Denoising Implicit Feedback for Cold-start Recommendation

# Denoising Implicit Feedback for Cold-start Recommendation
Source: [https://arxiv.org/html/2606.19658](https://arxiv.org/html/2606.19658)
,Shicheng WangKuaishou TechnologyBeijingChina[wangshicheng@kuaishou\.com](https://arxiv.org/html/2606.19658v1/mailto:[email protected]),Shikun LiHong Kong Baptist UniversityHong KongChina[shikunli\.ml@gmail\.com](https://arxiv.org/html/2606.19658v1/mailto:[email protected]),Rui HuangKuaishou TechnologyBeijingChina[huangrui06@kuaishou\.com](https://arxiv.org/html/2606.19658v1/mailto:[email protected]),Xinghua ZhangIndependent ResearcherBeijingChina[zxh\.zhangxinghua@gmail\.com](https://arxiv.org/html/2606.19658v1/mailto:[email protected]),Yunze LuoPeking UniversityBeijingChina[lyztangent@pku\.edu\.cn](https://arxiv.org/html/2606.19658v1/mailto:[email protected]),Shipeng LiNanjing UniversitySuzhouChina[shipengli\.nju@gmail\.com](https://arxiv.org/html/2606.19658v1/mailto:[email protected]),Shiming GeInstitute of Information Engineering, Chinese Academy of SciencesBeijingChina[geshiming@iie\.ac\.cn](https://arxiv.org/html/2606.19658v1/mailto:[email protected]),Ruina SunKuaishou TechnologyBeijingChina[sunruina@kuaishou\.com](https://arxiv.org/html/2606.19658v1/mailto:[email protected]),Yinjie JiangKuaishou TechnologyBeijingChina[jiangyinjie@kuaishou\.com](https://arxiv.org/html/2606.19658v1/mailto:[email protected])andJun ZhangKuaishou TechnologyBeijingChina[zhangjun08@kuaishou\.com](https://arxiv.org/html/2606.19658v1/mailto:[email protected])

\(2026\)

###### Abstract\.

Implicit feedback is widely used in recommender systems due to its accessibility and generality, yet it usually presents noisy samples \(e\.g\., clickbait, position bias\)\. Meanwhile, recommenders inevitably face the item cold\-start problem due to the continuous influx of new items\. We identify that cold items are more prone to noisy samples due to the aforementioned factors, and researchers often overlook the significance of denoising implicit feedback for cold items\. Previous denoising studies usually identify noisy samples based on heuristic patterns, such as higher loss values, and mitigate noise through sample selection or re\-weighting\. However, these methods have limited adaptability and are ineffective in cold\-start scenarios\. To achieve denoising implicit feedback for cold\-start recommendation, we propose a model\-agnostic denoising method called DIF\. First, user preferences for content remain stable, which allows us to infer pseudo\-labels indicating whether a user is interested in a cold item through content\-similar warm items\. We also elaborate on how to deploy industrial services to retrieve content\-similar warm items for the cold item and obtain their collaborative representations for pseudo\-labeling\. Furthermore, to improve pseudo\-label accuracy, we model the confidence of pseudo\-labels based on the content similarity between the cold item and warm items, and then aggregate multiple pseudo\-labels for each sample\. Finally, we explicitly estimate the uncertainty of the noisy sample label by considering its relative entropy and the cold\-start status of the item, which adaptively guides the role of pseudo\-labels to correct the noisy labels at the sample level\. DIF’s superiority is supported by both theoretical justification and extensive experiments on real\-world datasets\. The method has been deployed on a billion\-user scale short video application Kuaishou and has significantly improved various commercial metrics within cold\-start scenarios\.

Recommender System, Noisy Implicit Feedback, Cold\-start

††journalyear:2026††copyright:cc††conference:Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V\.2; August 09–13, 2026; Jeju Island, Republic of Korea††booktitle:Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V\.2 \(KDD ’26\), August 09–13, 2026, Jeju Island, Republic of Korea††doi:10\.1145/3770855\.3818367††isbn:979\-8\-4007\-2259\-2/2026/08††ccs:Information systems Recommender systems## 1\.Introduction

In the era of information explosion, recommender systems have been playing a critical role for mitigating information overload in various online applications such as short video\(Gonget al\.,[2022](https://arxiv.org/html/2606.19658#bib.bib6); Caiet al\.,[2023](https://arxiv.org/html/2606.19658#bib.bib10)\)and E\-commerce\(Zhouet al\.,[2018](https://arxiv.org/html/2606.19658#bib.bib12); Piet al\.,[2020](https://arxiv.org/html/2606.19658#bib.bib13)\)\. As the most successful technique for personalized recommender systems, collaborative filtering aims to predict items of interest to a specific user based on observed user–item interactions\(Chenet al\.,[2021](https://arxiv.org/html/2606.19658#bib.bib14); Zhanget al\.,[2019](https://arxiv.org/html/2606.19658#bib.bib20)\)\. While explicit user feedback \(e\.g\., ratings\) is the best fuel for recommender systems, its acquisition is often impeded by the need for active user participation\. Hence, implicit feedback \(e\.g\., view and click\) generated during user browsing is exploited as a viable substitute due to its large volume and availability\.

However, the item cold\-start problem\(Caoet al\.,[2022b](https://arxiv.org/html/2606.19658#bib.bib18); Zhouet al\.,[2023](https://arxiv.org/html/2606.19658#bib.bib19)\)may occur when the recommendation model faces a large volume of new items published daily\. For example, tens of millions of new videos are uploaded every day on Kuaishou, one of the most popular short\-video streaming platforms in China with hundreds of millions of active users\. A small portion of popular items tend to obtain more accurate recommendations and more impressions, which creates a strong feedback loop and the Matthew effect, namely “rich gets richer”\(Wanget al\.,[2023a](https://arxiv.org/html/2606.19658#bib.bib47)\)\. Concretely, for a large amount of new emerging items with limited interactions, their embeddings are insufficiently trained, resulting in new items that may miss the opportunity to be recommended or be recommended to inappropriate users\. Thus, the cold\-start problem has become a crucial obstacle for online recommendation\.

As is well known, industrial recommendation systems inevitably contain noisy implicit feedback\(Zhaoet al\.,[2024](https://arxiv.org/html/2606.19658#bib.bib48); Gaoet al\.,[2022](https://arxiv.org/html/2606.19658#bib.bib49)\)\. However, in this work, we highlight an overlooked fact that cold\-start items are more prone to having noisy implicit feedback\. We demonstrate various factors that generate noisy feedback on the Kuaishou platform to support our claim\. On the one hand, users may be influenced by the clickbait issue, where curiosity about the cover or title of a short video leads to clicks or views, resulting infalse positive samples\. Benefiting from the increasingly mature multi\-modal content understanding in the industry, we find that cold\-start videos in the top 10% scoring range of clickbait metrics exceed those in the bottom 10% scoring range by 37\.7%\. On the other hand, users might overlook certain items due to position bias or scroll too fast during browsing fatigue, resulting in a lack of exposure to items and leading tofalse negative samples\. We analyze 1 million user sessions and observe that the proportion of cold\-start videos placed in the tail three positions is 28\.3% higher than their proportion in the top three positions\. Hence, the implicit feedback on these short videos may not always indicate the actual satisfaction of users\. In other words, cold\-start videos are more likely to exhibit label noise due to these factors\.

However, most existing work fails to recognize the importance of denoising implicit feedback for cold\-start items\. Industrial recommendation systems primarily focus on target user exploration for items in the cold\-start phase, as these items lack well\-trained representations\. This often results in fewer positive feedback samples and more negative feedback samples\. Positive feedback serves as a critical signal for optimizing the representation of cold\-start items in recommendation models\. It guides these models to subsequently recommend these cold\-start items to more suitable users, helping them accumulate more positive feedback and exposure\. Regardless of whether the false positive or false negative samples, it misleads recommendation models and wastes trial\-and\-error costs, as exposure opportunities are both limited and valuable\. Such scenarios risk creating a feedback loop of inaccurate learning, which exacerbates the misrepresentation of cold\-start items\. This not only leads to a decline in user experience but also undermines the motivation of producers, potentially causing a missed opportunity for an item to become a breakout success\. As such, denoising implicit feedback for cold\-start recommendation becomes an imperative task\.

Considering the widespread use of implicit feedback and its significant impact on recommendation models, recent studies have noticed the importance of denoising implicit feedback\. Existing efforts on tackling this problem can be roughly divided into two categories: sample selection methods and sample re\-weighting methods\. Sample selection methods\(Wanget al\.,[2021a](https://arxiv.org/html/2606.19658#bib.bib45); Gaoet al\.,[2022](https://arxiv.org/html/2606.19658#bib.bib49)\)focus on improving model performance by selecting clean samples and discarding noisy ones during training\. In contrast, sample re\-weighting methods\(Huet al\.,[2021](https://arxiv.org/html/2606.19658#bib.bib44); Wanget al\.,[2021a](https://arxiv.org/html/2606.19658#bib.bib45)\)aim to assign lower weights to interactions identified as noisy, thereby reducing their influence on the model’s learning process\. The success of these denoising techniques largely depends on the accuracy of distinguishing between clean and noisy samples\. Consequently, various data patterns have been explored as noisy signals\. For example, loss value is one of the most commonly used signals, as noisy interactions tend to exhibit higher loss values compared to clean ones\(Heet al\.,[2024](https://arxiv.org/html/2606.19658#bib.bib46); Wanget al\.,[2021a](https://arxiv.org/html/2606.19658#bib.bib45)\)\. Additionally, other metrics, such as predicted scores\(Wanget al\.,[2022](https://arxiv.org/html/2606.19658#bib.bib41)\)and gradients\(Wanget al\.,[2023b](https://arxiv.org/html/2606.19658#bib.bib40)\)have also been investigated for identifying noisy samples\. However, cold\-start items typically inherently exhibit larger loss values, smaller predicted scores, and higher gradients in the recommendation model, making these denoising methods ineffective in cold\-start scenarios\. Moreover, these denoising methods face challenges in integrating effectively with industrial recommendation models and applying them to online streaming training scenarios\. As a result, existing denoising methods cannot be directly applied to industrial cold\-start recommendation tasks\.

To achieve theDenoisingImplicitFeedback for the cold\-start recommendation, we propose a model\-agnostic methodDIF\. Designing such a method still faces many unknowns: \(i\)How to design a reasonable pseudo\-labeling strategy for cold\-start items?User interests in the content are generally stable\. Thus, whether a user interacts with a cold\-start item can be reasonably inferred through the warm items that are content\-similar to the cold item\. The interest representation of user and the collaborative representations of warm items are well\-trained and can provide meaningful pseudo\-labels\. Moreover, we theoretically prove that even if content similarity between items cannot be fully guaranteed, pseudo\-labels can still approximate the true label as long as the number of warm items is sufficient\. \(ii\)How to generate pseudo\-labels for cold\-start items in online streaming training?Retrieving content\-similar warm items and their collaborative representations in industrial systems remains challenging due to the need for comprehensive candidate sets and real\-time updates\. We detail our practice experiences in applying our denoising method to online recommendation models\. \(iii\)How to further improve the accuracy of pseudo\-labels?We not only consider the top\-kkwarm items most similar to the cold\-start item to generate multiple pseudo\-labels for aggregation, but also explicitly model the confidence of each pseudo\-label based on content similarity during the pseudo\-label aggregation process\. \(iv\)How to correct the noisy sample label based on the final pseudo\-label?We model the uncertainty of the sample label via relative entropy and cold\-start status to adaptively guide the role of the pseudo\-label during the label correction process\.

