Idea
Model improving marketing decisions by integrating biased observational and scarce experimental data for optimal resource allocation.
Research Paper
Core Innovation
This paper introduces Bi-DFCL, a bi-level optimization framework that jointly leverages biased observational and scarce experimental data to train models focused on decision quality rather than just prediction accuracy. It uses an unbiased estimator of decision quality and implicit differentiation to correct learning directions, addressing bias-variance tradeoffs and prediction-decision misalignment in marketing optimization.
Why It Matters
Marketing teams on online platforms struggle with biased data and misaligned predictive models that do not translate into better decisions. This solution improves marketing ROI by directly optimizing decision quality using both observational and experimental data, enabling scalable, data-driven resource allocation. It transforms marketing workflows by reducing reliance on costly experiments and improving user retention and revenue.
Market Size (TAM)
$10–20B TAM for marketing optimization platforms; $2–5B SAM from large online marketplaces and digital advertisers. Driven by increasing digital ad spend and demand for data-driven marketing decisions.
Potential Customers & Pain Points
- Online marketplaces – Need better marketing ROI
- Digital advertisers – Struggle with biased data and scarce experiments
- E-commerce platforms – Require scalable user retention strategies
- Marketing analytics firms – Need integrated causal learning tools.
Business Model
SaaS subscription model targeting large online platforms and marketing agencies, with tiered pricing based on data volume and feature access.
Competitive Landscape
- Criteo
- Google Marketing Platform
- Adobe Experience Cloud
- Salesforce Marketing Cloud
Implementation Challenges
- Integration complexity with existing marketing systems
- Data privacy and compliance constraints
- High initial setup and experimentation costs
Validation Strategy
- Pilot deployments with major online marketplaces
- Large-scale A/B testing to measure uplift in marketing KPIs
- Benchmarking against state-of-the-art marketing optimization tools
Research Paper Overview
Bi-Level Decision-Focused Causal Learning for Large-Scale Marketing Optimization: Bridging Observational and Experimental Data
Summary
Online platforms face challenges in optimizing marketing strategies due to prediction-decision misalignment and bias-variance tradeoffs in data. Bi-DFCL integrates observational and experimental data through a bi-level optimization framework, improving decision quality and marketing outcomes. Deployed at Meituan, it shows significant gains in large-scale A/B tests and real-world marketing datasets.