ANCHOR framework turns recommendation denoising into supervised learning with LLM-simulated noise

A fundamental challenge in recommendation systems is extracting true user preferences from noisy implicit feedback (e.g., accidental clicks). Existing methods rely on heuristics or unsupervised side information, which often misidentify noise and degrade performance. To address this, researchers propose a paradigm shift: the Creation-Recognition approach. Instead of indirectly inferring noise, they proactively create labeled noisy interactions using LLM-based agents that simulate realistic user behaviors. This transforms denoising into a supervised learning task.

Their framework, ANCHOR, operates in two stages. First, it generates two types of noise: out-of-preference noise (via five extensible simulation mechanisms covering accidental clicks, curiosity exploration, etc.) and boundary-adjacent noise (via adversarial refinement to produce ambiguous interactions near the decision boundary). Second, a reusable parametric recognizer is trained on these synthesized labels, integrating collaborative signals and semantic representations to detect noise patterns in real interaction data. This approach eliminates reliance on expensive annotations and handcrafted heuristics, enabling more robust and accurate recommendation denoising.