IRENE algorithm boosts zero-shot retrieval accuracy by 15%
New IRENE algorithm dynamically builds classifiers for unseen items, boosting accuracy by 15%.
Conventional Siamese retrieval approaches use small encoders, limiting capacity for complex tasks, while extreme classification methods excel on seen items but fail for novel (zero-shot) items due to data/latency constraints. Researchers from Microsoft and academia (Sachin Yadav et al.) introduce EMMETT, a novel algorithmic framework that synthesizes classifiers for unseen items by leveraging observed item classifiers, and IRENE, a practical instantiation of EMMETT optimized for large-scale deployments.
In comprehensive experiments, IRENE improves zero-shot retrieval Recall@10 by up to 15% over leading encoders. In an online A/B test on a major search engine's ad retrieval pipeline, it boosted ad click-through rate by 4.2%. The work also provides a theoretical framework for generalization in zero-shot retrieval, guiding training and design decisions. Accepted at KDD 2024, the source code is publicly available.
- IRENE synthesizes classifiers for novel items on-the-fly, overcoming a key limitation of extreme classification methods.
- Achieves up to 15% improvement in Recall@10 for zero-shot retrieval across diverse tasks.
- Online A/B test on a major search engine's ad system shows 4.2% increase in ad click-through rate.
Why It Matters
Enables fast, accurate retrieval of never-before-seen items at scale, critical for search, ads, and recommendation systems.