Generative recommendation models predict items by autoregressively generating semantic identifiers, or SIDs. Because these identifiers are composed of tokens shared across semantically related items, the representation of one item is entangled with others. A new arXiv paper, “Neighborhood-Aware Unlearning for Generative Recommendation,” takes up the question of how to support unlearning in this setting.
Unlearning—removing the influence of specific data from a trained model—is straightforward in principle but complicated when tokens are shared. The abstract notes that the paradigm is gaining adoption, making the problem timely. The paper’s title indicates that the proposed solution is neighborhood-aware, meaning it likely accounts for the shared-token structure when removing information.
Because the abstract is truncated, details of the method and evaluation are not available in the source. The contribution, as described, is to define and address the unlearning problem for generative recommendation with shared semantic identifiers.