Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Practical Marketplace Optimization at Uber Using Causally-Informed Machine Learning

Bobby Chen, Siyu Chen, Jason Dowlatabadi, Yu Xuan Hong, Vinayak Iyer, Uday Mantripragada, Rishabh Narang, Apoorv Pandey, Zijun Qin, Abrar Sheikh, Hongtao Sun, Jiaqi Sun, Matthew Walker, Kaichen Wei, Chen Xu, Jingnan Yang, Allen T. Zhang, Guoqing Zhang

TL;DR

This work tackles high-dimensional, non-linear budget allocation in Uber’s marketplace under interference by marrying causal deep learning with a low-cost smoothing layer and a distributed optimizer. It introduces a novel S-Learner-based estimator that leverages temporal-spatial observations and experimental data, and a tensor B-Spline surface with a business penalty to capture the outcome surface while respecting monotonicity and convexity constraints. An ADMM-based optimization framework (implemented on Ray) solves the non-convex, multi-city problem efficiently, with a tailored Business Value Evaluation that accounts for network effects via elasticities and A/B data. Backtesting shows the approach improves predictive accuracy and economic efficiency, enabling weekly, scalable, data-driven budget allocations that align with Uber’s objectives. Key innovations include endogeneity handling, ASG-based surface smoothing, IOB-informed evaluation, and a distributed optimization pipeline that supports production deployment. $obj$, $B$, and $IOB$ are used to denote objective, total budget, and incremental budget return, respectively, throughout the discussion.

Abstract

Budget allocation of marketplace levers, such as incentives for drivers and promotions for riders, has long been a technical and business challenge at Uber; understanding lever budget changes' impact and estimating cost efficiency to achieve predefined budgets is crucial, with the goal of optimal allocations that maximize business value; we introduce an end-to-end machine learning and optimization procedure to automate budget decision-making for cities, relying on feature store, model training and serving, optimizers, and backtesting; proposing state-of-the-art deep learning (DL) estimator based on S-Learner and a novel tensor B-Spline regression model, we solve high-dimensional optimization with ADMM and primal-dual interior point convex optimization, substantially improving Uber's resource allocation efficiency.

Practical Marketplace Optimization at Uber Using Causally-Informed Machine Learning

TL;DR

This work tackles high-dimensional, non-linear budget allocation in Uber’s marketplace under interference by marrying causal deep learning with a low-cost smoothing layer and a distributed optimizer. It introduces a novel S-Learner-based estimator that leverages temporal-spatial observations and experimental data, and a tensor B-Spline surface with a business penalty to capture the outcome surface while respecting monotonicity and convexity constraints. An ADMM-based optimization framework (implemented on Ray) solves the non-convex, multi-city problem efficiently, with a tailored Business Value Evaluation that accounts for network effects via elasticities and A/B data. Backtesting shows the approach improves predictive accuracy and economic efficiency, enabling weekly, scalable, data-driven budget allocations that align with Uber’s objectives. Key innovations include endogeneity handling, ASG-based surface smoothing, IOB-informed evaluation, and a distributed optimization pipeline that supports production deployment. , , and are used to denote objective, total budget, and incremental budget return, respectively, throughout the discussion.

Abstract

Budget allocation of marketplace levers, such as incentives for drivers and promotions for riders, has long been a technical and business challenge at Uber; understanding lever budget changes' impact and estimating cost efficiency to achieve predefined budgets is crucial, with the goal of optimal allocations that maximize business value; we introduce an end-to-end machine learning and optimization procedure to automate budget decision-making for cities, relying on feature store, model training and serving, optimizers, and backtesting; proposing state-of-the-art deep learning (DL) estimator based on S-Learner and a novel tensor B-Spline regression model, we solve high-dimensional optimization with ADMM and primal-dual interior point convex optimization, substantially improving Uber's resource allocation efficiency.
Paper Structure (17 sections, 6 theorems, 34 equations, 4 figures, 3 tables, 1 algorithm)

This paper contains 17 sections, 6 theorems, 34 equations, 4 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Causal Marketplace Forecasting
  4. Related Work
  5. Methodology
  6. User Causal Effect Estimator
  7. Smoothing Layer
  8. Adaptive Sparse Grid
  9. B-Spline with Business Penalty
  10. Optimizer
  11. Business Value Evaluation
  12. Evaluation and Results
  13. Conclusion
  14. System Overview
  15. Metalearner Comparison
  16. ...and 2 more sections

Key Result

proposition 1

Given a univariate B-spline function with knots $\boldsymbol{t}=\{t_0,\dots,t_{n+1}\}$, the sufficient condition for $S(x)$ to be monotonically non-decreasing over $[t_r,t_{n}]$ is

Figures (4)

  • Figure 1: Driver model base learner architecture
  • Figure 2: Dense vs. Sparse Grid Representation
  • Figure 3: Budget differences between two models
  • Figure 4: Allocation system architecture: raw data processing into feature store, ML model training and serving, optimization, and evaluation.

Theorems & Definitions (12)

  • proposition 1
  • proof
  • proposition 2
  • proof
  • proposition 3
  • proof
  • lemma 1
  • proof
  • proposition 4
  • proof
  • ...and 2 more