Impact of Dataset Size & Data Source on Aviation Safety Incident Prediction Models with Natural Language Processing

This study investigates the influence of dataset size on machine learning model performance in aviation safety incident prediction. We compare two datasets: the Aviation Safety Network (ASN) dataset and the National Transport Safety Board (NTSB) dataset. These datasets cover the years 2000 to 2020. Employing neural network architectures (LSTM, GRU, CNN, and RNN) and with advanced natural language processing techniques (NLP), we process textual narratives. Our findings underscore the role of dataset size in shaping predictive model performance. Larger datasets consistently yield higher predictive accuracy. Notably, despite having the same number of instances, the NTSB dataset outperforms the ASN dataset. Additionally, CNN stands out with better performance across various evaluation metrics. Furthermore, we examine the performance of these models on datasets of different sizes, affirming that larger datasets lead to better model performance, further highlighting the significance of dataset size in aviation safety prediction. This research emphasizes dataset size as a determinant of predictive model efficacy in aviation safety and suggests improvements in the data collection process for the ASN dataset. These insights resonate within the industry, underlining the need for substantial data collection efforts and the potential benefits of advanced NLP techniques. This study contributes to enhancing safety protocols and guides the broader machine-learning community toward boundless possibilities at the intersection of data, technology, and innovation.