DoorDash: Applying Deep Learning to Ads Conversion Prediction in a Last-Mile Delivery Marketplace (Feb 2025)

Forum topic · 小凯 · 2026-07-05

Summary

This paper (arXiv:2502.10514) by Di Li, Xiaochang Miao, Huiyu Song, Chao Chu, Hao Xu, and Mandar Rahurkar of DoorDash describes how deep learning is applied to ads conversion prediction in DoorDash's last-mile delivery marketplace. Sponsored ads require accurate forecasts of whether an ad impression will lead to a conversion, balancing advertiser value, platform revenue, and user experience. The work outlines the problem formulation for conversion prediction in a three-sided marketplace of consumers, merchants, and Dashers, and discusses modeling approaches that capture user intent, context, and item relevance beyond simple click-through-rate estimation. It also covers practical engineering considerations for serving such models at scale in production. This post is an annotated index entry from a curated research list, situating the paper among related work on e-commerce search, query-product relevance, and behavior modeling. Readers should consult the original PDF for exact experimental figures.

DoorDash: Applying Deep Learning to Ads Conversion Prediction in a Last-Mile Delivery Marketplace

Paper: Applying Deep Learning to Ads Conversion Prediction in Last Mile Delivery Marketplace (arXiv:2502.10514, February 2025)
Authors: Di Li, Xiaochang Miao, Huiyu Song, Chao Chu, Hao Xu, Mandar Rahurkar (DoorDash)
Venue: arXiv preprint
Category: Verticals — Search / Ads / Personalization

Overview

This entry summarizes a February 2025 DoorDash technical paper on applying deep learning to ads conversion prediction in a last-mile delivery marketplace. In sponsored advertising systems, conversion prediction estimates the probability that an ad impression ultimately leads to a purchase or order — a harder and sparser signal than click-through rate (CTR), since conversions occur later in the funnel and depend on merchants, inventory, and user intent.

Key points

Problem setting: DoorDash operates a three-sided marketplace (consumers, merchants, Dashers). Ads must balance advertiser ROI, platform revenue, and user experience, making accurate conversion forecasting central to ad ranking and pricing.
Beyond CTR: The paper motivates modeling conversion directly rather than relying solely on click proxies, capturing delayed and lower-frequency conversion signals.
Deep learning approach: The work describes deep neural architectures that encode user context, query/intent signals, and ad/merchant features to predict conversion probability for ad ranking.
Production focus: As an industrial paper, it emphasizes practical deployment considerations — serving latency, data pipelines, and calibration — rather than purely academic benchmarks.

Related entries in this collection

An Interpretable Ensemble of Graph and Language Models
Automated Query-Product Relevance Labeling using LLMs (arXiv:2502.15990)
Behavior Modeling Space Reconstruction for E-Commerce Search (arXiv:2501.18216)
Behavior-driven Query Similarity Prediction with Pre-trained Language Models
Better to Ask in English: Cross-Lingual Evaluation of LLMs
Beyond Relevance: A Demand Balancer Model for Rental Platforms

Why it matters

Ads conversion prediction in delivery marketplaces differs from web search or e-commerce in important ways: demand is local and time-sensitive, catalogs are restaurant menus rather than web documents, and the advertiser (merchant) and platform must share value. This paper is a useful industrial reference for teams building conversion-aware ranking systems in on-demand marketplaces.

> Note: This is an annotated index entry. Quantitative results and full methodological details should be verified against the original paper at https://arxiv.org/abs/2502.10514.