In computational advertising, predicting the post-click conversion rate (CVR) using deep neural networks (DNNs) benefits a lot from privileged features, which can be collected for offline training but are unavailable during online serving. To utilize these privileged signals, privileged features distillation (PFD) methods incorporate a teacher model with privileged features to guide the CVR model. However, existing PFD approaches fail to put more emphasis on poorly predicted instances where the teacher's guidance is most crucial, and thus suffer from overconfidence on ''easy'' instances. In this work, we propose Hardness-aware Privileged Features Distillation (HA-PFD) for enhancing CVR prediction in real-world advertising recommender systems. We specifically design focal-style distillation losses that adaptively adjust the weight of each instance based on its ''hardness''. This method prioritizes poorly predicted instances during the distillation process, resulting in improved ranking performance and better model calibration. Additionally, we incorporate latent-level distillation into the PFD framework for the first time, which facilitates the student's representation learning through a straightforward layer alignment approach. We also propose a method for selecting privileged features based on their relevance to the conversion label. We conduct extensive offline experiments on large-scale, real-world datasets and online experiments on Douyin, a short video platform with billions of live users. In the offline evaluation, HA-PFD exhibits competitive performance and superior model calibration compared to existing state-of-the-art methods. In the online experiments, HA-PFD significantly improves advertiser value and conversions. Now we have deployed HA-PFD as the main online serving model on our short video platform.
论文检索
输入标题、作者或关键词，从 844 篇学术成果中精准定位
Relevance plays a crucial role in commercial search engines by identifying documents related to user queries and fulfilling their search needs. Traditional approaches employ encoder-only models like BERT, which process concatenated query-document pairs to predict relevance scores. While autoregressive large language models (LLMs) have revolutionized numerous NLP domains, their direct application to web-scale search systems presents significant challenges. On one hand, the relevance modeling capabilities of LLMs have not been fully explored. On the other, the high computational costs and inference times make deploying LLMs in online search systems, which demand extremely low latency, nearly impossible. In this work, we address these challenges through two key contributions. First, we develop a comprehensive evaluation framework to systematically assess the effectiveness of LLMs in query-document relevance ranking. By conducting assessment experiments to LLMs in four perspectives: ranking objectives, model size, domain-specific continuous pre-training, and the integration of prior knowledge, we identify the best resource allocation strategy given a restricted budget and develop practical LLMs in a more efficient way. Second, we propose a novel framework to transfer the capabilities of LLMs in the ranking aspect to existing BERT models to avoid directly deploying LLMs. Finally, to fully leverage the improvements in relevance ranking brought by LLMs, we successfully nearline deploy LLMs in Tencent QQ Browser search engine using query-based on-demand computing and quantization. Experiments on real-world datasets and online A/B tests demonstrate that our approach significantly enhances search engine performance while maintaining practical operational efficiency. Our findings provide actionable insights for integrating LLMs into production search engines.
Unified Game Moderation: Soft-Prompting and LLM-Assisted Label Transfer for Resource-Efficient Toxicity Detection
PDF ↗Toxicity detection in gaming communities faces significant scaling challenges when expanding across multiple games and languages, particularly in real-time environments where computational efficiency is crucial. We present two key findings to address these challenges while building upon our previous work on ToxBuster, a BERT-based real-time toxicity detection system. First, we introduce a soft-prompting approach that enables a single model to effectively handle multiple games by incorporating game-context tokens, matching the performance of more complex methods like curriculum learning while offering superior scalability. Second, we develop an LLM-assisted label transfer framework using GPT-4o-mini to extend support to seven additional languages. Evaluations on real game chat data across French, German, Portuguese, and Russian achieve macro F1-scores ranging from 32.96% to 58.88%, with particularly strong performance in German, surpassing the English benchmark of 45.39%. In production, this unified approach significantly reduces computational resources and maintenance overhead compared to maintaining separate models for each game and language combination. At Ubisoft, this model successfully identifies an average of 50 players, per game, per day engaging in sanctionable behavior.
