15th place solution (11th on public board)
Card Transaction data is available for 20,000 bank clients.
Data includes:
- GPS coordinates
- MCC code (merchant category)
- Transaction date
- Parsed address in text format
and some other features...
For the train part of 10,000 clients there is also home address and (for 5,000 clients) work address avaliable
To predict home and work for 10,000 clients in test dataset. Prediction is considered correct if error is less then 0.02 degress (~2km for local latitude)
Accuracy
| Idea | Description | Result |
|---|---|---|
| binary classification | 1 - for points near home(work), 0 - for other points | qucikly got near top with it, so chose it as the main approach |
| point candidates generation | extend dataset with new points (as a grid or as cluster centers) | tried cluster centers, it showed slight imporvement |
| convolutional networks / autoencoders | treat transaction map as an image | not tried |
| regression techniques | prediction as a combination of other points | not tried |
| clusterization | use different MCC-cluster centers as predictions | mediocre results, opted for more prospective classification approach |
- Transactions are aggregated into single points by (lonlat, mcc)
- Some customer level and point-level attributes are calculated
- Distance-based attributes are computed individually for each customer
- Catboost performed best among tried models
- Some averaging is done (very simple 5 model bagging combined with two top-scored points averaging)