Selective Learning-to-Rank for Product Analogs

Product analog discovery is a critical component of modern e-commerce systems, en- abling recommendations, catalog deduplication, and search diversification. Unlike classi- cal similarity search, many products in real-world catalogs do not admit valid substitutes, making forced ranking prone to false positives. This work extends selective prediction to learning-to-rank for analog discovery under partial coverage, introducing a simple yet effective confidence-aware reject mechanism based on score gap and absolute score. Experiments on a large proprietary catalog compris- ing 105 products across 50 categories and 106 labeled pairs show that the proposed method reduces false positives by 25% compared to a forced-ranking baseline while maintaining high coverage and product-level recall. Empirical evaluation across diverse product categories demonstrates a systematic recall– coverage trade-off induced by selective rejection. Price-aware features emerge as the most influential determinants of analog validity, often outweighing fine-grained specification similarity. Overall, selective ranking with abstention is an effective and practically imple- mentable strategy for robust analog discovery at scale.