Problem
eBay's "Similar Sponsored Items" module on the item page recommends promoted listings that resemble the item the buyer is viewing (the "seed"). A strong ranker feature was title similarity, computed with TF-IDF and Jaccard on title keywords. Both treat words as unrelated tokens, so they miss synonyms and handle long titles with many descriptors poorly. A BERT encoder fixes that but is too slow for the online time budget.
Example from the article: "red" and "crimson" share no words, so token-based measures rate two such listings as very dissimilar although the meaning is nearly the same.
Solution in brief
eBay pretrained eBERT on its item titles, distilled it into a small CPU-friendly MicroBERT, and fine-tuned that contrastively on items clicked after the same search. The cosine similarity of the two title embeddings became one more feature in the gradient-boosted ranker. Embeddings are precomputed daily and cached by title.
Questions this case answers
- How to add a semantic text-similarity feature to a learned ranker when token-overlap measures miss synonyms and a large encoder is too slow to run per request
- How to train text embeddings for item-to-item similarity from behavioral co-engagement signals instead of manual labels
- How to add semantic title similarity to a recommendation ranker when keyword overlap misses synonyms?
- How to serve a BERT-based feature in a ranker under a strict online latency limit?
Tools and methods
- eBERT — BERT pretrained on 250 million Wikipedia sentences and three billion eBay titles in five languages
- MicroBERT — small student model distilled from eBERT for CPU inference
- Siamese MicroBERT with InfoNCE — fine-tuning for title embeddings from click-based positive pairs
- Gradient Boosted Tree ranker — ranks candidates by purchase probability, with the title cosine similarity as a feature
- Daily batch job plus key-value cache (NuKV) — takes embedding inference off the request path
Google or a one-time email code. No password.