Ride-Hailing Service Pattern Recognition and Demand Prediction: A Reinforcement Ensemble Learning With Fuzzy C-Means Clustering Approach

Demand prediction is essential for enhancing the service quality of ride-hailing platforms. However, the imbalanced and highly skewed nature of ride-hailing demand poses challenges in achieving accurate predictions, especially at major transport hubs. Existing studies have predominantly focused on enhancing prediction models or algorithms, while the lack of attention given to understanding the underlying data characteristics inevitably leads to a mediocre prediction performance. To address this issue, this paper presents a practical approach-referred to Pattern Recognition and Prediction-that makes the trade-off between data intricacy and model flexibility. Firstly, the feature-weighted fuzzy c-means clustering algorithm is employed to assign appropriate pattern labels to daily demand order sequences. Subsequently, reinforcement learning assists ensemble learning to recognize the service pattern for new instance and predict the recognized patterns based on the multi-layer stacking model. In particular, the reinforcement learning dynamically identifies the most suitable combination of predictors, which are then efficiently stacked using a multi-layer stacking model. To quantify the uncertainty in predictions, an improved kernel density estimation is additionally developed for interval predictions. Extensive experiments on a real-world ride-hailing dataset from Beijing West Railway Station, China, demonstrate improvements in both point and interval prediction accuracy, with the former exhibiting a minimum increase of 3.37% compared to benchmark models without pattern recognition, and the latter achieving a more balanced interval width and coverage compared to traditional parametric methods.