To summarize, the key contributions are as follows:

- •We highlight that cold\-start items are especially vulnerable to noisy implicit feedback, underscoring the need for denoising\.
- •We propose a theoretically grounded method using multi\-modal semantic similarity to generate pseudo\-labels for cold items, along with practical deployment insights\.
- •We incorporate confidence modeling in the pseudo\-label aggregation process for higher accuracy, and explicitly measure label uncertainty during correction to adaptively control sample\-level pseudo\-label impact\.
- •Extensive offline experiments on three real\-world datasets validate the effectiveness of DIF and its generalizability across different tasks\. DIF has been deployed on a billion\-user scale short video application Kuaishou, yielding significant improvement on a series of commercial metrics in cold\-start scenarios\.

## 2\.Methodology

### 2\.1\.Preliminary

We give a formal description of the denoising implicit feedback for the cold\-start recommendation task\. Letu∈𝒰u\\in\\mathcal\{U\}andi∈ℐi\\in\\mathcal\{I\}denote users and items, with the observed implicit feedback matrix𝐘~∈ℝ\|𝒰\|×\|ℐ\|\\tilde\{\\mathbf\{Y\}\}\\in\\mathbb\{R\}^\{\|\\mathcal\{U\}\|\\times\|\\mathcal\{I\}\|\}\. Andy~u​i\\tilde\{y\}\_\{ui\}==11means that the user interacted with the item, andy~u​i\\tilde\{y\}\_\{ui\}==0means no interaction\. In previous work, the default assumption is that whenevery~u​i\\tilde\{y\}\_\{ui\}==11, it means that the user is interested in the item\. However, user implicit feedback may contain noise due to various factors \(i\.e\., clickbait or position bias\), resulting in noisy sample labels \(false positiveorfalse negative\)\. Thus, our goal is to learn the noise\-free representations of users and items from𝐘~\\tilde\{\\mathbf\{Y\}\}, specifically for cold\-start items\. In our application, a new or cold\-start item is defined as a short video released in less than 24 hours \(inclusive\) and viewed less than 50,000 times\.

Industrial recommenders usually follow a multi\-stage cascade architecture\(Zhanget al\.,[2023b](https://arxiv.org/html/2606.19658#bib.bib50); Gonget al\.,[2022](https://arxiv.org/html/2606.19658#bib.bib6)\)\(such asretrieval\(Yanget al\.,[2020](https://arxiv.org/html/2606.19658#bib.bib51)\),pre\-rankingandranking\(Zhouet al\.,[2018](https://arxiv.org/html/2606.19658#bib.bib12)\)\) to trade off accuracy and efficiency\. In theory, our method works for all the stages and is model\-agnostic\. In practice, we implement and deploy it based on the dual\-tower model in the retrieval stage\. Dual\-tower is the mainstream structure used by the industrial recommendation system in the retrieval stage\. Retrieval methods use a user tower and an item tower to transform user\-side and item\-side features into embeddings𝐞u\\mathbf\{e\}\_\{u\},𝐞i\\mathbf\{e\}\_\{i\}respectively\. The similarity between the user embedding𝐞u\\mathbf\{e\}\_\{u\}and the item embedding𝐞i\\mathbf\{e\}\_\{i\}is used to capture the relevance of useruuto itemiidenoted byy^u​i\\hat\{y\}\_\{ui\}\. During prediction, item embeddings\{𝐞i\}i∈ℐ\{\\left\\\{\\mathbf\{e\}\_\{i\}\\right\\\}\}\_\{i\\in\\mathcal\{I\}\}are calculated beforehand in an offline or near\-line system and indexed by the approximate nearest neighbor \(ANN\) search system, such as FAISS\(Johnsonet al\.,[2019](https://arxiv.org/html/2606.19658#bib.bib52)\)\. Thus, only the user tower needs to be forwarded in real\-time and the ANN system can efficiently retrieve the nearest items in sub\-linear time\.

### 2\.2\.Overview

![Refer to caption](https://arxiv.org/html/2606.19658v1/x1.png)Figure 1\.The overview of our proposed DIF integrated with a dual\-tower model\.![Refer to caption](https://arxiv.org/html/2606.19658v1/x2.png)Figure 2\.The industrial practice for pseudo\-labeling in online streaming training\.In Figure[1](https://arxiv.org/html/2606.19658#S2.F1), we present an overview of the proposed DIF integrated with a base recommendation model\. First, existing denoising methods in general cases typically perform pseudo\-labeling based on the predictions from model\(Tanakaet al\.,[2018](https://arxiv.org/html/2606.19658#bib.bib8); Arazoet al\.,[2019](https://arxiv.org/html/2606.19658#bib.bib9)\)\. However, due to representation issues, the predictions for cold\-start items are often inaccurate\. In contrast, the collaborative representations of some warm items are well\-trained\. Thus, we propose a novel denoising strategy for cold\-start recommendation: generating pseudo\-labels for the cold item based on the collaborative representations of content\-similar warm items and user interest representation\. This involves establishing services for retrieving content\-similar warm items for the cold item and obtaining real\-time collaborative representations of warm items\. We also provide detailed practical experience\. Second, by retrieving the top\-kkcontent\-similar warm items for a cold item, we obtainkkpseudo\-labels\. We further model the confidence of pseudo\-labels based on the content similarity, and then aggregate multiple pseudo\-labels for each sample to improve accuracy\. Finally, we estimate sample label uncertainty via relative entropy and cold\-start status to adaptively perform sample\-level label correction based on the final pseudo\-label\.

### 2\.3\.Pseudo\-labeling

#### 2\.3\.1\.Practical Experience

As shown in Figure[2](https://arxiv.org/html/2606.19658#S2.F2), We address two key practical questions in our industrial environment: how to retrieve content\-similar warm items for the cold item in sample stream and how to obtain collaborative representations of these warm items for subsequent pseudo\-labeling\.

First, for each newly uploaded itemii\(short video\), we can obtain its multi\-modal content representation through the Multi\-modal Inference Service and send it to the message queue\. We have an Embedding Service that reads the content representation of new items from the message queue at all times and stores it\. For each item in the data stream, we can request the Embedding Service to query the content representation of the item \(step①①\)\.

Moreover, we build anitem\-to\-itemANN service, where leftitemsare our queries and rightitemsare the warm items bucket we have constructed as candidates for retrieval\. To ensure sufficient coverage of warm item candidates, we utilize a Real\-time Runner Service to continuously retrieve eligible warm items \(e\.g\., videos with more than 50,000 views\) from the Index, and request the Embedding Service in the previous step to obtain content representations of warm items, which are sent together to the message queue for continuous updating of ANN service candidates\. For the itemiiand its content representations in the data stream, we feed them as input to request the ANN service \(step②②\), which quickly retrieves the top\-kkIDs of warm items with content similar to itemii, along with their similarity scores \(i\.e\.,SiS\_\{i\}==\[si1,si2,…,sik\]\\left\[s^\{1\}\_\{i\},s^\{2\}\_\{i\},\.\.\.,s^\{k\}\_\{i\}\\right\]\) relative to it\.

Furthermore, the online recommendation model actively sends the collaborative representations of users and items to a message queue\. We establish an Embedding Service to store the collaborative representations of all items\. Since an item \(rather than a sample\) may appear multiple times in the sample stream, the online recommendation model continuously updates its representation and sends it to the message queue to update the embedding storage in real time\. We feed the warm items IDs to query this Embedding Service, obtaining their collaborative representations𝐄i\\mathbf\{E\}\_\{i\}==\[𝐞i1,𝐞i2,…,𝐞ik\]\\left\[\\mathbf\{e\}^\{1\}\_\{i\},\\mathbf\{e\}^\{2\}\_\{i\},\.\.\.,\\mathbf\{e\}^\{k\}\_\{i\}\\right\]for subsequent pseudo\-label generation \(step③③\)\.

At this point, we obtain the collaborative representations𝐄i\\mathbf\{E\}\_\{i\}of the top\-kkwarm items with content similar to each itemiiin the data stream, which can be used for pseudo\-labeling in the online recommendation model \(step④④\)\.

#### 2\.3\.2\.Generation of Pseudo Labels

The user interest and the collaborative representations of warm items can be considered well\-trained\. Therefore, predicting interactions between the user and warm items provides valuable reference for predicting whether the user will interact with the content\-similar cold\-start item\. For a sample\(u,i,y~u​i\)\\left\(u,i,\\tilde\{y\}\_\{ui\}\\right\), a dual\-tower model typically predicts the interaction probability between useruuand itemiiby computing the inner product between𝐞u\\mathbf\{e\}\_\{u\}and𝐞i\\mathbf\{e\}\_\{i\}\. Thus, to ensure homogeneity between pseudo\-labels and the sample label, we also take the inner product to generate pseudo\-labels\. Forkkwarm items that have similar content to itemiiin a sample, we obtainkkpseudo\-labels as follows:

\(1\)\[yu​i,1p,yu​i,2p,…,yu​i,kp\]=𝐞u⋅\[𝐞i1,𝐞i2,…,𝐞ik\],\\displaystyle\\left\[y\_\{ui,1\}^\{p\},y\_\{ui,2\}^\{p\},\.\.\.,y\_\{ui,k\}^\{p\}\\right\]=\\mathbf\{e\}\_\{u\}\\cdot\\left\[\\mathbf\{e\}^\{1\}\_\{i\},\\mathbf\{e\}^\{2\}\_\{i\},\.\.\.,\\mathbf\{e\}^\{k\}\_\{i\}\\right\]\\ ,where⋅\\cdotdenotes inner product\.

### 2\.4\.Confidence Modeling

Although we later theoretically show that our method can approximate the true labels without strong similarity assumptions between cold and warm items, different similarity levels should still reflect different degrees of guidance strength\. Thus, it is intuitive to model the confidence of each pseudo\-label based on the similarity between the target item and warm items\. As shown below, after obtaining normalized confidence scorescic\_\{i\}, we compute a weighted sum to derive a more accurate pseudo\-labelyu​ipy\_\{ui\}^\{p\}:

\(2\)cij=exp​\(sij/τ\)∑z=1kexp​\(siz/τ\),j=1,2,…,k,\\displaystyle c\_\{i\}^\{j\}=\\frac\{\{\\rm exp\}\\left\(s\_\{i\}^\{j\}/\\tau\\right\)\}\{\\sum\_\{z=1\}^\{k\}\{\\rm exp\}\\left\(s\_\{i\}^\{z\}/\\tau\\right\)\},j=1,2,\.\.\.,k\\ ,\(3\)yu​ip=∑j=1kcij​yu​i,jp\.\\displaystyle y\_\{ui\}^\{p\}=\\sum\_\{j=1\}^\{k\}c\_\{i\}^\{j\}y\_\{ui,j\}^\{p\}\\ \.Here,τ\\tauis a temperature coefficient\. Benefiting from the existing advanced multi\-modal content understanding capabilities, online observations indicate that similarity scores are generally above 0\.9, andτ\\tauserves to enhance differentiation\.

