Reproduce “LiMAML: Personalization of Deep Recommender Models via Meta Learning” 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
Ruofan Wang, Prakruthi Prabhakar, Gaurav Srivastava, Tianqi Wang, Zeinab S. Jalali, Varun Bharill, Yunbo Ouyang, Aastha Nigam and 5 more
LiMAML, a production-ready framework from LinkedIn, adapts Model-Agnostic Meta-Learning (MAML) for large-scale recommender systems to enhance personalization for users with limited historical data. It achieved improvements such as a +2.0% increase in Notifications CTR for one application and significantly boosted engagement for infrequent users in online A/B tests.
Key insights
- A dual-block neural network architecture separates a small, adaptable 'Meta Block' for personalization from a larger, shared 'Global Block', improving computational efficiency and storage.
- A decoupled serving strategy allows meta embeddings to be generated offline and stored, eliminating the need for real-time fine-tuning during online inference and ensuring low latency.
- Defining 'tasks' at an entity level (e.g., per-user or per-advertiser) enables meta-learning to focus personalization efforts on relevant interaction patterns.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.