Bounded Personas Match Retrieval on Classification but Not Regression for a Frozen Agent

A compact user profile can sometimes replace repeated searches through that user's history. PersonaLink compresses past interactions into a fixed-size profile containing preferences, examples, and decision rules, then revises it using errors on reserved examples. With the same underlying model, it achieved news-category accuracy statistically indistinguishable from retrieving three to five relevant past items for each request. But retrieval was substantially stronger at predicting numeric product ratings, and supplying more retrieved examples improved results further. The study therefore supports compact profiles for the tested categorization setting, while showing that fine-grained numeric personalization can benefit from retaining concrete past examples. Methods and results.