### 2\.5\.Uncertainty Estimation

To effectively perform label correction, we estimate‌ the uncertainty of the noisy sample labely~u​i\\tilde\{y\}\_\{ui\}from two perspectives to adaptively guide the role of pseudo\-labelsyu​ipy\_\{ui\}^\{p\}at the sample level\.

First, we consider the relative entropy\(Coveret al\.,[1991](https://arxiv.org/html/2606.19658#bib.bib7)\)of the sample label with the predicted value \(e\.g\.,y^u​i\\hat\{y\}\_\{ui\}==\[𝐞u\]⊤⋅𝐞i\\left\[\\mathbf\{e\}\_\{u\}\\right\]^\{\\top\}\\cdot\\mathbf\{e\}\_\{i\}\) as a crucial indicator of the uncertainty, termed asru​ir\_\{ui\}\. The higher the relative entropy, the more likely the sample is mislabeled\. Its calculation is as below:

\(4\)ru​i=−y~u​i​log⁡y^u​i−\(1−y~u​i\)​log⁡\(1−y^u​i\)\.\\displaystyle r\_\{ui\}=\-\\tilde\{y\}\_\{ui\}\\log\\hat\{y\}\_\{ui\}\-\\left\(1\-\\tilde\{y\}\_\{ui\}\\right\)\\log\\left\(1\-\\hat\{y\}\_\{ui\}\\right\)\\ \.
Second, as previously mentioned, cold\-start items are more prone to generating label noise due to various factors\. Thus, we explicitly model the cold\-start stateγi\\gamma\_\{i\}of an itemiias another indicator for the uncertainty estimation:

\(5\)γi=e−α​gi,\\displaystyle\\gamma\_\{i\}=e^\{\-\\alpha g\_\{i\}\}\\ ,whereeedenotes the natural constant,gig\_\{i\}represents the interaction count of the itemii, andα\\alphacontrols the threshold for defining the cold\-start state\. A higher interaction count drivesγi\\gamma\_\{i\}closer to 0, while fewer interactions push it closer to 1\. This design effectively prevents any adverse impact on the training of warm items\.

Then, by integrating relative entropyru​ir\_\{ui\}and cold\-start statusγi\\gamma\_\{i\}, we assign higher uncertainty to cold\-start items with significant prediction discrepancies\. The final uncertainty estimationtu​it\_\{ui\}is formulated as shown below and constrained to be less than 1\.

\(6\)tu​i=ru​i​γi\.\\displaystyle t\_\{ui\}=r\_\{ui\}\\gamma\_\{i\}\\ \.

### 2\.6\.Model Optimization

To alleviate the impact of label noise, we adaptively correct the sample label based on its uncertainty at the sample level\. Specifically, taking the weight of the original sample label as 1, we normalize the uncertainty by scaling it within the range by Eq\. \([7](https://arxiv.org/html/2606.19658#S2.E7)\), and then we obtain the corrected soft labelyu​icy\_\{ui\}^\{c\}for each sample by Eq\. \([8](https://arxiv.org/html/2606.19658#S2.E8)\)\.

\(7\)w=tu​itu​i\+1,\\displaystyle w=\\frac\{t\_\{ui\}\}\{t\_\{ui\}\+1\}\\ ,\(8\)yu​ic=w​yu​ip\+\(1−w\)​y~u​i\.\\displaystyle y\_\{ui\}^\{c\}=wy\_\{ui\}^\{p\}\+\(1\-w\)\\tilde\{y\}\_\{ui\}\\ \.
Finally, we utilize the widely\-used cross\-entropy loss for training:

\(9\)ℒ=−∑\(u,i\)∈𝒟train\(yu​ic​log⁡y^u​i\+\(1−yu​ic\)​log⁡\(1−y^u​i\)\)\.\\displaystyle\\mathcal\{L\}=\-\\sum\_\{\(u,i\)\\in\\mathcal\{D\}\_\{\\text\{train\}\}\}\\left\(y\_\{ui\}^\{c\}\\log\\hat\{y\}\_\{ui\}\+\\left\(1\-y\_\{ui\}^\{c\}\\right\)\\log\\left\(1\-\\hat\{y\}\_\{ui\}\\right\)\\right\)\\ \.

### 2\.7\.Deployment

Our proposed DIF is trained on a large\-scale distributed nearline learning system of Kuaishou\. Every day, hundreds of millions of users visit Kuaishou, watch and interact with short videos, and yield tens of billions watch and interaction logs per day\. On the one hand, each log is collected in real\-time, preprocessed by the Kafka stream processing platform, and produces the sample stream\. In Section 2\.3, we provide a detailed practice experience of how to obtain the collaborative representations of content\-similar warm items for each item in the sample stream\. As training data, including the aforementioned features, is received by the downstream recommendation model through the pipeline, this near\-line training system incrementally updates the model parameters using the latest knowledge from user\-video interactions\. In streaming training, our approach effectively minimizes the impact on the training of warm items by leveraging uncertainty estimation while focusing on assisting cold\-start items in correcting noisy labels during the early stages, thereby enabling higher growth potential\.

### 2\.8\.Theoretical Justification

In this section, we present the theoretical analyses to justify the validity of our denoising approach\.

Letei∈ℰ⊂ℝde\_\{i\}\\in\\mathcal\{E\}\\subset\\mathbb\{R\}^\{d\}be the normalized multi\-modal embedding of a cold\-start itemii\. Let𝒩k​\(i\)\\mathcal\{N\}\_\{k\}\(i\)be the set ofkknearest warm items retrieved based on context similarity\. Letηu​i∗=ηu∗​\(𝐞i\)\\eta^\{\*\}\_\{ui\}=\\eta^\{\*\}\_\{u\}\(\\mathbf\{e\}\_\{i\}\)be the ground\-truth interaction probability between useruuand itemii\. Letyu​i,jp\{y\}^\{p\}\_\{ui,j\}be the pseudo label between useruuand warm neighbor itemj∈𝒩k​\(i\)j\\in\\mathcal\{N\}\_\{k\}\(i\)\. Letyu​ip\{y\}^\{p\}\_\{ui\}be the confidence\-weighed pseudo label between useruuand cold\-start itemii\.

###### Assumption 2\.1\.

The observed cold\-start item labely~u​i\\tilde\{y\}\_\{ui\}is a biased estimator of its true interaction probabilityηu​i∗\\eta^\{\*\}\_\{ui\}, i\.e\.,y~u​i=ηu​i∗\+ϵu​i\\tilde\{y\}\_\{ui\}=\\eta^\{\*\}\_\{ui\}\+\\epsilon\_\{ui\}, where𝔼​\[ϵu​i\]=bu​i\\mathbb\{E\}\[\\epsilon\_\{ui\}\]=b\_\{ui\}andVar​\[ϵu​i\]=σu​i2\\text\{Var\}\[\\epsilon\_\{ui\}\]=\\sigma^\{2\}\_\{ui\}\.

Remark 1\. Assumption[2\.1](https://arxiv.org/html/2606.19658#S2.ThmAssumption1)means that due to various factors, such as clickbait issues and position bias, the observed cold\-start item label has undeniable systematic bias noisebu​ib\_\{ui\}and relatively large noise varianceσu​i2\\sigma^\{2\}\_\{ui\}\.

###### Assumption 2\.2\.

The pseudo warm item labelyu​i,jp\{y\}^\{p\}\_\{ui,j\}is an unbiased estimator of its true interaction probabilityηu​j∗\\eta^\{\*\}\_\{uj\}, i\.e\.,yu​i,jp=ηu​j∗\+ϵu​i,j′\{y\}^\{p\}\_\{ui,j\}=\\eta^\{\*\}\_\{uj\}\+\\epsilon^\{\\prime\}\_\{ui,j\}, where𝔼​\[ϵu​i,j′\]=0\\mathbb\{E\}\[\\epsilon^\{\\prime\}\_\{ui,j\}\]=0andVar​\[ϵu​i,j′\]=σu′⁣2\\text\{Var\}\[\\epsilon^\{\\prime\}\_\{ui,j\}\]=\\sigma^\{\\prime 2\}\_\{u\}\.

Remark 2\. Since the user interest representation and collaborative representations of warm items can be considered well\-trained with a sufficiently large sample, Assumption[2\.2](https://arxiv.org/html/2606.19658#S2.ThmAssumption2)posits that the pseudo warm item label has negligible approximation error \(𝔼​\[ϵu​i,j′\]\\mathbb\{E\}\[\\epsilon^\{\\prime\}\_\{ui,j\}\]\) and small varianceσu′⁣2\\sigma^\{\\prime 2\}\_\{u\}\.

###### Assumption 2\.3\.

The context representation spaceℰ\\mathcal\{E\}is a compact subset ofℝd\\mathbb\{R\}^\{d\}, and the probability density functionp​\(𝐞\)p\(\\mathbf\{e\}\)is bounded away from zero onℰ\\mathcal\{E\}, i\.e\., there exists a constantpinf\>0p\_\{\\inf\}\>0such thatp​\(𝐞\)≥pinfp\(\\mathbf\{e\}\)\\geq p\_\{\\inf\}for all𝐞∈ℰ\\mathbf\{e\}\\in\\mathcal\{E\}\.

###### Assumption 2\.4\.

The true interaction probability functionηu∗​\(𝐞\)\\eta^\{\*\}\_\{u\}\(\\mathbf\{e\}\)isα\\alpha\-Hölder continuous on the content representations space, i\.e\., there exists0<α<10<\\alpha<1andL\>0L\>0such that\|ηu∗​\(𝐞i\)−ηu∗​\(𝐞j\)\|≤L​‖𝐞i−𝐞j‖α\|\\eta\_\{u\}^\{\*\}\(\\mathbf\{e\}\_\{i\}\)\-\\eta\_\{u\}^\{\*\}\(\\mathbf\{e\}\_\{j\}\)\|\\leq L\\\|\\mathbf\{e\}\_\{i\}\-\\mathbf\{e\}\_\{j\}\\\|^\{\\alpha\}, for any𝐞i,𝐞j∈ℰ\\mathbf\{e\}\_\{i\},\\mathbf\{e\}\_\{j\}\\in\\mathcal\{E\}\.

Remark 3\. Assumption[2\.3](https://arxiv.org/html/2606.19658#S2.ThmAssumption3)ensures valid nearest neighbors exist within a small radius around any item when having a sufficiently large sample, and Assumption[2\.4](https://arxiv.org/html/2606.19658#S2.ThmAssumption4)makes a smoothness assumption onηu∗​\(𝐞\)\\eta^\{\*\}\_\{u\}\(\\mathbf\{e\}\)\. Both of them are standard conditions for the analysis under k\-NN classification\(Bahriet al\.,[2020](https://arxiv.org/html/2606.19658#bib.bib3); Gaoet al\.,[2018](https://arxiv.org/html/2606.19658#bib.bib4); Györfiet al\.,[2002](https://arxiv.org/html/2606.19658#bib.bib5)\)\.

