Problem
A bank and a telecom provider each hold a behavior log of the same people: card transactions on one side, visited page categories on the other. The two logs share no identifier. Comparing single events by hand-made features is too narrow, and production has more candidate pairs than a contest. The VTB team tested whether whole sequences can be matched, using Data Fusion Contest 2022 data.
Example from the article: Transactions from VTB (timestamp, MCC code, currency, amount) were matched with Rostelecom clickstream (timestamp, device ID, category of the visited page). Clickstream sequences reach thousands or tens of thousands of events per client.
Solution in brief
Train one transformer encoder per source, map both outputs into a common space with DSSM and a metric loss on true pairs. Pick and rank candidates on these embeddings with gradient boosting. Cut memory with grouping, filtering, mixed precision and linear attention, and add a classifier for clients with no pair.
Questions this case answers
- How to link records of the same entity across two heterogeneous event logs that share no identifier, when the sequences are long and compute is limited
- How to learn a joint embedding of two behavior sequences from different sources so that nearest-neighbor candidates can be ranked by a downstream model
- How to match the same customers across bank transaction logs and clickstream logs without a shared ID?
- How to train sequence embeddings that put one person's records from two different sources close together?
Tools and methods
- Transformer encoder per source + DSSM — maps both sequences into a common embedding space.
- Metric losses (triplet margin, multi-similarity) and miners — train the space so true pairs are close.
- Hard negative mining — adds pairs the model often gets wrong to random negatives.
- CatBoost — ranks candidates on embeddings and baseline aggregates.
- Linear attention (Linformer), xFormers — attention memory linear in sequence length.
- PyTorch mixed precision, gradient accumulation — less memory, larger virtual batch.
Google or a one-time email code. No password.