[Paper Review] DxNAT - Deep Neural Networks for Explaining Non-Recurring Traffic Congestion

Fangzhou Sun, Dubey Abhishek+1 more|arXiv (Cornell University)|Jan 30, 2018

Traffic Prediction and Management Techniques3 citations

TL;DR

This paper proposes DxNAT, a deep learning model that converts regional traffic speed data into scaled images and fuses them with temporal and event data to detect non-recurring traffic congestion (NRC) caused by events like football games, hockey games, and accidents. Using a custom crossover-based data augmentation technique and a convolutional neural network, DxNAT achieves 98.73% accuracy in identifying NRC from football games, outperforming traditional methods like random forests.

ABSTRACT

Non-recurring traffic congestion is caused by temporary disruptions, such as accidents, sports games, adverse weather, etc. We use data related to real-time traffic speed, jam factors (a traffic congestion indicator), and events collected over a year from Nashville, TN to train a multi-layered deep neural network. The traffic dataset contains over 900 million data records. The network is thereafter used to classify the real-time data and identify anomalous operations. Compared with traditional approaches of using statistical or machine learning techniques, our model reaches an accuracy of 98.73 percent when identifying traffic congestion caused by football games. Our approach first encodes the traffic across a region as a scaled image. After that the image data from different timestamps is fused with event- and time-related data. Then a crossover operator is used as a data augmentation method to generate training datasets with more balanced classes. Finally, we use the receiver operating characteristic (ROC) analysis to tune the sensitivity of the classifier. We present the analysis of the training time and the inference time separately.

Motivation & Objective

• To address the challenge of detecting non-recurring traffic congestion (NRC) in large urban areas using real-time traffic and event data.
• To overcome limitations of traditional methods that lose spatial information by using feature vectors and suffer from class imbalance in training data.
• To develop a deep learning model capable of identifying NRC caused by diverse events such as sports games and traffic incidents.
• To improve detection accuracy and reduce false positive and false negative rates compared to baseline models like random forests.

Proposed method

• Traffic data in TMC format is converted into multi-dimensional images to preserve spatial relationships across road segments.
• Temporal features (time of day, day of week) and event data are fused with image representations to form multi-modal input.
• A crossover operator is designed to generate synthetic training samples, balancing underrepresented classes like hockey games and accidents.
• A convolutional neural network (CNN) is trained to classify traffic patterns as normal or anomalous, with anomaly detection focused on event-induced congestion.
• Receiver operating characteristic (ROC) analysis is used to tune the classifier’s sensitivity for optimal performance.
• The model is evaluated across three scenarios: football games, hockey games, and traffic incidents, using real-world data from Nashville, TN.

Experimental results

Research questions

• RQ1 Can deep neural networks effectively detect non-recurring traffic congestion caused by large-scale events like football games using spatially encoded traffic data?
• RQ2 How effective is a custom crossover-based data augmentation technique in improving model performance on imbalanced datasets of traffic events?
• RQ3 To what extent can a CNN model trained on image-converted traffic data outperform traditional models like random forests in detecting event-induced congestion?
• RQ4 How does the model’s performance vary across different types of events with varying impact magnitudes, such as football vs. hockey games?
• RQ5 Can the model generalize to detect congestion from rare events like traffic accidents, despite limited training data?

Key findings

• DxNAT achieved 98.73% accuracy in detecting non-recurring congestion caused by football games, significantly outperforming the random forest baseline (84.06% accuracy).
• For hockey games, which have lower attendance and smaller impact, DxNAT achieved 90.76% accuracy, with a higher false negative rate (23.19%) due to the subtler congestion patterns.
• In detecting congestion from traffic accidents, DxNAT achieved 86.59% accuracy, with a false positive rate of 13.71% and a false negative rate of 4.44%.
• The model’s performance was consistently better than the random forest model across all scenarios, with lower false positive and false negative rates.
• The use of image-based representation and crossover-based data augmentation significantly improved model robustness and class balance handling.
• Training time and inference time were measured separately, indicating the model’s feasibility for real-time deployment in smart city applications.