Blogs›Lyft›Bayesian Trees for Rider Conversion Prediction

Bayesian Trees for Rider Conversion Prediction

posts

2026

Lyft developed and implemented a novel modeling framework using Bayesian Trees to predict rider conversion in sparse data environments. This system addresses the challenge of data sparsity by leveraging a hierarchical tree structure and Bayesian smoothing with Gaussian priors. It allows for robust and accurate predictions even with limited data for specific contexts, while also ensuring behavioral consistency and interpretability by enforcing monotonic relationships where domain knowledge dictates. The platform has been extended to support Augmented Inverse Propensity Weighting (AIPW) for quasi-experimental causal inference, enabling validation of observational data against randomized experiments. This includes rigorous confounder management, propensity score correction, outcome reweighting to address downsampling bias, and detailed model diagnostics like common support and covariate balance checks. The system has been empirically validated against randomized experiments for weekly ride challenges, demonstrating its ability to provide reliable causal estimates from observational data.

2026

Predicting Rider Conversion in Sparse Data Environments with Bayesian Trees

3/30/2026

This post details the development of a Bayesian Tree framework for predicting rider conversion. It introduces a hierarchical tree structure for data decomposition and explains the application of Bayesian smoothing with Gaussian priors to handle data sparsity. The post also discusses how the framework enforces behavioral consistency and monotonicity in predictions, contributing a new capability for real-time predictive modeling in sparse environments.

Trusting the Untestable: Validation and Diagnostics for the Doubly Robust Models

2/17/2026

This post details the implementation and validation of Augmented Inverse Propensity Weighting (AIPW) models for quasi-experimental causal inference at Lyft. It introduces the concept of doubly robust estimation, the critical need for validation in non-randomized settings, and the platform's features for confounder management (pre-defined and customizable sets, leakage prevention) and model diagnostics (common support/propensity overlap, covariate balance). It also describes the scientific refinements for correcting downsampling bias through propensity score correction and outcome reweighting, and an empirical validation using weekly ride challenges to compare AIPW estimates with randomized experiment results.