From historical patterns to hourly dynamics: interpretable deep learning for cycling estimation

View at publisher Summarise with VIRAT
Authors: Yuxin Wang, Kim N. Dirks, Guohua Song, Lei Yu, Yizheng Wu
Cited by: 0
References: 48

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Accurate hourly cycling volume estimation supports short-term infrastructure management and demand-responsive policy design. While existing estimation models focus on aggregated historical patterns, this study proposes an interpretable deep learning framework for estimating hourly cycling volumes using historical, temporal, weather, and spatial features. A multi-branch residual network, guided by counter-specific feature dependency types derived from eXtreme Gradient Boosting (XGBoost) and K-means clustering, captures nonlinear and spatially heterogeneous patterns. Two years of data from 39 permanent counters in Auckland, New Zealand (2022–2024) are used for validation. The proposed framework outperforms traditional machine learning models in predictive accuracy (MAE = 4.00 cyclists/h) and explanatory power (R2 = 0.925). SHapley Additive exPlanations (SHAP) and spatial feature ablation reveal threshold and interaction effects from features. Spatial extrapolations also demonstrate robust transferability (R2 = 0.903 with warm-up, 0.839 without). Embedding feature impact mechanisms effectively improves predictive performance, offering insights into data-driven cycling policy and infrastructure management.

Fields of Study & Subjects

Urban Transport and Accessibility Traffic Prediction and Management Techniques Transportation Planning and Optimization

Publication details

Journal / source
Transportation Letters
Volume
18
Issue
9
Pages
1943–1968
Publication date
26 March 2026
Language
EN