The OTTO session dataset is a large-scale dataset intended for multi-objective recommendation research. We collected the data from anonymized behavior logs of the OTTO webshop and the app. The mission of this dataset is to serve as a benchmark for session-based recommendations and foster research in the multi-objective and session-based recommender systems area. We also launched a Kaggle competition with the goal to predict clicks, cart additions, and orders based on previous events in a user session.
Use case
Onsite product discovery for an e-commerce marketplace
| Example observations | Session meta: time since session start, time since last event, number of past clicks / carts / orders Last-K interacted items (IDs) with their event types Item content embeddings (category, brand, price bucket, etc.) Position of last shown item, page type (search, product detail, landing) |
| Example actions | Discrete choice of 1 article ID to display next (or a slate of N items; the dataset lists 1.8 M possibilities) Historical policy: the logs implicitly record which item was actually surfaced/available and then clicked, carted or ordered. |
| Example reward | +5 if the chosen item is ordered +1 if it is added to cart but not ordered +0.1 if it is clicked only 0 otherwise |