Business Entity Resolution (Amazon ML Challenge 2026)

Built with Team EC2 Bombers for the Amazon ML Challenge 2026, this is an entity-resolution system that decides which vendor records refer to the same real-world business as records in a master list, even when names and addresses are written in different languages and scripts.

The Pipeline

1. Normalisation: Names and addresses are transliterated to a common form using anyascii plus a learned Indic transliteration step, so the same business written in Devanagari and Latin script lines up.
2. Blocking: An exact TF-IDF search from the vendor side keeps 98.25% of true matches while producing only about 2.6 candidates per record, which keeps every later stage cheap.
3. Feature model: 48 hand-built fuzzy, token and field-level features (RapidFuzz) feed a LightGBM classifier.
4. Cross-encoder reranking: Any pair the first model gives at least a 1% chance goes through a bge-reranker-base cross-encoder, trained on SageMaker GPUs.
5. Stacking and assignment: A second LightGBM model combines both scores, and each vendor record is assigned at most one match above a 0.75 threshold.

Results

• 0.9836 macro F0.5 on the public leaderboard, up from 0.842 on the first submission.
• Ranking AUC on the hardest, most uncertain pairs rose from 0.856 to 0.981 as the reranker and stacker were added.
• About 0.954 on a country the model never saw in training (France), showing the pipeline generalises beyond its training markets.