A competition archive combining time-window customer features, boosting models, ensemble blending and Optuna experiments for customer-behavior modeling.
Historical archive
Python
pandas
NumPy
scikit-learn
LightGBM
XGBoost
CatBoost
Optuna
Original code-generated category illustration; not a product screenshot.
Overview
The ING Türkiye Datathon repository contains Python experiments for customer-behavior modeling. It includes exploration, historical feature engineering, LightGBM/XGBoost/CatBoost models, ensemble blending and hyperparameter search. An earlier feature file and optimized train/test outputs have separate experiment contracts. This is a research archive requiring authorized competition data and a reproducible run setup, rather than a deployed prediction service.
Problem and objectives
Turning customer history into model inputs can introduce leakage if records after a reference date are included. Multiple time windows and model families generate many experiment paths, making train/test schema alignment and traceable outputs important. Optimizing a competition metric also requires more detailed evaluation than a single generic accuracy figure, while customer overlap and temporal splits can change what validation results mean.
My role
Development of data preparation, feature engineering and model-comparison experiments.
Solution approach
The optimized feature path filters customer history before each reference date and creates aggregate features over six-, twelve- and twenty-four-month windows. Its train/test feature files feed ensemble and Optuna scripts. LightGBM, XGBoost, CatBoost and classical models appear in comparison and blending flows; metric code combines Gini, recall and lift components. Exploration and an integrated Colab-oriented experiment remain separate entry points. New runs require explicit input schemas, split choices and recorded seeds.
Implementation details
01
Carrying the reference date into features
For each customer/reference row, the optimized feature script first cuts transaction history to records before ref_date. It then computes summaries such as totals, means and variability over six-, twelve- and twenty-four-month windows, combining them with customer attributes into train/test feature files. This exposes the transformation from raw history to a model table. Skipped customers without usable history and empty-window behavior are part of the contract; output row counts should not automatically be assumed to match raw inputs.
02
Temporal folds for model comparison
ensemble_model.py sorts ref_date values and builds folds whose training records precede each validation date group. It clones the model for each fold, computes AUC, Gini, recall@10 and lift@10 from predict_proba, then fits the final model on all training data. This evaluates a different temporal question from a random split. Category encoding fitted on combined train/test data and customers repeated across dates still require review; temporal folds alone do not establish a leakage-free evaluation.
03
How ensemble weights follow the metric
The weighted ensemble combines model probabilities using normalized ING scores from temporal validation; average and median alternatives also exist. The score function combines Gini, recall@10 and lift@10 through defined baseline ratios. This explicitly connects model weighting to the competition’s ranking-oriented objective. Reusing validation scores for weights does not remove the need for independent final evaluation. The narrative describes the implemented calculation path, without asserting a new leaderboard result or a verified improvement over another model.