Graph-based methods have proven effective in financial fraud detection by modeling relationships between entities, yet they often fail to leverage the rich textual information present in real-world data. With the ability to understand semantic information, large language models (LLMs) offer a promising solution to enhance fraud detection by incorporating textual data, such as user profiles and transaction descriptions. However, integrating LLMs with graph-based methods introduces two key challenges: (1) the neighborhood camouflage problem, where fraudulent nodes disguise themselves within normal network structures, and (2) the input size constraints of LLMs, making it difficult to process large, complex graphs with extensive textual data. In this paper, we propose a novel framework, Fraud Detection with LLM-enhanced Graph Neural Networks (FLAG), to address these challenges. FLAG integrates LLMs with graph-based fraud detection by introducing two main modules: semantic similarity neighbor sampling, which reduces the input size and further alleviates the influence of camouflaged neighbors by selecting neighbors having high semantic similarity with the target nodes, and LLM-based node enhancement, which extracts discriminative textual features by LLM to enhance node robustness against camouflaged neighbors. To further improve the model, we design a fine-tuning approach that enables the LLM to extract discriminative text more closely related to the node labels, enhancing the model's ability to differentiate between fraudulent and normal nodes. Extensive experiments on public datasets highlight the superiority of FLAG, showing average improvements of 3.14% in F1-macro and 6.97% in AUC. Furthermore, we have deployed FLAG in Alipay's credit risk assessment system and evaluated its performance on a real-world dataset. The results indicate a 0.9% improvement in the KS criterion, further underscoring FLAG's effectiveness.
Internet live streaming is widely used in online entertainment and e-commerce, where live advertising is an important marketing tool for anchors. An advertising campaign hopes to maximize the effect (such as conversions) under constraints (such as budget and cost-per-click). The mainstream control of campaigns is auto-bidding, where the performance depends on the decision of the bidding algorithm in each request. The most widely used auto-bidding algorithms include Proportional-Integral-Derivative (PID) control, linear programming (LP), reinforcement learning (RL), etc. Existing methods either do not consider the entire time traffic, or have too high computational complexity. In this paper, the live advertising has high requirements for real-time bidding (second-level control) and faces the difficulty of unknown future traffic. Therefore, we propose a lightweight bidding algorithm Binary Constrained Bidding (BiCB), which neatly combines the optimal bidding formula given by mathematical analysis and the statistical method of future traffic estimation, and obtains good approximation to the optimal result through a low complexity solution. In addition, we complement the form of upper and lower bound constraints for traditional auto-bidding modeling and give theoretical analysis of BiCB. Sufficient offline and online experiments prove BiCB's good performance and low engineering cost.
In video recommendation systems, accurately predicting watch time is crucial for enhancing user engagement and retention. Traditional methods typically apply label transformations or mitigate duration bias to improve performance but overlook that erroneous instance representations are the primary cause of significant prediction errors. Moreover, these approaches predominantly rely on point perdition, limiting their robustness. To address these challenges, we propose LDA, a novel prediction paradigm that optimizes instance representations by explicitly leveraging label distributions as anchors within the model, enabling more accurate and robust predictions. Our analysis reveals that watch ratio across different duration groups exhibit distinct multi-peak distributions, reflecting the strong aggregation of user behavior. Based on this finding, we employ Vector Quantized Variational Auto-encoder (VQ-VAE) to convert the continuous watch ratio distribution into representative anchors that capture these multi-peak characteristics within each duration group. Subsequently, we project both instance representations and anchors into a common space and utilize Optimal Transport (OT) to generate pseudo-labels aligned with the anchor distribution, allowing instances to obtain structured coordinates within this space during training. Finally, we derive optimized instance representations for watch time prediction by aggregating anchor vectors through weighted integration. Extensive offline experiments on two datasets and large-scale online A/B testing on a short-video platform with over 300 million DAUs demonstrate the consistent superiority of LDA in watch time prediction.
