Denoising Implicit Feedback for Cold-start Recommendation

The paper introduces a model-agnostic denoising method called DIF for improving cold-start recommendations by addressing the noise in implicit feedback. DIF infers pseudo-labels for cold items using content-similar warm items, enhances label accuracy through confidence modeling based on content similarity, and estimates label uncertainty to adaptively correct noisy samples. This approach has demonstrated significant improvements in commercial metrics when deployed in a billion-user short video application, highlighting its practical relevance for practitioners dealing with cold-start scenarios in recommendation systems.