We created two recommendation systems, which recommend shopping baskets for the following week. The recommendation systems were implemented on two domains. First one is based on Instacart data set, provided in Kaggle competition. Second system is based on data set from Mercator grocery stores.
We used gradient boosting tree models for prediction of product orders. We build multiple models using different input data and parameters. The models predict, which products will be bought by a customer.
For evaluation of model performance we used F-score, recall and precision measures. On Mercator data set we managed to reach F-score of 0,164. This represents a 6 percent increase in comparison to baseline reference models. On Instacart data set we reached F-score of 0,301 on local test cases and 0,379 on public test cases. This result is comparable to other Kaggle competitors and ranks us on 980. place among 2572 total competitors.
|
