Loading...
Recall informed self-attention representation learning and generative AI for improved car accident detection in highly unbalanced datasets
Watters, Jacob Daniel
Watters, Jacob Daniel
Citations
Altmetric:
Advisor
Editor
Date
Date Issued
2024
Date Submitted
Collections
Files
Research Projects
Organizational Units
Journal Issue
Embargo Expires
Abstract
This thesis explores the application of self-attention mechanisms in conjunction with recall-informed loss functions and generative models to the domain of traffic safety analysis and car accident predictions. Utilizing a comprehensive dataset compiled from the Colorado Department of Transportation, supplemented with historical weather data, this study introduces a novel approach that leverages the dynamic and complex nature of traffic data while focusing on rare but critical events such as car accidents. We present a self-representation learning model to learn latent representations of traffic data that effectively capture the time-series sequence and periodicity nuances of vehicle traffic. This model works in conjunction with two novel loss functions that prioritize recall over accuracy. The incorporation of these recall-informed loss functions into the classification task addresses the problem of class imbalance which is prevalent in car accident detection tasks. This enables car accidents, which are uncommon but critical events to be more effectively detected when combined with self-representation learning methods. Further, this study explores the use of self-attention in generative models and its efficacy for utilization in data balancing to increase successful detection of uncommon and critical events. The experiments presented in this paper contribute to the advancement of machine learning in the field of traffic safety and demonstrate that the combination of self-representation learning and recall-informed loss functions have the potential to improve the performance of car accident and traffic anomaly models in terms of ``critical-event'' detection.
Associated Publications
Rights
Copyright of the original work is retained by the author.