Action-On-Item Preference Flow: A Shared Event Schema for Predictive and Generative Personalization
PerTIDE transfers user-history updates from news to movies, beating a random core by 15.23 MRR.
The paper studies whether movie, news, and dialogue histories can train one reusable user-memory update. An action-on-item schema pairs a mapped interaction role with a content embedding so shared parameters update separate user states. PerTIDE implements the Multi-Timescale State Hypothesis with action gating, three state-space traces, fusion, and command-conditioned readout. On PENS the same encoder supports next-news prediction and personalized headlines; a frozen source core gains 15.23 MRR over a random core on PENS-to-MovieLens, and PerTIDE leads a same-input control by 4.12 MRR on MIND.
- An action-on-item schema shares update parameters across separate user states.
- PerTIDE uses action gating, three state-space traces, fusion, and conditioned readout.
- A frozen source core beats a random core by 15.23 MRR on PENS-to-MovieLens.
- On MIND, PerTIDE keeps a 4.12-point MRR lead over a three-branch control.
From the original article
1 min read at arxiv.orgA user's movie, news, and dialogue histories differ in their native actions and outputs, yet each interaction supplies evidence that can update user memory. We study whether these histories can train one reusable update mechanism. An action-on-item schema pairs a mapped interaction role with a content embedding, allowing shared update parameters to operate on separate user states.
By Parthiv Chatterjee · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.