###### Theorem 2\.1\.

Suppose Assumptions[2\.2](https://arxiv.org/html/2606.19658#S2.ThmAssumption2),[2\.3](https://arxiv.org/html/2606.19658#S2.ThmAssumption3), and[2\.4](https://arxiv.org/html/2606.19658#S2.ThmAssumption4)hold\. For any cold\-start itemi∈ℐi\\in\\mathcal\{I\}, useruu∈\\in𝒰\\mathcal\{U\},δ∈\(0,1\)\\delta\\in\(0,1\), we have

\|yu​ip−ηu​i∗\|≤𝒪​\(L​\(kNw​a​r​m\)αd\)\+𝒪​\(σu′k\),\|y^\{p\}\_\{ui\}\-\\eta^\{\*\}\_\{ui\}\|\\leq\\mathcal\{O\}\\left\(L\\left\(\\frac\{k\}\{N\_\{warm\}\}\\right\)^\{\\frac\{\\alpha\}\{d\}\}\\right\)\+\\mathcal\{O\}\\left\(\\frac\{\\sigma^\{\\prime\}\_\{u\}\}\{\\sqrt\{k\}\}\\right\),wherekkis the number of context\-similar neighbors,Nw​a​r​mN\_\{warm\}is the number of warm items,ddis the dimension of the context spaceℰ\\mathcal\{E\}\.

###### Proof\.

Under Assumption[2\.2](https://arxiv.org/html/2606.19658#S2.ThmAssumption2), we first decompose the error as follows:

\(10\)\|yu​ip−ηu​i∗\|\\displaystyle\|y^\{p\}\_\{ui\}\-\\eta^\{\*\}\_\{ui\}\|=\|∑j∈𝒩k​\(i\)cij​y~u​i,j−ηu​i∗\|=\|∑j∈𝒩k​\(i\)cij​\(y~u​i,j−ηu​i∗\)\|\\displaystyle=\\left\|\\sum\_\{j\\in\\mathcal\{N\}\_\{k\}\(i\)\}c\_\{i\}^\{j\}\\tilde\{y\}\_\{ui,j\}\-\\eta^\{\*\}\_\{ui\}\\right\|=\\left\|\\sum\_\{j\\in\\mathcal\{N\}\_\{k\}\(i\)\}c\_\{i\}^\{j\}\(\\tilde\{y\}\_\{ui,j\}\-\\eta^\{\*\}\_\{ui\}\)\\right\|=\|∑j∈𝒩k​\(i\)cij​\(ηu​j∗\+ϵu​i,j′−ηu​i∗\)\|\\displaystyle=\\left\|\\sum\_\{j\\in\\mathcal\{N\}\_\{k\}\(i\)\}c\_\{i\}^\{j\}\(\\eta^\{\*\}\_\{uj\}\+\\epsilon^\{\\prime\}\_\{ui,j\}\-\\eta^\{\*\}\_\{ui\}\)\\right\|≤\|∑j∈𝒩k​\(i\)cij​\(ηu​j∗−ηu​i∗\)\|\+\|∑j∈𝒩k​\(i\)cij​ϵu​i,j′\|\.\\displaystyle\\leq\{\\left\|\\sum\_\{j\\in\\mathcal\{N\}\_\{k\}\(i\)\}c\_\{i\}^\{j\}\(\\eta^\{\*\}\_\{uj\}\-\\eta^\{\*\}\_\{ui\}\)\\right\|\}\+\{\\left\|\\sum\_\{j\\in\\mathcal\{N\}\_\{k\}\(i\)\}c\_\{i\}^\{j\}\\epsilon^\{\\prime\}\_\{ui,j\}\\right\|\}\.
For the first term\|∑j∈𝒩k​\(i\)cij​\(ηu​j∗−ηu​i∗\)\|\{\\left\|\\sum\_\{j\\in\\mathcal\{N\}\_\{k\}\(i\)\}c\_\{i\}^\{j\}\(\\eta^\{\*\}\_\{uj\}\-\\eta^\{\*\}\_\{ui\}\)\\right\|\}, it can be further decomposed under Assumption[2\.4](https://arxiv.org/html/2606.19658#S2.ThmAssumption4):

\(11\)\|∑j∈𝒩k​\(i\)cij​\(ηu​j∗−ηu​i∗\)\|≤∑j∈𝒩k​\(i\)cij​L​‖𝐞j−𝐞i‖α≤L⋅maxj∈𝒩k​\(i\)⁡‖𝐞j−𝐞i‖α\.\\left\|\\sum\_\{j\\in\\mathcal\{N\}\_\{k\}\(i\)\}c\_\{i\}^\{j\}\(\\eta^\{\*\}\_\{uj\}\-\\eta^\{\*\}\_\{ui\}\)\\right\|\\leq\\sum\_\{j\\in\\mathcal\{N\}\_\{k\}\(i\)\}c\_\{i\}^\{j\}L\\\|\\mathbf\{e\}\_\{j\}\-\\mathbf\{e\}\_\{i\}\\\|^\{\\alpha\}\\leq L\\cdot\\max\_\{j\\in\\mathcal\{N\}\_\{k\}\(i\)\}\\\|\\mathbf\{e\}\_\{j\}\-\\mathbf\{e\}\_\{i\}\\\|^\{\\alpha\}\.According to non\-parametric regression theory \(See Theorem 6\.2 in\(Györfiet al\.,[2002](https://arxiv.org/html/2606.19658#bib.bib5)\)\), in add\-dimensional space withNw​a​r​mN\_\{warm\}samples and a probability density function bounded away from zero \(Assumption[2\.3](https://arxiv.org/html/2606.19658#S2.ThmAssumption3)\), the distance between a target point and itskk\-th nearest neighbormaxj∈𝒩k​\(i\)⁡‖𝐞j−𝐞i‖\\max\_\{j\\in\\mathcal\{N\}\_\{k\}\(i\)\}\\\|\\mathbf\{e\}\_\{j\}\-\\mathbf\{e\}\_\{i\}\\\|is bounded by𝒪​\(\(k/Nw​a​r​m\)1/d\)\\mathcal\{O\}\(\(k/N\_\{warm\}\)^\{1/d\}\)\. Substituting this into the above inequality, we obtain:

\(12\)\|∑j∈𝒩k​\(i\)cij​\(ηu​j∗−ηu​i∗\)\|≤𝒪​\(L​\(kNw​a​r​m\)αd\)\.\\left\|\\sum\_\{j\\in\\mathcal\{N\}\_\{k\}\(i\)\}c\_\{i\}^\{j\}\(\\eta^\{\*\}\_\{uj\}\-\\eta^\{\*\}\_\{ui\}\)\\right\|\\leq\\mathcal\{O\}\\left\(L\\left\(\\frac\{k\}\{N\_\{warm\}\}\\right\)^\{\\frac\{\\alpha\}\{d\}\}\\right\)\.
For the second term\|∑j∈𝒩k​\(i\)cij​ϵu​i,j′\|\{\\left\|\\sum\_\{j\\in\\mathcal\{N\}\_\{k\}\(i\)\}c\_\{i\}^\{j\}\\epsilon^\{\\prime\}\_\{ui,j\}\\right\|\}, by assumingϵu​i,j′\\epsilon^\{\\prime\}\_\{ui,j\}represents bounded independent zero\-mean random noise with variance \(Assumption[2\.2](https://arxiv.org/html/2606.19658#S2.ThmAssumption2)\), by applying Hoeffding’s Inequality\(Boucheronet al\.,[2013](https://arxiv.org/html/2606.19658#bib.bib1)\), we have:

\(13\)\|∑j∈𝒩k​\(i\)cij​ϵu​i,j′\|≤𝒪​\(σu′ke​f​f\)≈𝒪​\(σu′k\),\{\\left\|\\sum\_\{j\\in\\mathcal\{N\}\_\{k\}\(i\)\}c\_\{i\}^\{j\}\\epsilon^\{\\prime\}\_\{ui,j\}\\right\|\}\\leq\\mathcal\{O\}\\left\(\\frac\{\\sigma^\{\\prime\}\_\{u\}\}\{\\sqrt\{k\_\{eff\}\}\}\\right\)\\approx\\mathcal\{O\}\\left\(\\frac\{\\sigma^\{\\prime\}\_\{u\}\}\{\\sqrt\{k\}\}\\right\),whereke​f​f=1/\(∑\(cij\)2\)k\_\{eff\}=1/\\left\(\\sum\\left\(c\_\{i\}^\{j\}\\right\)^\{2\}\\right\), and in the high\-dimensional spaces with dense sampling,ke​f​f≈kk\_\{eff\}\\approx kdue to the concentration of measure\.

Finally, substituting Eq\. \([12](https://arxiv.org/html/2606.19658#S2.E12)\) and Eq\. \([13](https://arxiv.org/html/2606.19658#S2.E13)\) into Eq\. \([10](https://arxiv.org/html/2606.19658#S2.E10)\), we can obtain:

\(14\)\|yu​ip−ηu​i∗\|≤𝒪​\(L​\(kNw​a​r​m\)αd\)\+𝒪​\(σu′k\)\\left\|y^\{p\}\_\{ui\}\-\\eta^\{\*\}\_\{ui\}\\right\|\\leq\\mathcal\{O\}\\left\(L\\left\(\\frac\{k\}\{N\_\{warm\}\}\\right\)^\{\\frac\{\\alpha\}\{d\}\}\\right\)\+\\mathcal\{O\}\\left\(\\frac\{\\sigma^\{\\prime\}\_\{u\}\}\{\\sqrt\{k\}\}\\right\)∎

Remark 4\. Theorem[2\.1](https://arxiv.org/html/2606.19658#S2.Thmtheorem1)clearly shows that, when the number of warm itemsNw​a​r​mN\_\{warm\}is sufficiently large, by choosing a suitablekk, the error bound betweenyu​ipy^\{p\}\_\{ui\}andηu​i∗\\eta^\{\*\}\_\{ui\}will become small, which clearly justifies the cleanness of the pseudo labelsyu​ipy^\{p\}\_\{ui\}\.

###### Theorem 2\.2\.

Suppose Assumptions[2\.1](https://arxiv.org/html/2606.19658#S2.ThmAssumption1)and[2\.2](https://arxiv.org/html/2606.19658#S2.ThmAssumption2)hold\. Assume the observation noiseϵu​i\\epsilon\_\{ui\}and the pseudo\-label noiseϵu​i,j′\\epsilon^\{\\prime\}\_\{ui,j\}are uncorrelated\. There exists an optimal weightw∗∈\[0,1\]w^\{\*\}\\in\[0,1\]that minimizes the MSE riskR​\(yu​ic\)=𝔼​\[\(yu​ic−ηu​i∗\)2\]R\(y^\{c\}\_\{ui\}\)=\\mathbb\{E\}\[\(y^\{c\}\_\{ui\}\-\\eta^\{\*\}\_\{ui\}\)^\{2\}\]\. Furthermore, the optimal weight is

w∗=λu​iλu​i\+1,and​λu​i=R​\(y~u​i\)R​\(yu​ip\),w^\{\*\}=\\frac\{\\lambda\_\{ui\}\}\{\\lambda\_\{ui\}\+1\},\\text\{and\}~\\lambda\_\{ui\}=\\frac\{R\(\\tilde\{y\}\_\{ui\}\)\}\{R\(y^\{p\}\_\{ui\}\)\},whereλu​i\\lambda\_\{ui\}represents the noise ratio ofy~u​i\\tilde\{y\}\_\{ui\}toyu​ipy^\{p\}\_\{ui\}\.

