Predicting road accident risks using web data: A classification approach

With the rapidly increasing rates of vehicle usage during recent decades, road accidents have become a significant concern, posing not only risks of injuries but also ranking among the leading causes of fatalities for young and middle-aged individuals. Several factors influence the occurrence of accidents, including careless driving, atmospheric conditions, speeding, and driving under the influence. Understanding the circumstances that lead to a greater risk of road accidents is very helpful for their prevention. The primary goal of this work is to explore patterns in road accidents that have occurred within the state of Albania based on web data scraped from news portals and reports from governmental institutions. The data mining pipeline first involves an intensive data preprocessing phase where scraping, filtering, and refining techniques are employed. Subsequently, several classification models are built on the preprocessed data. These models are developed using various methodologies, including naïve Bayes, random forests, XGBoost, and LightGBM. The constructed classification models are evaluated based on training-test splitting of the preprocessed data using various performance measures. Finally, these models can be used to predict the likelihood of accidents based on certain circumstances.