A machine learning tool for collecting and analyzing subjective road safety data from Twitter

Several measures have been proposed in the road safety field to measure safety objectively. However, studies focused on measuring subjective road safety (i.e., the cognitive safety of road environment by drivers) are limited due to the lack and difficulty of collecting subjective safety data. Still, subjective safety is important and should be monitored as much as objective safety as it can negatively impact road users' mobility and road agencies draw information on it in determining policies (ultimately affecting collision occurrence). Nowadays, crowdsourcing big data related to subjective road safety can be done in social media platforms like Twitter. Therefore, this research aims at developing a tool for extracting, classifying, and studying drivers' affective states from road safety-related tweets using keyword filtering, geo-boundaries, natural language processing, and machine-learning (ML) classification. The ML classification algorithms used in this study were naive Bayes (NB), logis-tic regression (LR), support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGBoost). Also, word count, word level TF-IDF, N-gram level TF-IDF, and character level TF-IDF were used as features. Metro Vancouver (Canada) was selected as the geographic region for tweet extraction and creation of training and test data sets. The study focused on the year 2019, where 13,226 unique tweets were extracted after removing duplicates. The performance of proposed ML models was compared by estimating accuracy, precision, recall, and F1-scores. The results showed that the trained RF model with count vector, and SVM classifier with word-level TF-IDF performed best in separating road safety-related from unrelated tweets (accuracy = 0.935, F1-score = 0.937) and determining the proposed classification tags (accuracy = 0.881, F1-score = 0.879), respectively. Finally, sentiment analysis was conducted to investigate the polarity of tweets in each group of the proposed classification to better understand drivers' affective states.