffner, J., Richter, J., Studerus, E., Casalicchio, G., Jones, Z. M., & 818,. (2016). mlr: Machine Learning in R. The Journal of Machine Learning Research, 17(1), 5938–5942.
Bucsky, P. (2020). Modal share changes due to COVID-19: The case of Budapest. Transportation Research Interdisciplinary Perspectives, 8, Article 100141.
Carter, D. (2020). Effects of COVID-19 Shutdown on Crashes and Travel in NC. Chan, A. (2020). Saanich sees spike in speeders as roads empty due to COVID-19. CTV
News. https://vancouverisland.ctvnews.ca/saanichsees-spike-in-speeders-as-roads-e mpty-due-tocovid-19-1.4886785. Chang, I., Park, H., Hong, E., Lee, J., & Kwon, N. (2022). Predicting effects of built
environment on fatal pedestrian accidents at location-specific level: Application of XGBoost and SHAP. Accident Analysis and Prevention, 166(August 2021), 106545. https://doi.org/10.1016/j.aap.2021.106545.
Chen, T., He, T., Benesty, M., & Khotilovich, V. (2021). Package 'xgboost'. R, version, Article 1.4.1.1.
Chen, T., & Guestrin, C. (2016). Xgboost: A scalable tree boosting system. In In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining (pp. 785–794).
City of Toronto. (2020). City of Toronto Urges Drivers to Obey Rules of the Road. Toronto: Ontario.
Gupta, M., Pawar, N. M., & Velaga, N. R. (2021). Impact of lockdown and change in mobility patterns on road fatalities during COVID-19 pandemic. Transportation Letters, 1–14.
Hale, T., Anania, J., Angrist, N., Boby, T., Cameron-Blake, E., Folco, M. Di, Ellen, L., Goldszmidt, R., Hallas, L., KIra, B., Luciano, M., Majumadar, S., Nagesh, R., Petherick, A., Phillips, T., Tatlow, H., Webster, S., Wood, A., & Zhang, Y. (2021). Variation in government responses to COVID-19, Version 12.0. In BSG Working Paper Series. https://covidtracker.bsg.ox.ac.uk/stringency-scatter.
Hale, T., Angrist, N., Cameron-Blake, E., Hallas, L., Kira, B., Majumdar, S., Petherick, A., Phillips, T., Tatlow, H., & Webster, S. (2020). Variation in government responses to COVID-19. Version 8.0: University of Oxford.
Heidenreich, P. (2020). Recent uptick in speeding on Edmonton roads concerns mayor and city official. Global News. https://globalnews.ca/news/6739220/edmonton-dr ivers-speedingcoronavirus-covid-19-iveson/.
Hughes, J.E., Kaffine, D., Kaffine, L. (2021). Decline in traffic congestion increased accident severity in the wake of COVID-19.
Jaekel, B., & Muley, D. (2022). Transport impacts in Germany and State of Qatar: An assessment during the first wave of COVID-19. Transportation research interdisciplinary perspectives, 13, Article 100540.
Katrakazas, C., Michelaraki, E., Sekadakis, M., & Yannis, G. (2020). A descriptive analysis of the effect of the COVID-19 pandemic on driving behavior and road safety. Transportation research interdisciplinary perspectives, 7, Article 100186.
Katrakazas, C., Michelaraki, E., Sekadakis, M., Ziakopoulos, A., Kontaxi, A., & Yannis, G. (2021). Identifying the impact of the COVID-19 pandemic on driving behavior using naturalistic driving data and time series forecasting. Journal of Safety Research.
Kontaxi, A., Ziakopoulos, A., & Yannis, G. (2021). Trip characteristics impact on the frequency of harsh events recorded via smartphone sensors. IATSS Research.
Kontaxi, A., Ziakopoulos, A., & Yannis, G. (2021). Investigation of the speeding behavior of motorcyclists through an innovative smartphone application. Traffic injury prevention, 22(6), 460–466.
Liu, Y., & Just, A. (2021). SHAPforxgboost: SHAP Plots for 'XGBoost'. Available online: R package version, (1), 1 Https://cran.r-project.org/web/packages/SHAPforxgboost
Lundberg, S. M., & Lee, S. I. (2017). A unified approach to interpreting model predictions. In In Proceedings of the 31st international conference on neural information processing systems (pp. 4768–4777).
Michelaraki, E., Sekadakis, M., Katrakazas, C., Ziakopoulos, A., & Yannis, G. (2021). A four-country comparative overview of the impact of COVID-19 on traffic safety behavior. In Proceedings of 10th International Congress on Transportation Research (ICTR).
Molnar, C. (2020). Interpretable machine learning. E-book version (2023) available online: https://christophm.github.io/interpretable-ml-book/. Lastly accessed 10-01- 2024.
Muley, D., Ghanim, M. S., Mohammad, A., & Kharbeche, M. (2021). Quantifying the impact of COVID–19 preventive measures on traffic in the State of Qatar. Transport Policy, 103, 45–59.
