As a global leader in the gig economy, Uber's platform efficiency is severely compromised by high post-match cancellation rates, which disrupt driver utilization and revenue streams. This study aims to predict and mitigate these cancellations using the National Capital Region (NCR) Ride Booking dataset (2024), specifically challenging the prevailing industry reliance on static booking metadata. We demonstrate that conventional models based solely on spatiotemporal features (e.g., timestamp, location IDs) are insufficient, yielding a near-random baseline Area Under the Receiver Operating Characteristic Curve (ROC-AUC) of approximately 0.50. To address this, we introduce a hybrid feature engineering framework that combines K-Fold Regularized Target Encoding with specific transactional context features, namely Ride Distance and Booking Value. Our experimental results indicate that the proposed Light Gradient Boosting Machine (LightGBM) architecture achieves a ROC-AUC of 0.81, representing a 60% performance improvement over the baseline. These findings suggest that cancellations are fundamentally economic decisions driven by the trade-off between ride utility and cost, rather than fixed environmental conditions. Furthermore, this study draws parallels between ride cancellations and social media engagement abandonment, offering a novel perspective for future cross-domain predictive modeling.