Practical Application of Interpretable Machine Learning in Public Transport Passenger Flow Prediction

This article introduces a station-level public transport passenger flow prediction project. The project combines Random Forest and XGBoost algorithms, and uses interpretable tools such as SHAP and PDP, as well as a fairness audit mechanism, to ensure that model decisions are transparent and fair to all operational groups. It aims to build a responsible AI system to support operational decision-making.

In modern urban public transport systems, traditional prediction methods rely on empirical rules or simple statistical models, which struggle to capture complex non-linear relationships. Meanwhile, most machine learning models are "black boxes"—operation staff cannot understand the prediction logic, which may hide bias issues and reduce the credibility of decisions.

The project aims for station-level passenger flow prediction, emphasizing interpretability and fairness. The dataset includes station information (ID, historical average passenger flow), time features (month, day of the week, whether it is a weekend), shift features (morning/evening shift), weather conditions, etc. After cleaning and preprocessing, it is used for model training.

A dual-model strategy is adopted: Random Forest (handles non-linear interactions and provides feature importance) and XGBoost (gradient boosting tree with regularization to prevent overfitting). Hyperparameter tuning and cross-validation are used to ensure the model's generalization ability.

SHAP (feature contribution value) is used to answer the reasons for predictions, PDP (marginal effect) to show the trend of features on results, and ICE (individual conditional expectation) to identify the heterogeneity of feature impacts—helping operation staff understand and trust the model.

Through group fairness analysis (performance differences across different stations/regions), bias detection (inappropriate reliance on socioeconomic factors), and limitation documentation (clarifying applicable scope), the model is ensured not to treat specific groups unfairly.

The prediction results support staff allocation optimization, resource deployment, operational planning, and emergency response. The project demonstrates a responsible AI construction framework, and the end-to-end methodology from data preparation to fairness audit is worth promoting.