Predicting Rider Conversion in Sparse Data Environments with Bayesian Trees

This post details the implementation of a Bayesian Tree framework for predicting rider conversion in sparse data environments. It introduces a hierarchical tree structure for organizing data by context (e.g., spatial, temporal, congestion) and employs Bayesian smoothing with Gaussian priors to leverage parent node statistics for data-sparse leaf nodes. The training process involves L2 regularization to balance local data with parent priors. The framework also enforces monotonic relationships between input features (like historical conversion rate) and predicted conversion probability to ensure behavioral consistency and interpretability.