Nielsen, D. (2016). Tree boosting with xgboost-why does xgboost win" every" machine learning competition? (Master's thesis, NTNU).
Our World in Data (OWD). (2020). Coronavirus Pandemic (COVID-19). https:// ourworldindata.org/coronavirus.
Papadimitriou, E., Argyropoulou, A., Tselentis, D. I., & Yannis, G. (2019). Analysis of driver behaviour through smartphone data: The case of mobile phone use while driving. Safety Science, 119(May), 91–97. https://doi.org/10.1016/j. ssci.2019.05.059
Parsa, A. B., Movahedi, A., Taghipour, H., Derrible, S., & Mohammadian, A. K. (2020). Toward safer highways, application of XGBoost and SHAP for real-time accident detection and feature analysis. Accident Analysis & Prevention, 136, Article 105405.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., & Duchesnay, E. (2011). Scikit-learn: Machine learning in Python. the Journal of machine Learning research, 12, 2825–2830.
Petraki, V., Ziakopoulos, A., & Yannis, G. (2020). Combined impact of road and traffic characteristic on driver behavior using smartphone sensor data. Accident Analysis & Prevention, 144(November 2019), 105657. https://doi.org/10.1016/j. aap.2020.105657.
R Core Team, 2019. R: A Language and Environment for Statistical Computing.
Roe, C. M., Rosnick, C. B., Colletta, A., & Babulal, G. M. (2021). Reaction to a pandemic: Social distancing and driving among older adults during COVID-19. Journal of Applied Gerontology, 40(3), 263–267.
Rosenman, R., Tennekoon, V., & Hill, L. G. (2011). Measuring bias in self-reported data. International Journal of Behavioural and Healthcare Research, 2(4), 320–332.
Saladi´e, O., ` Bustamante, E., & Guti´errez, A. (2020). COVID-19 lockdown and reduction of traffic accidents in Tarragona province. Spain. Transportation research interdisciplinary perspectives, 8, Article 100218.
Sekadakis, M., Katrakazas, C., Michelaraki, E., Kehagia, F., & Yannis, G. (2021). Analysis of the impact of COVID-19 on collisions, fatalities and injuries using time series forecasting: The case of Greece. Accident Analysis & Prevention, 162, Article 106391.
Shapley, L. (1953) 'A value fo n-person Games', Ann. Math. Study28, Contributions to the Theory of Games, ed. by HW Kuhn, and AW Tucker, pp. 307–317.
Sharifi, A., & Khavarian-Garmsir, A. R. (2020). The COVID-19 pandemic: Impacts on cities and major lessons for urban planning, design, and management. Science of the Total Environment, 749, Article 142391.
Stavrinos, D., McManus, B., Mrug, S., He, H., Gresham, B., Albright, M. G., & White, D. M. (2020). Adolescent driving behavior before and during restrictions related to COVID-19. Accident Analysis & Prevention, 144, Article 105686.
Tweedie, M. C. (1984). An index which distinguishes between some important exponential families. In Statistics: Applications and new directions: Proc. Indian statistical institute golden Jubilee International conference (Vol. 579, pp. 579-604).
Vanlaar, W. G. M., Woods-Fry, H., Barrett, H., Lyon, C., Brown, S., Wicklund, C., & Robertson, R. D. (2021). The impact of COVID-19 on road safety in Canada and the United States. Accident Analysis & Prevention, 160, Article 106324.
Vingilis, E., Beirness, D., Boase, P., Byrne, P., Johnson, J., Jonah, B., & Wiesenthal, D. L. (2020). Coronavirus disease 2019: What could be the effects on Road safety? Accident Analysis & Prevention, 144, Article 105687.
Wagner, E., Atkins, R. G., Berning, A., Robbins, A., Watson, C., & Anderle, J. (2020). Examination of the traffic safety environment during the second quarter of 2020: Special report (No. DOT HS 813 011). United States. National Highway Traffic Safety Administration. Office of Behavioral Safety Research.
Wegman, F., & Katrakazas, C. (2021). Did the COVID-19 pandemic influence traffic fatalities in 2020? A presentation of first findings. IATSS Research, 45(4), 469–484. https://doi.org/10.1016/j.iatssr.2021.11.005
Wen, X., Xie, Y., Jiang, L., Li, Y., & Ge, T. (2022). On the interpretability of machine learning methods in crash frequency modeling and crash modification factor development. Accident Analysis and Prevention, 168(February), Article 106617. https://doi.org/10.1016/j.aap.2022.106617
Wen, X., Xie, Y., Wu, L., & Jiang, L. (2021). Quantifying and comparing the effects of key risk factors on various types of roadway segment crashes with LightGBM and SHAP. Accident Analysis and Prevention, 159(June), Article 106261. https://doi.org/ 10.1016/j.aap.2021.106261
XGBoost developer team (2021). XGBoost Documentation. Official website. Available from: https://xgboost.readthedocs.io/en/latest/index.html [Accessed 07/10/2021].