In natural language processing (NLP), the focus has shifted from encoder-only tiny language models like BERT to decoder-only large language models(LLMs) such as GPT-3. However, LLMs' practical application in the financial sector has revealed three limitations: (1) LLMs often perform worse than fine-tuned BERT on discriminative tasks despite costing much higher computational resources, such as market sentiment analysis in financial reports; (2) Application on generative tasks heavily relies on retrieval augmented generation (RAG) methods to provide current and specialized information, with general retrievers showing suboptimal performance on domain-specific retrieval tasks; (3) There are additional inadequacies in other feature-based scenarios, such as topic modeling. We introduce FinBERT2, a specialized bidirectional encoder pretrained on a high-quality, financial-specific corpus of 32b tokens. This represents the largest known Chinese financial pretraining corpus for models of this parameter size. As a better backbone, FinBERT2 can bridge the gap in the financial-specific deployment of LLMs through the following achievements: (1) Discriminative fine-tuned models (Fin-Labelers) outperform other (Fin)BERT variants by 0.4%-3.3% and leading LLMs by 9.7%-12.3% on average across five financial classification tasks. (2) Contrastive fine-tuned models (Fin-Retrievers) outperform both open-source (e.g., +6.8% avg improvement over BGE-base-zh) and proprietary (e.g., +4.2% avg improvement over OpenAI's text-embedding-3-large) embedders across five financial retrieval tasks; (3) Building on FinBERT2 variants, we construct the Fin-TopicModel, which enables superior clustering and topic representation for financial titles. Our work revisits financial BERT models through comparative analysis with contemporary LLMs and offers practical insights for effectively utilizing FinBERT in the LLMs era.
Scientific research heavily depends on suitable datasets for method validation, but existing academic platforms with dataset management like PapersWithCode suffer from inefficiencies in their manual workflow. To overcome this bottleneck, we present a system, called ChatPD, that utilizes Large Language Models (LLMs) to automate dataset information extraction from academic papers and construct a structured paper-dataset network. Our system consists of three key modules: paper collection, dataset information extraction, and dataset entity resolution to construct paper-dataset networks. Specifically, we propose a Graph Completion and Inference strategy to map dataset descriptions to their corresponding entities. Through extensive experiments, we demonstrate that ChatPD not only outperforms the existing platform PapersWithCode in dataset usage extraction but also achieves about 90% precision and recall in entity resolution tasks. Moreover, we have deployed ChatPD to continuously extract which datasets are used in papers, and provide a dataset discovery service, such as task-specific dataset queries and similar dataset recommendations. We open source ChatPD and the current paper-dataset network on GitHub.
The online advertising platform serves as the intermediary between advertisers and users to facilitate the product trading process and generate profits. While the vitality of both the user and advertiser community is crucial to the long-term development of platforms, compared with the enduring efforts to improve user experience, little attention has been given to methods for attracting or retaining advertisers. To address this oversight, in this work, we investigate the first intelligent solution to the advertiser retention problem in online advertising. Motivated by advertiser surveys and data analytics, we first define a set of short-term ad performance metrics as the key retention metrics we need to optimize for advertisers. In consideration of the desirable properties of a retention framework, we propose an Auto-Coupon module to distribute impression-level virtual coupons to selected advertisers during the bidding process. We decompose the resulted retention optimization problem into two stages, and adopt feedback control as well as reinforcement learning methods to realize personalized coupon distribution. The offline experiments demonstrate the effectiveness of our methods in boosting the ad performance of churning advertisers, and the large-scale online A/B test further justifies the improvement in advertisers' willingness to stay in the platform.