###### Proof\.

To minimize the Mean Squared Error \(MSE\) of the corrected labelyu​icy^\{c\}\_\{ui\}with respect to the true probabilityηu​i∗\\eta^\{\*\}\_\{ui\}:

\(15\)R​\(yu​ic\)\\displaystyle R\(y^\{c\}\_\{ui\}\)=𝔼​\[\(w​yu​ip\+\(1−w\)​y~u​i−ηu​i∗\)2\]\\displaystyle=\\mathbb\{E\}\\left\[\\left\(wy^\{p\}\_\{ui\}\+\(1\-w\)\\tilde\{y\}\_\{ui\}\-\\eta^\{\*\}\_\{ui\}\\right\)^\{2\}\\right\]=w2​𝔼​\[\(yu​ip−ηu​i∗\)2\]\+\(1−w\)2​𝔼​\[\(y~u​i−ηu​i∗\)2\]\\displaystyle=w^\{2\}\\mathbb\{E\}\[\(y^\{p\}\_\{ui\}\-\\eta^\{\*\}\_\{ui\}\)^\{2\}\]\+\(1\-w\)^\{2\}\\mathbb\{E\}\[\(\\tilde\{y\}\_\{ui\}\-\\eta^\{\*\}\_\{ui\}\)^\{2\}\]\+2​w​\(1−w\)​𝔼​\[\(yu​ip−ηu​i∗\)​\(y~u​i−ηu​i∗\)\]\\displaystyle\+2w\(1\-w\)\\mathbb\{E\}\[\(y^\{p\}\_\{ui\}\-\\eta^\{\*\}\_\{ui\}\)\(\\tilde\{y\}\_\{ui\}\-\\eta^\{\*\}\_\{ui\}\)\]=w2​R​\(yu​ip\)\+\(1−w\)2​R​\(y~u​i\),\\displaystyle=w^\{2\}R\(y^\{p\}\_\{ui\}\)\+\(1\-w\)^\{2\}R\(\\tilde\{y\}\_\{ui\}\),where the last equation holds because we assume the observation noise and the pseudo\-label noise are uncorrelated\. Taking the derivative with respect towwand setting it to zero:

2​w​R​\(yu​ip\)−2​\(1−w\)​R​\(y~u​i\)=0\.2wR\(y^\{p\}\_\{ui\}\)\-2\(1\-w\)R\(\\tilde\{y\}\_\{ui\}\)=0\.Solving for the optimal weight:

w∗=R​\(y~u​i\)R​\(y~u​i\)\+R​\(yu​ip\)=λu​iλu​i\+1\.w^\{\*\}=\\frac\{R\(\\tilde\{y\}\_\{ui\}\)\}\{R\(\\tilde\{y\}\_\{ui\}\)\+R\(y^\{p\}\_\{ui\}\)\}=\\frac\{\\lambda\_\{ui\}\}\{\\lambda\_\{ui\}\+1\}\.∎

Remark 5\. Theorem[2\.2](https://arxiv.org/html/2606.19658#S2.Thmtheorem2)derives an optimal weightw∗=λu​iλu​i\+1w^\{\*\}=\\frac\{\\lambda\_\{ui\}\}\{\\lambda\_\{ui\}\+1\}, which mathematically aligns with the weighting scheme proposed in Eq\. \([7](https://arxiv.org/html/2606.19658#S2.E7)\)\. SinceR​\(y~u​i\)R\(\\tilde\{y\}\_\{ui\}\)andR​\(yu​ip\)R\(y^\{p\}\_\{ui\}\)are unknown, we utilize the uncertainty termtu​it\_\{ui\}to measure the observation noise level by the relative entropyru​ir\_\{ui\}and cold\-start statusγi\\gamma\_\{i\}\. It can be regarded as an effective estimator of the noise ratio for denoising\.

## 3\.Experiments

### 3\.1\.Experimental Settings

#### 3\.1\.1\.Dataset

Table 1\.Statistics of experimented datasets with multi\-modal item Visual\(V\), Acoustic\(A\), Textual\(T\) contents\.DatasetAmazonTiktokSportsBabyModalityVTVTVATEmbed Dim4096102440961024128128768User35,59819,4459,319Item18,3577,0506,710Interactions256,308139,11059,541Sparsity99\.961%99\.899%99\.904%Following\(Chenet al\.,[2024](https://arxiv.org/html/2606.19658#bib.bib15)\), we conduct offline experiments on three real\-world multi\-modal recommendation datasets111All datasets are publicly available at[https://github\.com/HKUDS/MMSSL/tree/main](https://github.com/HKUDS/MMSSL/tree/main)\., i\.e\., Amazon\-Sports, Amazon\-Baby, Tiktok\. Data statistics with multi\-modal feature embedding dimensionality are reported in Table[1](https://arxiv.org/html/2606.19658#S3.T1)\.

- •Amazon\.We adopt two benchmark datasets from Amazon with two item categories Baby and Sports\. In those datasets, textual feature embeddings are generated via Sentence\-Bert\(Reimers and Gurevych,[2019](https://arxiv.org/html/2606.19658#bib.bib28)\)based on the extracted text from the product title, description, brand, and categorical information\. The product images are used to generate 4096\-ddvisual feature embeddings of items\.
- •TikTok\.This data is collected from TikTok platform to log the viewed short\-videos of users\. The multi\-modal features are visual, acoustic, and title textual features of videos\. The textual embeddings are also encoded with Sentence\-Bert\.

#### 3\.1\.2\.Evaluation Protocols

For each dataset, we used the ratio 8:1:1 to randomly split the historical interactions of each user and constituted the training set, validation set, and testing set\. Moreover, following the widely\-used evaluation metrics\(Chenet al\.,[2021](https://arxiv.org/html/2606.19658#bib.bib14); Weiet al\.,[2023](https://arxiv.org/html/2606.19658#bib.bib29)\), we adopted Precision@N, Recall@N, and NDCG@N to evaluate the performance of methods\. By default, we set N = 20 and reported the average values of the three metrics for all users in the testing set\.

Table 2\.Performance comparison of baselines on different datasets in terms of Recall@20 and NDCG@20\. The best performance is highlighted inboldand the second to best is highlighted byunderlines\.Base ModelMethodAmazon\-SportsAmazon\-BabyTiktokRecall@20NDCG@20Recall@20NDCG@20Recall@20NDCG@20ColdNeuMFNormal0\.002440\.002280\.001420\.001570\.001040\.00010RINCE0\.002710\.002220\.001970\.001720\.006740\.00298DECL0\.002660\.001910\.001860\.001490\.006220\.00234MWUF0\.003020\.002500\.001660\.001580\.000520\.00013Ours0\.004610\.002590\.003850\.001880\.008810\.00227LightGCNNormal0\.001910\.001920\.001430\.001220\.001060\.00026RINCE0\.002660\.002940\.002580\.002480\.004150\.00125DECL0\.002770\.003060\.001790\.001680\.002310\.00065MWUF0\.002530\.002220\.001860\.001610\.001550\.00050Ours0\.003250\.003410\.003030\.002610\.005180\.00117SimGCLNormal0\.002180\.002100\.001800\.001840\.001550\.00040RINCE0\.002930\.003120\.002850\.002890\.004660\.00101DECL0\.003170\.003610\.002190\.001840\.002590\.00072MWUF0\.003190\.002700\.002360\.002270\.002070\.00060Ours0\.003470\.003790\.003500\.003560\.005700\.00154WarmNeuMFNormal0\.045350\.021340\.055340\.024560\.100950\.04033RINCE0\.066190\.027550\.054350\.022200\.096900\.03936DECL0\.060730\.025950\.054870\.022360\.086670\.03324MWUF0\.045940\.020580\.057680\.024890\.111190\.04685Ours0\.062810\.028480\.059800\.028320\.110950\.04440LightGCNNormal0\.063810\.027690\.058820\.024380\.094290\.04429RINCE0\.096950\.043140\.088520\.038380\.121190\.05403DECL0\.099370\.044360\.062500\.026320\.115000\.04423MWUF0\.069530\.030290\.063160\.026960\.110240\.04769Ours0\.108620\.048690\.094940\.042570\.126670\.05134SimGCLNormal0\.065500\.029720\.059340\.025460\.095710\.04785RINCE0\.098150\.044440\.093160\.041320\.121670\.05183DECL0\.107490\.049430\.061770\.026090\.098330\.04649MWUF0\.072350\.032270\.061890\.026140\.125240\.05732Ours0\.114100\.053120\.100980\.044270\.128810\.05758OverallNeuMFNormal0\.029320\.013750\.042610\.018910\.069490\.02763RINCE0\.042700\.017760\.041910\.017100\.068520\.02791DECL0\.039220\.016740\.042310\.017230\.061340\.02351MWUF0\.029870\.013350\.044420\.019170\.076350\.03214Ours0\.041140\.018290\.046740\.021820\.078790\.03113LightGCNNormal0\.040730\.017680\.045240\.018750\.064930\.03043RINCE0\.061890\.027540\.068090\.029520\.084340\.03741DECL0\.063440\.028320\.048070\.020240\.079120\.03039MWUF0\.044450\.019350\.048580\.020730\.076020\.03283Ours0\.069340\.031090\.073030\.032750\.088420\.03555SimGCLNormal0\.042070\.019040\.045640\.019580\.066070\.03291RINCE0\.062660\.028370\.071650\.031780\.084830\.03583DECL0\.068620\.031560\.047510\.020070\.068190\.03208MWUF0\.046410\.020660\.047600\.020110\.086460\.03946Ours0\.072870\.033920\.077670\.034050\.090050\.03993
#### 3\.1\.3\.Baseline Methods

To demonstrate the efficacy of our proposed DIF in denoising implicit feedback for cold\-start recommendation, we compare DIF with the state\-of\-the\-art model\-agnostic denoising and cold\-start methods\. In particular, 1\)DECL\(Qinet al\.,[2022](https://arxiv.org/html/2606.19658#bib.bib56)\)captures the uncertainty caused by noise and separates clean and noisy data, employing them differently for learning with noisy correspondence\. 2\)RINCE\(Chuanget al\.,[2022](https://arxiv.org/html/2606.19658#bib.bib54)\)introduces a new contrastive learning objective designed to be robust to noisy data views\. 3\)MWUF\(Zhuet al\.,[2021](https://arxiv.org/html/2606.19658#bib.bib24)\)proposes meta\-scaling and shifting networks to generate initial embedding representations for cold\-start items within the warm feature space, demonstrating greater stability from noisy samples\. We implement DIF and the aforementioned model\-agnostic baselines to three representative backend models\. 1\)NeuMF\(Heet al\.,[2017](https://arxiv.org/html/2606.19658#bib.bib34)\)models the relationship between users and items by combining GMF and a MLP\. 2\)LightGCN\(Heet al\.,[2020](https://arxiv.org/html/2606.19658#bib.bib35)\)leverages high\-order neighbors information to enhance the user and item representations\. 3\)SimGCL\(Yuet al\.,[2022](https://arxiv.org/html/2606.19658#bib.bib55)\)simplifies the contrastive loss and optimizes uniformity representations\.

