Scholarlink

This paper presents a machine learning approach to predict ride-hailing demand based on weather conditions. We conduct an in-depth Exploratory Data Analysis (EDA) to identify and handle outliers, refine features through correlation analysis, and optimize the dataset of 10,886 observations with 11 column variables (excluding the datetime column—which has been used as an index) with feature engineering. Various machine learning models, from simpler Linear Regression and Regularization techniques (Ridge, Lasso) to complex ensemble methods (Random Forest, Gradient Boost, AdaBoost), are evaluated to improve prediction accuracy. We test different model structures, including unified and separate models for working and non-working days, and experiment with feature encoding methods. Additionally, we implement a stacking technique where predictions from individual models serve as inputs to a second-level model, enhancing accuracy through ensemble learning. This comprehensive approach yields a robust model capable of accurately predicting ride-hailing demand based on varying weather conditions and time periods. The Random Forest model resulted in the best Root Mean Squared Logarithmic Error (RMSLE) of 0.43, with humidity and temperature as the most impactful features indicated by Shapley Additive exPlanations (SHAP) analysis. These results underscore the project’s potential to revolutionize urban mobility by contributing to smarter, more sustainable transportation systems.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimizing Ride-Hailing Demand Forecasting with Explainable AI and Ensemble Learning Techniques

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Vedanti Patil,
  • Manas Choudhary,
  • T. V. Sumithra

摘要

This paper presents a machine learning approach to predict ride-hailing demand based on weather conditions. We conduct an in-depth Exploratory Data Analysis (EDA) to identify and handle outliers, refine features through correlation analysis, and optimize the dataset of 10,886 observations with 11 column variables (excluding the datetime column—which has been used as an index) with feature engineering. Various machine learning models, from simpler Linear Regression and Regularization techniques (Ridge, Lasso) to complex ensemble methods (Random Forest, Gradient Boost, AdaBoost), are evaluated to improve prediction accuracy. We test different model structures, including unified and separate models for working and non-working days, and experiment with feature encoding methods. Additionally, we implement a stacking technique where predictions from individual models serve as inputs to a second-level model, enhancing accuracy through ensemble learning. This comprehensive approach yields a robust model capable of accurately predicting ride-hailing demand based on varying weather conditions and time periods. The Random Forest model resulted in the best Root Mean Squared Logarithmic Error (RMSLE) of 0.43, with humidity and temperature as the most impactful features indicated by Shapley Additive exPlanations (SHAP) analysis. These results underscore the project’s potential to revolutionize urban mobility by contributing to smarter, more sustainable transportation systems.