Problem
A bank and an internet provider each hold behavior logs of their customers: card transactions on one side, website clickstream on the other. There is no shared identifier, but the business wants to know which provider customer is which bank customer. For each bank customer the model must rank all provider customers by the probability of being the same person. The task comes from Data Fusion Contest 2022 (VTB and Rostelecom).
Example from the article: the organizers' baseline (bank and provider logs, binary classifier on pairs) scored a little over 0.22 on the public leaderboard; the third-place solution reached an R1 of almost 0.30.
Solution in brief
Treat matching as binary classification of customer pairs: known correct pairs plus sampled wrong pairs. Describe each customer with many aggregated features (activity over time, shares by category, hour of day, currency, device), then add 512-number embeddings of the raw event sequences from pytorch-lifestream. Average an ensemble of 11 CatBoost models.
Questions this case answers
- How to rank candidate records from one event log against records from another log when no common key exists and only behavior can link them
- How to link the same person across a bank's transaction log and a telecom provider's clickstream with no shared identifier?
- How to build features from raw event logs for a pairwise matching classifier?
- How to add learned sequence embeddings to a gradient boosting model and check that they help?
Tools and methods
- Pair classification with negative sampling — each correct pair gets 8 or 10 random wrong pairs; the predicted probability ranks candidates
- Aggregated behavior features — time activity, category shares, hour, currency, device
- CatBoost — classifier on pair features, ensemble of 11 models
- PyTorch-Lifestream — neural embeddings of transaction and clickstream sequences
- Feature selection by importance — drop features with zero importance in several seeds
Google or a one-time email code. No password.