Generative recommendation unlearning improved by controlling neighborhood effects
Neighborhood-Aware Unlearning for Generative Recommendation
Information RetrievalMachine Learning
Summary
When recommendation systems predict items using shared word-like tokens, removing specific user interactions (like bad or unwanted ones) can accidentally affect related items nearby. The authors show that deleting these interactions may not change overall accuracy much but can hurt recommendations close to the removed items. They propose a new method called Neighborhood-Aware Unlearning that carefully manages how probability is spread to these neighboring items, recovering lost performance locally without hurting overall results. This helps systems forget bad data more precisely while keeping good predictions intact.
What this means in practice
- •For e-commerce platform engineers: Remove unwanted or spammy interactions from training data while preserving recommendation quality around related items.
- •For online content recommendation teams: Maintain local recommendation quality after deleting user-item interactions by explicitly controlling token probability shifts.
Authors
Pierre Lubitzsch, Yubao Tang, Maarten de Rijke, Sebastian Schelter
Abstract
Generative recommendation predicts items by autoregressively generating semantic identifiers (SIDs), whose tokens are shared across semantically related items. With this paradigm gaining adoption, the question of how to support unlearning in it has become pressing. If training interactions must later be removed, such as injected spam or interactions with items a user no longer wants, this sharing lets an unlearning update unintentionally affect the predictions for the interactions' neighbors. We show that this effect is locally structured and hidden by global metrics: unlearning can leave recommendation utility on the whole test set nearly unchanged while substantially degrading it on neighborhoods of the forgotten interactions' SIDs. We propose Neighborhood-Aware Unlearning (NAU), which augments interaction-level unlearning with explicit control over the probability mass assigned to the forgotten items' neighborhoods. Across three datasets and four SID tokenizers, NAU restores spam-induced exposure toward the retrained reference while retaining more than 95% of retrained utility, and recovers the neighborhood utility that forget-and-repair unlearning loses at no global cost. We identify local probability redistribution as an important failure mode of unlearning in SID-based generative recommenders and show that explicitly controlling this redistribution can mitigate its effects.