[FEA] Synthetic data generation with users preference patterns

[gabrielspmoreira]

🚀 Feature request

Currently, our synthetic data generation is simple, ensuring that user features are the same for the user id, and the same for item features with respect to the item id. But the relationships among users and items are totally random, simulating users interacting with random items.

Motivation

Such approach prevents testing model generalization, as the model might overfit training data (for many epochs training) but provides random predictions for eval set (e.g. AUC=0.5 for binary classification).

Your contribution

I suggest the following approach to introduce in the synthetic data generation process some patterns on users interactions (where users have some preference on some categories of items):

• Check in the schema if there are item features other than the item id
• If not, fallbacks to the current methodology, but if there is other available item features, picks the one that has the one with highest cardinality (i.e. which is more granular, as it might be potentially a feature that groups items, like the item category, which I name here as item group feature).
• When generating users or session interactions, first sample 1 or 2 values of item group feature for each user/session, and then only sample items from those item group for the user/session.