#### 3\.1\.4\.Implementation Details

We set the embedding dimensionddfixed to 64 for all models\. We optimize all models with the Adam\(Kingma and Ba,[2014](https://arxiv.org/html/2606.19658#bib.bib32)\)optimizer, where the batch size is fixed at 1024\. We set the learning rate to 0\.001, the number of similar warm itemskkto 5, and the temperature coefficientτ\\tauto 0\.3\. We use the Xavier initializer\(Glorot and Bengio,[2010](https://arxiv.org/html/2606.19658#bib.bib33)\), and distinguish warm items and cold items based on the 20th percentile of interaction counts for each dataset and determine the specific values ofα\\alphaandγ\\gammain uncertainty estimation\.

### 3\.2\.Performance Comparison

We conduct comprehensive experiments in natural noise setting to compare DIF’s performance with other referenced baselines\. The results, including on cold, warm, and overall items, are illustrated in Table[2](https://arxiv.org/html/2606.19658#S3.T2), yields several key observations:

- •All the denoising methods show better results compared with normal training, especially in the improvement of performance on cold items, which indicates the necessity of denoising implicit feedback for cold\-start recommendations\.
- •Jointly analyzing the performance of different backbones in recommender systems, we observe that NeuMF demonstrates more significant performance on cold items compared to graph\-based methods such as LightGCN and SimGCL\. This advantage arises because graph\-based methods rely heavily on the propagation of collaborative signals along edges, making them more effective in data\-rich scenarios\. Nevertheless, our method is model\-agnostic and achieves the best results across various backbones on cold items, which demonstrates the superior scalability of DIF\.
- •DECL and RINCE, as general denoising methods directly transferred to recommendation tasks, improve performance on both cold and warm items, though the gains are not substantial\. MWUF performs reasonably well on cold items but shows limited improvement on warm items and overall performance\. In contrast, our method focuses on denoising implicit feedback for cold items without causing unacceptable impacts on warm items or overall performance, even achieving performance enhancements\.
- •Our DIF does not achieve the optimal NDCG on cold items for Tiktok\. This is partly due to the small and sparser nature of the TikTok\. Moreover, our primary focus is on retrieve optimization rather than ranking, which remains one of future directions\.
- •The proposed DIF effectively enhances the performance of all base models and outperforms most denoising methods across three datasets\. We attribute these improvements to DIF’s content\-based denoising approach using warm items for cold items: \(1\) By modeling the confidence of multiple pseudo\-labels, DIF further improves the accuracy of aggregated pseudo\-labels based on content similarity scores\. \(2\) Benefiting from our uncertainty estimation, DIF accurately utilizes pseudo\-labels and guides the sample label correction process\.

### 3\.3\.Ablation Study

Table 3\.Ablation study on key components of DIF\.DatasetsMethodColdWarmOverallTiktok𝚠/𝚘\\mathtt\{w/o\}CM0\.003110\.108570\.07537𝚠/𝚘\\mathtt\{w/o\}RE0\.001520\.122140\.08385𝚠/𝚘\\mathtt\{w/o\}CS0\.001040\.106190\.07308DIF0\.005180\.126670\.08842Sports𝚠/𝚘\\mathtt\{w/o\}CM0\.002930\.098040\.06259𝚠/𝚘\\mathtt\{w/o\}RE0\.002780\.096850\.06182𝚠/𝚘\\mathtt\{w/o\}CS0\.002640\.093180\.05954DIF0\.003250\.108620\.06934To evaluate the effectiveness of each component in our method, we perform ablation studies on Tiktok and Amazon Sports using LightGCN as the backbone\. We present the results of Recall@20 in Table[3](https://arxiv.org/html/2606.19658#S3.T3)\. We can see that:

- •Without the Confidence Modeling \(𝚠/𝚘\\mathtt\{w/o\}CM\), we aggregate multiple pseudo\-labels for each sample using average pooling, without considering content similarity distinctions\. This variant has minimal impact on the performance of cold items, suggesting that current multi\-modal representations are already highly reliable, with well\-guaranteed similarity among the retrieved warm items\.
- •We ablate the relative entropy component from the uncertainty estimation with the variant𝚠/𝚘\\mathtt\{w/o\}RE, which significantly affects cold items but has a limited impact on warm items\. High relative entropy indicates a greater likelihood of label noise\. However, since the collaborative representations of warm items are already well\-trained, situations with high relative entropy are rare\.
- •We make another comparison between DIF and the variant \(𝚠/𝚘\\mathtt\{w/o\}CS\) without consideration of cold\-start status, which demonstrates the most significant impact for cold items\. This can be attributed to two factors: First, the failure to account for the varying probabilities of label noise across different cold\-start states\. Second, the potential disruption to representation learning for warm items, which may subsequently affect user modeling and ultimately influence recommendations for cold items\.

### 3\.4\.Noise Robustness

![Refer to caption](https://arxiv.org/html/2606.19658v1/x3.png)Figure 3\.Performance comparison of denoise training with random noises in Amazon datasets\.We conduct random noisy training on Amazon datasets using the backbones of NeuMF and LightGCN to evaluate the noise resistance capability of DIF, comparing it with the three competitive model\-agnostic methods, DECL, RINCE and MWUF\. The proportion of noise in our training settings spanned from 20% to 80%\. We report the results on cold items in Figure[3](https://arxiv.org/html/2606.19658#S3.F3)\. Similar results are seen with Tiktok and other backbones, but figures are omitted for brevity\. The results show that: 1\) As the noise ratio increases, the performance of DECL, RINCE, MWUF, and DIF generally declines\. This decline is attributed to the intensifying corruption of data due to the escalating noise level, making the representation learning for cold\-start items and the modeling of user preferences increasingly challenging\. 2\) DIF consistently outperforms all other denoising methods across different noise ratio settings\. This highlights the remarkable noise robustness of DIF in cold\-start scenarios, which can be attributed to the effectiveness of its strategy of generating pseudo\-labels through content\-similar warm items\.

### 3\.5\.Online A/B Testing

Table 4\.The results of A/B testing in online scenario\.Effective ViewWatch TimeLong ViewShort View\+2\.327%\+2\.921%\+2\.688%\-2\.030%LikeCommentFollowShare\+2\.921%\+2\.435%\+2\.790%\+0\.876%We carried out rigorous online A/B testing in our short\-video streaming scenario from Jan\. 12, 2025, to Jan\. 16, 2025, with hundreds of millions of users per day\. The results of the online A/B test for cold\-start recommendation are shown in Table[4](https://arxiv.org/html/2606.19658#S3.T4), we focus on several commercial metrics, which can be divided into view\-related and action\-related\. For view\-related metrics, we introduce metrics such aseffective view,watch time,long view, andshort view\. We define an “effective view” label as user watches a video longer than a threshold \(e\.g\., 5 seconds\), and videos in different duration intervals have different thresholds\. Meanwhile, the “long view” label follows stricter criterias compared to “effective view”\. For action\-related metrics, includelike,comment,follow, andshare\. For example, a “like” is defined as the user likes current video by clicking the like button or double tapping/long pressing screen\. For company privacy, we cannot report the real metrics of the original online models\. Instead, we report the performance gain ratio improved by our approach DIF\. It is worth noting that one percent improvement ratio usually indicates a large improvement of the recommendation capacity in real\-world application scenario, when tested on a large population of users\. By effectively denoising cold\-start item samples using our method, both view\-related and action\-related metrics show significant improvement, contributing to a substantial enhancement in the recommendation efficiency for cold\-start items\.

## 4\.Related Works

Cold\-start Recommendation\.Cold\-start is one of the main challenges in recommender systems\. Specifically, we focus on the more challenging scenario where cold\-start items are newly uploaded and lack any user feedback\. The common solution to this issue can be categorized into two types, namely content\-based and transfer learning based methods\. The first type of methods\(Weiet al\.,[2023](https://arxiv.org/html/2606.19658#bib.bib29); Chenet al\.,[2024](https://arxiv.org/html/2606.19658#bib.bib15); Wanget al\.,[2024](https://arxiv.org/html/2606.19658#bib.bib66); Jeonet al\.,[2024](https://arxiv.org/html/2606.19658#bib.bib67); Liuet al\.,[2024](https://arxiv.org/html/2606.19658#bib.bib70)\)aims to exploit side information, especially content features of items to compensate the absence of collaborative signals\. The core idea is to approximate the well\-trained collaborative embeddings via content information by modeling their correlation\. Another way to alleviate the cold\-start problem is to transfer knowledge from other domains, such as cross\-domain recommendation\(Caoet al\.,[2022a](https://arxiv.org/html/2606.19658#bib.bib61); Zanget al\.,[2022](https://arxiv.org/html/2606.19658#bib.bib21); Chenet al\.,[2023a](https://arxiv.org/html/2606.19658#bib.bib62)\), transfer learning methods\(Zhanget al\.,[2021](https://arxiv.org/html/2606.19658#bib.bib63); Shenget al\.,[2021](https://arxiv.org/html/2606.19658#bib.bib22); Zhanget al\.,[2023a](https://arxiv.org/html/2606.19658#bib.bib64)\), meta\-learning methods\(Donget al\.,[2020](https://arxiv.org/html/2606.19658#bib.bib23); Zhuet al\.,[2021](https://arxiv.org/html/2606.19658#bib.bib24)\), and prompt learning\(Jianget al\.,[2024](https://arxiv.org/html/2606.19658#bib.bib17); Kusano,[2024](https://arxiv.org/html/2606.19658#bib.bib65)\)\. However, few existing studies on cold\-start recommendation recognize the importance of implicit feedback denoising and design practical deployment solutions\.

Denoising Recommendation\.Existing recommender systems are typically trained with implicit feedback\. Recently, some studies\(Wanget al\.,[2021a](https://arxiv.org/html/2606.19658#bib.bib45),[b](https://arxiv.org/html/2606.19658#bib.bib42); Zhaoet al\.,[2024](https://arxiv.org/html/2606.19658#bib.bib48); Wanget al\.,[2023b](https://arxiv.org/html/2606.19658#bib.bib40)\)have noticed that implicit feedback could be easily corrupted by different factors \(e\.g\., popularity bias\(Chenet al\.,[2023b](https://arxiv.org/html/2606.19658#bib.bib57)\)and user unconscious behaviors\(Huet al\.,[2008](https://arxiv.org/html/2606.19658#bib.bib58)\)\), and the inevitable noise would dramatically degrade the recommendation performance\. As a result, some efforts have been dedicated to solving the noisy implicit feedback problem, which can be categorized into sample selection methods\(Dinget al\.,[2019](https://arxiv.org/html/2606.19658#bib.bib59); Park and Chang,[2019](https://arxiv.org/html/2606.19658#bib.bib60); Yu and Qin,[2020](https://arxiv.org/html/2606.19658#bib.bib43); Liet al\.,[2026](https://arxiv.org/html/2606.19658#bib.bib2)\)and sample re\-weighting methods\(Huet al\.,[2021](https://arxiv.org/html/2606.19658#bib.bib44); Wanget al\.,[2022](https://arxiv.org/html/2606.19658#bib.bib41)\)\. The sample selection methods aim to select clean and informative samples, which exhibit high performance variation since they heavily depend on the sampling distribution\. Sample re\-weighting methods focus on assigning lower weights to inaccurate interactions by the learning process signals of models \(e\.g\., loss values and predictions\)\. However, these methods heavily rely on predefined heuristic assumptions, resulting in poor generalization for cold\-start models\.

