Study on Multi-Modal Shift to Dhaka Metro: Application of Machine Learning and Explainable AI in Transportation

This study focuses on Dhaka Metro in Bangladesh, using machine learning and explainable AI technologies to analyze the behavioral patterns of passengers shifting from other transportation modes to the metro. The research aims to understand the key factors driving this shift and provide a basis for transportation planning. The original author is tasnooba, and the source is the GitHub repository (Assessing_multi_modal_shift). The associated paper title is "Assessing Multi Modal Shift to Metro using machine learning and explainable AI: A case study of Dhaka Metro rail".

As the capital of Bangladesh, Dhaka faces challenges such as extreme congestion (ranking among the most congested cities globally), deteriorating air quality, insufficient public transport, and fragmented travel modes. The launch of Dhaka Metro is a key solution to alleviate these issues, but its success depends on whether it can attract passengers to shift. Understanding shift behavior is crucial for line optimization, service improvement, policy formulation, and environmental assessment.

Traditional traffic demand forecasting methods (e.g., discrete choice models) have limitations such as assumption constraints, difficulty capturing non-linearity, and insufficient ability to process high-dimensional data. Machine learning methods (random forest, gradient boosting, neural networks, etc.) can automatically learn complex relationships and handle high-dimensional data. Explainable AI (e.g., SHAP, LIME, feature importance analysis, partial dependence plots) solves the "black box" problem of ML models, providing a basis for policy making, enhancing public acceptance, and improving model credibility.

Data sources include on-site passenger surveys (socio-economic characteristics, travel features, metro usage, etc.) and GIS data (distance from residence/workplace to metro stations, bus network coverage, etc.). Key features are categorized into accessibility (walking distance, transfer convenience), time (travel time comparison, waiting time), cost (fare, cost comparison), service quality (comfort, punctuality), socio-economic (income, private car ownership), and psychological factors (new technology acceptance, environmental awareness).

The study considers multiple ML models: logistic regression (baseline), random forest, gradient boosting trees (XGBoost/LightGBM), support vector machines, and neural networks. Evaluation metrics include accuracy, precision/recall/F1, ROC-AUC, and confusion matrix. Stratified K-fold cross-validation is used to ensure robust results, with a focus on regional, population differences, and temporal stability.

Based on similar studies, key influencing factors may include distance to the metro station (most important), time savings, income level, and travel purpose (commuting is more likely to shift). SHAP analysis may reveal non-linear effects (e.g., distance thresholds), interaction effects (income and fare), and individual differences. Policy recommendations include first/last mile solutions (feeder buses, bicycle facilities), fare strategies (time-based fares, subsidies for low-income groups), and service improvements (peak-hour frequency, transfer experience).

Contributions: Introduce explainable AI into traffic behavior research, bridging the gap between prediction accuracy and interpretability; fill the literature gap for developing countries. Limitations: Cross-sectional data makes it difficult to establish causal relationships, self-reported data bias, sample representativeness issues; models may overfit local patterns. Future directions: Longitudinal tracking, dynamic models, multi-modal integration, and expansion to other cities.