Reproduce “MAMO: Memory-Augmented Meta-Optimization for Cold-start Recommendation” 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
Manqing Dong, Feng Yuan, Lina Yao, Xiwei Xu, Liming Zhu
MAMO introduces a memory-augmented meta-optimization framework to address the cold-start problem in recommender systems. This approach generates personalized parameter initializations for new users by integrating feature-specific and task-specific memory components, consistently outperforming existing meta-learning baselines in various cold-start scenarios on MovieLens 1M and Book-crossing datasets.
Key insights
- A single global initialization for meta-learning models is often insufficient for diverse user preferences and can lead to local optima; personalized initialization is crucial.
- Explicit memory mechanisms can effectively store and retrieve "task-specific" and "feature-specific" knowledge, providing a more nuanced and adaptive guide for meta-optimization.
- Augmenting meta-optimization with personalized bias terms and fast weights improves personalization and accelerates adaptation to new users in cold-start scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.