A Multi-Agent Collaborative Framework for Denoising Sequential Recommendation

Abstract

In open environments, interaction data are often contaminated with noisy behaviors that obscure genuine user preferences, leading
to degraded performance in sequential recommenders. Traditional denoising or robust training methods usually rely on pre-defined anomaly
types or extensive supervised signals, which limit their adaptability to unseen or complex data distributions. To address this limitation, we propose a Multi-Agent Collaborative Framework for Denoising Sequential Recommendation (MACDSR for short), which enables adaptive noise
detection through interactive reasoning among agents without requiring prior noise annotation or pre-training. Specifically, MACDSR consists
of three agents with distinct roles: a support agent that argues for the presence of noise in a user sequence, an oppose agent that argues against
it, and a judge agent that synthesizes their arguments to determine which items are likely noisy. This collaborative process allows the framework to dynamically identify and filter out unreliable interactions, thereby producing cleaner input for downstream recommendation models.
By integrating multi-agent reasoning with large language models (LLMs), MACDSR provides a generalized and training-free solution for
denoising sequential recommendation. Experiments on multiple benchmark datasets demonstrate that MACDSR consistently enhances recommendation robustness and accuracy under noisy conditions.