Yang, C., Chen, M., & Yuan, Q. (2021). The application of XGBoost and SHAP to examining the factors in freight truck-related crashes: An exploratory analysis. Accident Analysis & Prevention, 158, Article 106153.
Yu, H., & Kim, S. (2012). SVM Tutorial: Classification, Regression, and Ranking. Handbook of Natural Computing.
Yuan, C., Li, Y., Huang, H., Wang, S., Sun, Z., & Wang, H. (2022). Application of explainable machine learning for real-time safety analysis toward a connected vehicle environment. Accident Analysis & Prevention, 171, Article 106681.
Zhu, N., Zhang, D., Wang, W., Li, X., Yang, B., Song, J., ... & Tan, W. (2020). A novel coronavirus from patients with pneumonia in China, 2019. New England journal of medicine.
Ziakopoulos, A., Kontaxi, A., & Yannis, G. (2023). Analysis of mobile phone use engagement during naturalistic driving through explainable imbalanced machine learning. Accident Analysis & Prevention, 181, Article 106936.
Ziakopoulos, A., Vlahogianni, E., Antoniou, C., & Yannis, G. (2022). Spatial predictions of harsh driving events using statistical and machine learning methods. Safety science, 150, Article 105722.
Ziakopoulos, A., Kontaxi, A., Yannis, G., Fortsakis, P., Kontonasios, K. N., & Kostoulas, G. (2020). Advanced driver monitoring using smartphone applications: The BeSmart project. In Proceedings of 8th Transport Research Arena (TRA).
Dr. Apostolos Ziakopoulos, PhD, MSc DIC is a Research Associate at the Department of Transportation Planning and Engineering at the School of Civil Engineering of the National Technical University of Athens (NTUA). He holds a Civil Engineering Diploma from NTUA majoring in Transportation Engineering (2013). He also holds a Master of Science – DIC in Transport from Imperial College London and University College London (2014). His Ph.D. dissertation was in spatial analysis of road safety and traffic behavior from NTUA (2020). His main research interests involve road crash analyses, statistical and spatial modelling and traffic engineering. He has participated in 15 research and engineering projects and studies and has published 90 papers in scientific journals and conferences.
Mr. Marios Sekadakis is a Civil-Transportation Engineer, PhD Candidate and Research Associate at the Department of Transportation Planning and Engineering at the School of Civil Engineering of the National Technical University of Athens (NTUA). In 2020, he graduated in Civil Engineering from the National Technical University of Athens (NTUA) majoring in Transportation Engineering. He has more than 4 years of experience in several aspects of traffic engineering and road safety. He has participated in 2 research projects and has published 21 scientific papers.
Dr Christos Katrakazas is a Civil Engineer, PhD, and a Research Associate at the Department of Transportation Planning and Engineering at the School of Civil Engineering
of the National Technical University of Athens (NTUA). He holds a Civil Engineering Diploma from NTUA majoring in Transportation Engineering (2013) and a Ph.D. in safety of autonomous vehicles from Loughborough University (2017). After finishing his doctoral studies, he worked for 8 months as a Research Associate at Loughborough University, and for two years at the Chair of Transportation Systems Engineering at the Technical University of Munich (TUM), Germany. He has more than 9 years of experience in several aspects of transport safety and engineering.
Ms. Marianthi Kallidoni is a Civil-Transportation Engineer and PhD student at the Technical University of Denmark (DTU). Since 2021, she holds a Civil Engineering Diploma from the National Technical University of Athens (NTUA) majoring in Transportation Engineering. She has more than 2 years of experience in several aspects of transport engineering and behavioural modelling. She has participated in 7 research projects and has published eight (8) scientific papers.
Ms. Eva Michelaraki is a Civil Engineer, PhD Candidate and Researcher at the Department of Transportation Planning and Engineering at the School of Civil Engineering of the National Technical University of Athens (NTUA). Since 2019, she holds a Civil Engineering Diploma from NTUA majoring in Transportation Engineering. Her PhD research is on ''Road accident risk factors and big data". She has more than 3 years of experience in several aspects of traffic engineering and road safety. She has participated in 3 research projects and has published 14 scientific papers.
Prof. George Yannis is a Professor in Traffic Safety and Management with particular focus on data management and analysis and currently Director of the Department of Transportation Planning and Engineering of the School of Civil Engineering at the NTUA. He has contributed extensively in more than 245 research and engineering projects and studies and in several scientific committees of the European Commission and other International Organizations (UN/ECE, OECD, WHO, World Bank, EIB, CEDR, ERF, UITP, ETSC, ECTRI, WCTR, TRB). He has published more than 750 scientific papers (226 in scientific journals) widely cited worldwide.

ziakopoulosexplainablemacroscopicmicroscopicinfluences2025.md 116 KB Historia Czysty