This project is to predict the user's next action(click, add to cart, or order)on a very large set of items. To perform the time series sequence-to-sequence prediction, Otto-recommender uses the transformer encoder architecture and was trained based on a large-scale real-world dataset known as OTTO (https://github.com/otto-de/recsys-dataset).
- 12M real-world anonymized user sessions
- 220M events, consisting of clicks, carts, and orders
- 1.8M unique articles in the catalogue
Data Preprocessing Pipeline: filter out items with low frequency --> encoding--> split data w.r.t. sequence length --> train/test split The transformer encoder uses 4 layers with 4 multiple heads. Negative sampling is used to compute the Loss function, defined as cross-entropy. NDCG is used to measure the effectiveness of this ranking system.