PRAG - Prompting for RAG-styled Recommendation
· 2 min read
We are thrilled to announce the release of PRAG, a model family that extends the capabilities of existing encoder-only models to generate personalized recommendations.
PRAG is trained:
- On a mix of public and internal datasets across e - commerce, media, and retail.
- For diverse tasks: next product, similar content, complementary items, etc.
- To minimize feature - engineering.
- For example, shopping history is a de facto input and requires sorting. When using PRAG, you do not need to sort it by timestamp and simply add timestamp as a descriptive feature. If timestamp is not available, PRAG can still implicitly capture temporal patterns.
Versions:
- prag-v1: Base model
| Models | Recall@5/10 | NDCG@5/10 | ||
|---|---|---|---|---|
| Toys and Games | All Beauty | Toys and Games | All Beauty | |
| PRAG-v1 | 0.0452/0.0921 | 0.0452/0.0921 | 0.0267/0.0414 | 0.0267/0.0414 |
| TIGER | 0.0531/0.0712 | 0.0454/0.0648 | 0.0371/0.0432 | 0.0321/0.0384 |
The technical report will be soon released.