## 5\.Conclusion

To address the implicit feedback denoising in cold\-start recommendation, this paper introduces a model\-agnostic approach called DIF, which can be flexibly applied to various online models\. Inspired by the item collaborative filtering, we assign pseudo\-labels to cold\-start items through the user and warm items with similar content, and we present a detailed industrial implement solution\. We further model the confidence of each pseudo\-label based on content similarity to improve the accuracy of the final generated pseudo label\. Furthermore, we estimate the uncertainty of the noisy sample label to guide the participation of the pseudo\-label during the label correction process, thereby avoiding overcorrection\. We share the successful experience of deploying DIF on Kuaishou\. Extensive offline experiments and online A/B testing further verify its effectiveness in cold\-start recommendation tasks\.

## References

- E\. Arazo, D\. Ortego, P\. Albert, N\. O’Connor, and K\. McGuinness \(2019\)Unsupervised label noise modeling and loss correction\.InInternational conference on machine learning,pp\. 312–321\.Cited by:[§2\.2](https://arxiv.org/html/2606.19658#S2.SS2.p1.2)\.
- D\. Bahri, H\. Jiang, and M\. Gupta \(2020\)Deep k\-nn for noisy labels\.InInternational Conference on Machine Learning,pp\. 540–550\.Cited by:[§2\.8](https://arxiv.org/html/2606.19658#S2.SS8.p5.1)\.
- S\. Boucheron, G\. Lugosi, and P\. Massart \(2013\)Concentration inequalities \- a nonasymptotic theory of independence\.OUP Oxford\.Cited by:[§2\.8](https://arxiv.org/html/2606.19658#S2.SS8.3.p3.2)\.
- Q\. Cai, Z\. Xue, C\. Zhang, W\. Xue, S\. Liu, R\. Zhan, X\. Wang, T\. Zuo, W\. Xie, D\. Zheng,et al\.\(2023\)Two\-stage constrained actor\-critic for short video recommendation\.InProceedings of the ACM Web Conference 2023,pp\. 865–875\.Cited by:[§1](https://arxiv.org/html/2606.19658#S1.p1.1)\.
- J\. Cao, J\. Sheng, X\. Cong, T\. Liu, and B\. Wang \(2022a\)Cross\-domain recommendation to cold\-start users via variational information bottleneck\.In2022 IEEE 38th International Conference on Data Engineering \(ICDE\),pp\. 2209–2223\.Cited by:[§4](https://arxiv.org/html/2606.19658#S4.p1.1)\.
- Y\. Cao, S\. Hu, Y\. Gong, Z\. Li, Y\. Yang, Q\. Liu, and S\. Ji \(2022b\)Gift: graph\-guided feature transfer for cold\-start video click\-through rate prediction\.InProceedings of the 31st ACM International Conference on Information & Knowledge Management,pp\. 2964–2973\.Cited by:[§1](https://arxiv.org/html/2606.19658#S1.p2.1)\.
- G\. Chen, R\. Sun, Y\. Jiang, J\. Cao, Q\. Zhang, J\. Lin, H\. Li, K\. Gai, and X\. Zhang \(2024\)A multi\-modal modeling framework for cold\-start short\-video recommendation\.InProceedings of the 18th ACM Conference on Recommender Systems,pp\. 391–400\.Cited by:[§3\.1\.1](https://arxiv.org/html/2606.19658#S3.SS1.SSS1.p1.1),[§4](https://arxiv.org/html/2606.19658#S4.p1.1)\.
- G\. Chen, X\. Zhang, Y\. Su, Y\. Lai, J\. Xiang, J\. Zhang, and Y\. Zheng \(2023a\)Win\-win: a privacy\-preserving federated framework for dual\-target cross\-domain recommendation\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.37,pp\. 4149–4156\.Cited by:[§4](https://arxiv.org/html/2606.19658#S4.p1.1)\.
- G\. Chen, X\. Zhang, Y\. Zhao, C\. Xue, and J\. Xiang \(2021\)Exploring periodicity and interactivity in multi\-interest framework for sequential recommendation\.InProceedings of the Thirtieth International Joint Conference on Artificial Intelligence,pp\. 1426–1433\.Cited by:[§1](https://arxiv.org/html/2606.19658#S1.p1.1),[§3\.1\.2](https://arxiv.org/html/2606.19658#S3.SS1.SSS2.p1.1)\.
- J\. Chen, H\. Dong, X\. Wang, F\. Feng, M\. Wang, and X\. He \(2023b\)Bias and debias in recommender system: a survey and future directions\.ACM Transactions on Information Systems41\(3\),pp\. 1–39\.Cited by:[§4](https://arxiv.org/html/2606.19658#S4.p2.1)\.
- C\. Chuang, R\. D\. Hjelm, X\. Wang, V\. Vineet, N\. Joshi, A\. Torralba, S\. Jegelka, and Y\. Song \(2022\)Robust contrastive learning against noisy views\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 16670–16681\.Cited by:[§3\.1\.3](https://arxiv.org/html/2606.19658#S3.SS1.SSS3.p1.1)\.
- T\. M\. Cover, J\. A\. Thomas,et al\.\(1991\)Entropy, relative entropy and mutual information\.Elements of information theory2\(1\),pp\. 12–13\.Cited by:[§2\.5](https://arxiv.org/html/2606.19658#S2.SS5.p2.4)\.
- J\. Ding, G\. Yu, X\. He, F\. Feng, Y\. Li, and D\. Jin \(2019\)Sampler design for bayesian personalized ranking by leveraging view data\.IEEE transactions on knowledge and data engineering33\(2\),pp\. 667–681\.Cited by:[§4](https://arxiv.org/html/2606.19658#S4.p2.1)\.
- M\. Dong, F\. Yuan, L\. Yao, X\. Xu, and L\. Zhu \(2020\)Mamo: memory\-augmented meta\-optimization for cold\-start recommendation\.InProceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining,pp\. 688–697\.Cited by:[§4](https://arxiv.org/html/2606.19658#S4.p1.1)\.
- W\. Gao, X\. Niu, and Z\. Zhou \(2018\)On the consistency of exact and approximate nearest neighbor with noisy data\.Arxiv, abs/1607\.07526\.Cited by:[§2\.8](https://arxiv.org/html/2606.19658#S2.SS8.p5.1)\.
- Y\. Gao, Y\. Du, Y\. Hu, L\. Chen, X\. Zhu, Z\. Fang, and B\. Zheng \(2022\)Self\-guided learning to denoise for robust recommendation\.InProceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval,pp\. 1412–1422\.Cited by:[§1](https://arxiv.org/html/2606.19658#S1.p3.1),[§1](https://arxiv.org/html/2606.19658#S1.p5.1)\.
- X\. Glorot and Y\. Bengio \(2010\)Understanding the difficulty of training deep feedforward neural networks\.InProceedings of the thirteenth international conference on artificial intelligence and statistics,pp\. 249–256\.Cited by:[§3\.1\.4](https://arxiv.org/html/2606.19658#S3.SS1.SSS4.p1.5)\.
- X\. Gong, Q\. Feng, Y\. Zhang, J\. Qin, W\. Ding, B\. Li, P\. Jiang, and K\. Gai \(2022\)Real\-time short video recommendation on mobile devices\.InProceedings of the 31st ACM international conference on information & knowledge management,pp\. 3103–3112\.Cited by:[§1](https://arxiv.org/html/2606.19658#S1.p1.1),[§2\.1](https://arxiv.org/html/2606.19658#S2.SS1.p2.8)\.
- L\. Györfi, M\. Kohler, A\. Krzyżak, and H\. Walk \(2002\)A distribution\-free theory of nonparametric regression\.Springer\.Cited by:[§2\.8](https://arxiv.org/html/2606.19658#S2.SS8.2.p2.6),[§2\.8](https://arxiv.org/html/2606.19658#S2.SS8.p5.1)\.
- X\. He, K\. Deng, X\. Wang, Y\. Li, Y\. Zhang, and M\. Wang \(2020\)Lightgcn: simplifying and powering graph convolution network for recommendation\.InProceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval,pp\. 639–648\.Cited by:[§3\.1\.3](https://arxiv.org/html/2606.19658#S3.SS1.SSS3.p1.1)\.
- X\. He, L\. Liao, H\. Zhang, L\. Nie, X\. Hu, and T\. Chua \(2017\)Neural collaborative filtering\.InProceedings of the 26th international conference on world wide web,pp\. 173–182\.Cited by:[§3\.1\.3](https://arxiv.org/html/2606.19658#S3.SS1.SSS3.p1.1)\.
- Z\. He, Y\. Wang, Y\. Yang, P\. Sun, L\. Wu, H\. Bai, J\. Gong, R\. Hong, and M\. Zhang \(2024\)Double correction framework for denoising recommendation\.InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining,pp\. 1062–1072\.Cited by:[§1](https://arxiv.org/html/2606.19658#S1.p5.1)\.
- K\. Hu, L\. Li, Q\. Xie, J\. Liu, and X\. Tao \(2021\)What is next when sequential prediction meets implicitly hard interaction?\.InProceedings of the 30th ACM International Conference on Information & Knowledge Management,pp\. 710–719\.Cited by:[§1](https://arxiv.org/html/2606.19658#S1.p5.1),[§4](https://arxiv.org/html/2606.19658#S4.p2.1)\.
- Y\. Hu, Y\. Koren, and C\. Volinsky \(2008\)Collaborative filtering for implicit feedback datasets\.In2008 Eighth IEEE international conference on data mining,pp\. 263–272\.Cited by:[§4](https://arxiv.org/html/2606.19658#S4.p2.1)\.
- H\. Jeon, J\. Lee, J\. Yun, and U\. Kang \(2024\)Cold\-start bundle recommendation via popularity\-based coalescence and curriculum heating\.InProceedings of the ACM Web Conference 2024,pp\. 3277–3286\.Cited by:[§4](https://arxiv.org/html/2606.19658#S4.p1.1)\.
- Y\. Jiang, G\. Chen, W\. Zhang, J\. Wang, Y\. Jiang, Q\. Zhang, J\. Lin, P\. Jiang, and K\. Bian \(2024\)Prompt tuning for item cold\-start recommendation\.InProceedings of the 18th ACM Conference on Recommender Systems,pp\. 411–421\.Cited by:[§4](https://arxiv.org/html/2606.19658#S4.p1.1)\.
- J\. Johnson, M\. Douze, and H\. Jégou \(2019\)Billion\-scale similarity search with gpus\.IEEE Transactions on Big Data7\(3\),pp\. 535–547\.Cited by:[§2\.1](https://arxiv.org/html/2606.19658#S2.SS1.p2.8)\.
