DIF Denoises Implicit Feedback for Cold-Start Recommendations.

Implicit feedback, such as clicks or views, is a prevalent data source for recommender systems due to its widespread availability. However, this data often contains noise, stemming from factors like clickbait or position bias. This issue is particularly pronounced for "cold-start" items—new items with limited interaction data—which are more susceptible to noisy samples. Existing denoising techniques, which often rely on heuristic patterns or loss values, struggle with adaptability and are largely ineffective in these cold-start situations. To address this, researchers propose DIF (Denoising Implicit Feedback), a novel model-agnostic method specifically designed for cold-start recommendation. DIF leverages the stability of user preferences for content to infer "pseudo-labels" for cold items by drawing connections to content-similar "warm" (well-established) items. To enhance the accuracy of these pseudo-labels, DIF models their confidence based on content similarity and aggregates multiple pseudo-labels for each sample. Furthermore, DIF explicitly estimates the uncertainty of the noisy sample label by considering their relative entropy and the item's cold-start status. This adaptive mechanism guides the pseudo-labels in correcting noisy labels at a sample level. The method's effectiveness is supported by theoretical justification and extensive experiments on real-world datasets, demonstrating significant improvements in commercial metrics within a billion-user scale short video application, Kuaishou, particularly for cold-start scenarios.