PRISM: Personalized Item Representations in User History
What it is
This paper proposes PRISM, a method for generating personalized item embeddings in sequential recommendation systems. It uses an autoencoder to compress layer states from user and item towers into a single embedding per scored pair, allowing the same item to have different representations based on the user's relationship to it.
The problem it solves
Standard sequential recommenders use shared embedding maps, causing the same item to have identical representations for all users regardless of their specific interaction history or relationship with that item.
What you could use it for
- Improving sequential recommendation accuracy by providing richer, user-specific history inputs
- Scaling recommendation quality by leveraging longer user interaction histories through personalized embeddings
Why it matters
Establishes personalized item representations as a new scaling dimension for recommendation systems, potentially leading to significant improvements in ranking quality as history length increases.
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.