- D\. P\. Kingma and J\. Ba \(2014\)Adam: a method for stochastic optimization\.arXiv preprint arXiv:1412\.6980\.Cited by:[§3\.1\.4](https://arxiv.org/html/2606.19658#S3.SS1.SSS4.p1.5)\.
- G\. Kusano \(2024\)Data augmentation using reverse prompt for cost\-efficient cold\-start recommendation\.InProceedings of the 18th ACM Conference on Recommender Systems,pp\. 861–865\.Cited by:[§4](https://arxiv.org/html/2606.19658#S4.p1.1)\.
- S\. Li, Z\. Yang, S\. Li, X\. Xia, H\. Liu, X\. Zhang, G\. Chen, D\. Fang, Y\. Tai, and Z\. Peng \(2026\)LearnAlign: data selection for llm reinforcement learning with improved gradient alignment\.InFindings of the Association for Computational Linguistics: ACL 2026,Cited by:[§4](https://arxiv.org/html/2606.19658#S4.p2.1)\.
- R\. Liu, H\. Chen, Y\. Bei, Q\. Shen, F\. Zhong, S\. Wang, and J\. Wang \(2024\)Fine tuning out\-of\-vocabulary item recommendation with user sequence imagination\.Advances in Neural Information Processing Systems37,pp\. 8930–8955\.Cited by:[§4](https://arxiv.org/html/2606.19658#S4.p1.1)\.
- D\. H\. Park and Y\. Chang \(2019\)Adversarial sampling and training for semi\-supervised information retrieval\.InThe World Wide Web Conference,pp\. 1443–1453\.Cited by:[§4](https://arxiv.org/html/2606.19658#S4.p2.1)\.
- Q\. Pi, G\. Zhou, Y\. Zhang, Z\. Wang, L\. Ren, Y\. Fan, X\. Zhu, and K\. Gai \(2020\)Search\-based user interest modeling with lifelong sequential behavior data for click\-through rate prediction\.InProceedings of the 29th ACM International Conference on Information & Knowledge Management,pp\. 2685–2692\.Cited by:[§1](https://arxiv.org/html/2606.19658#S1.p1.1)\.
- Y\. Qin, D\. Peng, X\. Peng, X\. Wang, and P\. Hu \(2022\)Deep evidential learning with noisy correspondence for cross\-modal retrieval\.InProceedings of the 30th ACM International Conference on Multimedia,pp\. 4948–4956\.Cited by:[§3\.1\.3](https://arxiv.org/html/2606.19658#S3.SS1.SSS3.p1.1)\.
- N\. Reimers and I\. Gurevych \(2019\)Sentence\-bert: sentence embeddings using siamese bert\-networks\.InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing \(EMNLP\-IJCNLP\),pp\. 3982–3992\.Cited by:[1st item](https://arxiv.org/html/2606.19658#S3.I1.i1.p1.1)\.
- X\. Sheng, L\. Zhao, G\. Zhou, X\. Ding, B\. Dai, Q\. Luo, S\. Yang, J\. Lv, C\. Zhang, H\. Deng,et al\.\(2021\)One model to serve all: star topology adaptive recommender for multi\-domain ctr prediction\.InProceedings of the 30th ACM International Conference on Information & Knowledge Management,pp\. 4104–4113\.Cited by:[§4](https://arxiv.org/html/2606.19658#S4.p1.1)\.
- D\. Tanaka, D\. Ikami, T\. Yamasaki, and K\. Aizawa \(2018\)Joint optimization framework for learning with noisy labels\.InProceedings of the IEEE conference on computer vision and pattern recognition,pp\. 5552–5560\.Cited by:[§2\.2](https://arxiv.org/html/2606.19658#S2.SS2.p1.2)\.
- J\. Wang, H\. Lu, J\. Caverlee, E\. H\. Chi, and M\. Chen \(2024\)Large language models as data augmenters for cold\-start item recommendation\.InCompanion Proceedings of the ACM Web Conference 2024,pp\. 726–729\.Cited by:[§4](https://arxiv.org/html/2606.19658#S4.p1.1)\.
- J\. Wang, H\. Lu, S\. Zhang, B\. Locanthi, H\. Wang, D\. Greaves, B\. Lipshitz, S\. Badam, E\. H\. Chi, C\. J\. Goodrow,et al\.\(2023a\)Fresh content needs more attention: multi\-funnel fresh content recommendation\.InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining,pp\. 5082–5091\.Cited by:[§1](https://arxiv.org/html/2606.19658#S1.p2.1)\.
- W\. Wang, F\. Feng, X\. He, L\. Nie, and T\. Chua \(2021a\)Denoising implicit feedback for recommendation\.InProceedings of the 14th ACM international conference on web search and data mining,pp\. 373–381\.Cited by:[§1](https://arxiv.org/html/2606.19658#S1.p5.1),[§4](https://arxiv.org/html/2606.19658#S4.p2.1)\.
- Y\. Wang, X\. Xin, Z\. Meng, J\. M\. Jose, F\. Feng, and X\. He \(2022\)Learning robust recommenders through cross\-model agreement\.InProceedings of the ACM Web Conference 2022,pp\. 2015–2025\.Cited by:[§1](https://arxiv.org/html/2606.19658#S1.p5.1),[§4](https://arxiv.org/html/2606.19658#S4.p2.1)\.
- Z\. Wang, Q\. Xu, Z\. Yang, X\. Cao, and Q\. Huang \(2021b\)Implicit feedbacks are not always favorable: iterative relabeled one\-class collaborative filtering against noisy interactions\.InProceedings of the 29th ACM International Conference on Multimedia,pp\. 3070–3078\.Cited by:[§4](https://arxiv.org/html/2606.19658#S4.p2.1)\.
- Z\. Wang, M\. Gao, W\. Li, J\. Yu, L\. Guo, and H\. Yin \(2023b\)Efficient bi\-level optimization for recommendation denoising\.InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining,pp\. 2502–2511\.Cited by:[§1](https://arxiv.org/html/2606.19658#S1.p5.1),[§4](https://arxiv.org/html/2606.19658#S4.p2.1)\.
- W\. Wei, C\. Huang, L\. Xia, and C\. Zhang \(2023\)Multi\-modal self\-supervised learning for recommendation\.InProceedings of the ACM Web Conference 2023,pp\. 790–800\.Cited by:[§3\.1\.2](https://arxiv.org/html/2606.19658#S3.SS1.SSS2.p1.1),[§4](https://arxiv.org/html/2606.19658#S4.p1.1)\.
- J\. Yang, X\. Yi, D\. Zhiyuan Cheng, L\. Hong, Y\. Li, S\. Xiaoming Wang, T\. Xu, and E\. H\. Chi \(2020\)Mixed negative sampling for learning two\-tower neural networks in recommendations\.InCompanion proceedings of the web conference 2020,pp\. 441–447\.Cited by:[§2\.1](https://arxiv.org/html/2606.19658#S2.SS1.p2.8)\.
- J\. Yu, H\. Yin, X\. Xia, T\. Chen, L\. Cui, and Q\. V\. H\. Nguyen \(2022\)Are graph augmentations necessary? simple graph contrastive learning for recommendation\.InProceedings of the 45th international ACM SIGIR conference on research and development in information retrieval,pp\. 1294–1303\.Cited by:[§3\.1\.3](https://arxiv.org/html/2606.19658#S3.SS1.SSS3.p1.1)\.
- W\. Yu and Z\. Qin \(2020\)Sampler design for implicit feedback data by noisy\-label robust learning\.InProceedings of the 43rd international ACM SIGIR conference on research and development in information retrieval,pp\. 861–870\.Cited by:[§4](https://arxiv.org/html/2606.19658#S4.p2.1)\.
- T\. Zang, Y\. Zhu, H\. Liu, R\. Zhang, and J\. Yu \(2022\)A survey on cross\-domain recommendation: taxonomies, methods, and future directions\.ACM Transactions on Information Systems41\(2\),pp\. 1–39\.Cited by:[§4](https://arxiv.org/html/2606.19658#S4.p1.1)\.
- S\. Zhang, L\. Yao, A\. Sun, and Y\. Tay \(2019\)Deep learning based recommender system: a survey and new perspectives\.ACM computing surveys \(CSUR\)52\(1\),pp\. 1–38\.Cited by:[§1](https://arxiv.org/html/2606.19658#S1.p1.1)\.
- Y\. Zhang, D\. Z\. Cheng, T\. Yao, X\. Yi, L\. Hong, and E\. H\. Chi \(2021\)A model of two tales: dual transfer learning framework for improved long\-tail item recommendation\.InProceedings of the web conference 2021,pp\. 2220–2231\.Cited by:[§4](https://arxiv.org/html/2606.19658#S4.p1.1)\.
- Y\. Zhang, R\. Wang, D\. Z\. Cheng, T\. Yao, X\. Yi, L\. Hong, J\. Caverlee, and E\. H\. Chi \(2023a\)Empowering long\-tail item recommendation through cross decoupling network \(cdn\)\.InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining,pp\. 5608–5617\.Cited by:[§4](https://arxiv.org/html/2606.19658#S4.p1.1)\.
- Y\. Zhang, X\. Dong, W\. Ding, B\. Li, P\. Jiang, and K\. Gai \(2023b\)Divide and conquer: towards better embedding\-based retrieval for recommender systems from a multi\-task perspective\.InCompanion Proceedings of the ACM Web Conference 2023,pp\. 366–370\.Cited by:[§2\.1](https://arxiv.org/html/2606.19658#S2.SS1.p2.8)\.
- J\. Zhao, W\. Wenjie, Y\. Xu, T\. Sun, F\. Feng, and T\. Chua \(2024\)Denoising diffusion recommender model\.InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval,pp\. 1370–1379\.Cited by:[§1](https://arxiv.org/html/2606.19658#S1.p3.1),[§4](https://arxiv.org/html/2606.19658#S4.p2.1)\.
- G\. Zhou, X\. Zhu, C\. Song, Y\. Fan, H\. Zhu, X\. Ma, Y\. Yan, J\. Jin, H\. Li, and K\. Gai \(2018\)Deep interest network for click\-through rate prediction\.InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining,pp\. 1059–1068\.Cited by:[§1](https://arxiv.org/html/2606.19658#S1.p1.1),[§2\.1](https://arxiv.org/html/2606.19658#S2.SS1.p2.8)\.
- Z\. Zhou, L\. Zhang, and N\. Yang \(2023\)Contrastive collaborative filtering for cold\-start item recommendation\.InProceedings of the ACM Web Conference 2023,pp\. 928–937\.Cited by:[§1](https://arxiv.org/html/2606.19658#S1.p2.1)\.
- Y\. Zhu, R\. Xie, F\. Zhuang, K\. Ge, Y\. Sun, X\. Zhang, L\. Lin, and J\. Cao \(2021\)Learning to warm up cold item embeddings for cold\-start recommendation with meta scaling and shifting networks\.InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval,pp\. 1167–1176\.Cited by:[§3\.1\.3](https://arxiv.org/html/2606.19658#S3.SS1.SSS3.p1.1),[§4](https://arxiv.org/html/2606.19658#S4.p1.1)\.