Drift-triggered Evidence Switching for Cold-start Interactive Recommendation under Non-stationary Preferences
Year: 2026
Type: article
Abstract: A demographic prior serves a new user well and turns harmful once taste drifts.DES-UCB keeps two estimators apart.F is fitted on the user's own feedback since the last detected change.C pools the prior rows with that feedback and is warm LinUCB while nothing moves.A controller picks one of the two when they disagree.On a Page-Hinkley firing the old feedback is dropped, C is rebuilt, its credit is wiped and F runs alone for a burn-in.12 claims were preregistered and 10 hold under one tuning rule and seed-clustered intervals.On abrupt drift with a stale prior the regret is 24.7 against 78.0 without suppression, 73.0 for passive decay and 31.3 for a detector-triggered restart. 2 claims fail.Without drift it costs 1.28 times the regret of warm LinUCB.MovieLens-1M shows no precision gain.The contribution is a mechanism, its diagnostics and a preregistered protocol.
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7235040856
