Action-On-Item Preference Flow: A Shared Event Schema for Predictive and Generative Personalization

TL;DR

The paper introduces a unified framework called PerTIDE that uses a shared action-on-item schema to update user states across diverse interaction types like news, movies, and dialogue. It demonstrates that a single history encoder can support both prediction and generation tasks, outperforming specialized models when adapted to new domains.

Why It Matters

This approach enables efficient transfer of user modeling across different platforms and modalities, reducing the need for task-specific architectures and training. It advances personalization systems by unifying diverse interaction histories under a common update mechanism.

Key Findings

• A shared action-on-item schema allows consistent state updates across different interaction types despite varying native actions and outputs.
• PerTIDE, implementing the Multi-Timescale State Hypothesis, achieves superior performance in both next-item prediction and personalized generation tasks.
• Frozen core models trained on one domain transfer effectively to another, outperforming random initialization by significant margins on downstream tasks.