DIF: Denoising Implicit Feedback for Cold-start Recommendation

DIF, a model-agnostic denoising method for implicit feedback, targets the item cold-start problem by inferring pseudo-labels from content-similar warm items and using those labels to correct noisy signals. The method was submitted to arXiv on 17 Jun 2026 (arXiv:2606.19658) by Gaode Chen and 10 coauthors and was accepted by KDD 2026 ADS Track; it has been deployed on a billion-user scale short video application Kuaishou and "has significantly improved various commercial metrics within cold-start scenarios."

What is DIF and how does it work?

DIF is a model-agnostic denoising pipeline that creates and weights pseudo-labels for cold-item implicit feedback, then uses an uncertainty estimate to adaptively correct noisy sample labels. The method first infers whether a user is interested in a cold item by referencing content-similar warm items, then models the confidence of each pseudo-label based on content similarity, aggregates multiple pseudo-labels per sample, and finally estimates label uncertainty by considering relative entropy and the item's cold-start status to guide correction.

In practice DIF treats user preferences for content as relatively stable and leverages that stability to transfer signal from warm to cold items. The confidence modeling and aggregation aim to improve pseudo-label accuracy. The final uncertainty computation uses two explicit factors from the paper: relative entropy of the noisy sample label and the cold-start status of the item, which together determine how strongly pseudo-labels should correct original implicit feedback.

How was DIF evaluated and where was it deployed?

The paper provides theoretical justification and reports extensive experiments on real-world datasets, and it notes a production deployment on Kuaishou where the approach improved business metrics in cold-start scenarios. The authors state that DIF "has significantly improved various commercial metrics within cold-start scenarios" after being deployed on a billion-user scale short video application Kuaishou.

The submission metadata identifies the work as arXiv:2606.19658, submitted 17 Jun 2026, and accepted to the KDD 2026 ADS Track. The manuscript emphasizes two evaluation strands: theoretical analysis supporting the denoising approach, and empirical experiments across real-world recommendation datasets; the abstract highlights that prior denoising methods relied on heuristic patterns such as higher loss values and were limited in cold-start adaptability, motivating DIF's design choices.

Why does denoising matter for cold-start recommendation?

Cold items are more prone to noisy implicit feedback because of factors such as clickbait and position bias, and recommenders have historically under-focused on cleaning signals for newly arrived items. By creating content-based pseudo-labels and weighing them by similarity and uncertainty, DIF directly addresses the asymmetry: warm items carry usable preference signal that cold items lack, and treating those signals with confidence-aware aggregation reduces the impact of noisy clicks on model training.

Practically, that matters for systems with large inflows of new content where commercial outcomes depend on early exposure and accurate initial ranking. The paper argues that heuristic denoising (for example, selecting samples by loss magnitude) lacks adaptability in cold-start settings, while DIF’s per-sample confidence and uncertainty make corrections adaptive to each cold item and its observed implicit feedback.

What to watch

Look for the KDD 2026 ADS Track presentation and the full paper details for the empirical results and implementation notes; the arXiv record is arXiv:2606.19658 (submitted 17 Jun 2026). Also watch whether the authors publish code, datasets, or ablation studies that quantify which components (pseudo-label aggregation, confidence modeling, uncertainty estimation) drive the deployment gains on Kuaishou.