Replication Data for: Spatial Platoon Context vs. Architectural Complexity: Predicting Headway Degradation in High-Frequency Bus Networks

Description

This repository contains the dataset, evaluation code, and result outputs associated with the study "Spatial Platoon Context vs. Architectural Complexity: Predicting Headway Degradation in High-Frequency Bus Networks." The core dataset encompasses over five months (October 2025 to March 2026) of continuous, high-resolution General Transit Feed Specification Realtime (GTFS-RT) telemetry from 30 diverse high-frequency bus routes in Melbourne, Australia. Unlike traditional transit datasets that treat vehicles in isolation, this data has been heavily engineered to include a Platoon Spatial Matrix. Every row captures synchronous 5-minute snapshots of the physical kinematic state of an interacting transit platoon. The target variable is defined as Kinematic Convergence (Δ Gap): the physical change in the spatial gap between consecutive buses exactly 15 minutes into the future. The repository provides the exact Python framework used to evaluate whether predicting this headway degradation requires deep historical sequence memory (Long Short-Term Memory, Multi-Headed Transformers) or if a simple instantaneous snapshot of the platoon (Deep Neural Networks, Linear Regression) is sufficient. Keywords (Tags) Bus Bunching; GTFS-Realtime; Headway Prediction; Deep Learning; Spatial Interaction; Public Transport; Transit Telemetry

Steps to reproduce

To replicate the findings, feature importance extractions, and model evaluation charts presented in the associated manuscript: 1. Download both academic_model_evaluation_v2.py and bus_training_dataset_v2.zip. 2. Extract the dataset so that bus_training_dataset_v2.csv is located in the same root directory as the Python script. 3. Ensure your Python environment has the following dependencies installed: pandas, numpy, scikit-learn, tensorflow, matplotlib, and seaborn. (Note: The script automatically detects and accelerates training on Apple Silicon MPS if available, but standard CPU/GPU works universally). 4. Run the script: python academic_model_evaluation_v2.py. 5. The script will automatically filter the dataset for active topographical interaction zones using a route-specific 95th-percentile dynamic threshold, train the four architectural models (Linear Regression, Spatial DNN, LSTM, and Transformer) chronologically, and output the comparative evaluation metrics (MAE) alongside multi-route visualizations and feature importance graphs to your local directory.

Institutions

• Monash University

  Victoria, Melbourne

Categories

Licence