Reproduce “Finding Lookalike Customers for E-Commerce Marketing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yang Peng, Changzheng Liu, Wei Shen
Walmart Global Tech developed a two-stage system for e-commerce marketing that identifies lookalike customers by generating interpretable embeddings from multimodal customer data. It uses a deep learning model with a two-tower architecture and FAISS for efficient approximate nearest neighbor search, achieving an optimal model quality with an 86% relative MAE using BERT-based location embeddings, albeit at a higher inference cost.
Key insights
- Integrating location embeddings, especially via BERT, substantially improves the quality of customer embeddings (reducing Mean Absolute Error).
- A significant trade-off exists between achieving higher model quality (e.g., using BERT for location) and maintaining efficient inference times for large-scale production deployments.
- Defining similarity metrics directly aligned with business interests enhances the interpretability and adaptability of the lookalike identification system.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.