M2-MFP: A Multi-Scale and Multi-Level Memory Failure Prediction Framework for Reliable Cloud Infrastructure
PDF ↗As cloud services become increasingly integral to modern IT infrastructure, ensuring hardware reliability is essential to sustain high-quality service. Memory failures pose a significant threat to overall system stability, making accurate failure prediction through the analysis of memory error logs (i.e., Correctable Errors) imperative. Existing memory failure prediction approaches have notable limitations: rule-based expert models suffer from limited generalizability and low recall rates, while automated feature extraction methods exhibit suboptimal performance. To address these limitations, we propose M2-MFP: a Multi-scale and Multi-Level Memory Failure Prediction framework designed to enhance the reliability and availability of cloud infrastructure. M2-MFP converts correctable errors (CEs) into multi-level binary matrix representations and introduces a Binary Spatial Feature Extractor (BSFE) to automatically extract high-order features at both DIMM-level and bit-level. Building upon the BSFE outputs, we develop a dual-path temporal modeling architecture: 1) a time-patch module that aggregates multi-level features within observation windows, and 2) a time-point module that employs interpretable rule-generation trees trained on bit-level patterns. Experiments on both benchmark datasets and real-world deployment show the superiority of M2-MFP as it outperforms existing state-of-the-art methods by significant margins. Code and data are available at this repository: https://github.com/hwcloud-RAS/M2-MFP.
In logistic systems, demand prediction is an essential task providing the basis for improving the quality of terminal services, such as pick-up and delivery efficiency. However, the geographical scope of operations across multiple cities brings challenges due to the sparsity of user behavior data, hindering accurate predictions. Despite cross-city prediction methods potentially solving this problem by relying on the label of overlapping users in different cities, annotating these overlapping users is expensive. Additionally, the dynamic and diverse nature of user behaviors complicates feature transfer between cities. In this work, we define the logistics demand prediction problem as forecasting pick-up and delivery demand for zones, the smallest operational units in logistics systems, in different cities. To address the challenge, we propose TSTL, a Transferable Spatio-Temporal Learning framework for cross-city logistics prediction with sparse user data. TSTL advances existing methods from two aspects: (1) User-level invariant representation module extracts consistent user representations for overlapping and non-overlapping users across cities. (2) User-zone graph aggregation module enhances user embeddings by integrating dynamic interactions, such as logistics behaviors, into inherent user relations. Finally, the multi-city transfer module fine-tunes model parameters for city-invariant knowledge adoption and predicts future logistics demand. We implement and evaluate TSTL on one of the largest logistics systems. Extensive offline experiments and real-world deployment demonstrate the effectiveness of TSTL.
Recommender systems (RSs) often suffer from the feedback loop phenomenon, i.e., RSs are trained on data biased by their recommendations. This leads to the filter bubble effect that reinforces homogeneous content and reduces user satisfaction. To this end, serendipity recommendations, which offer unexpected yet relevant items, are proposed. Recently, large language models (LLMs) have shown potential in serendipity prediction due to their extensive world knowledge and reasoning capabilities. However, they still face challenges in aligning serendipity judgments with human assessments, handling long user behavior sequences, and meeting the latency requirements of industrial RSs. To address these issues, we propose SERAL (Serendipity Recommendations with Aligned Large Language Models), a framework comprising three stages: (1) Cognition Profile Generation to compress user behavior into multi-level profiles; (2) SerenGPT Alignment to align serendipity judgments with human preferences using enriched training data; and (3) Nearline Adaptation to integrate SerenGPT into industrial RSs pipelines efficiently. Online experiments demonstrate that SERAL improves exposure ratio (PVR), clicks, and transactions of serendipitous items by 5.7%, 29.56%, and 27.6%, enhancing user experience without much impact on overall revenue. Now, it has been fully deployed in ''Guess What You Like'' on the Taobao App homepage.
In this paper, we introduce the Generalized Auto Segmentation (GAS), an end-to-end, cross-domain personalization platform developed at Meta. GAS is designed to optimize personalization policies for online experiments, operating independently of any specific server infrastructure. Our process begins with a seed experiment to collect user data, which is followed by feature logging and selection. User segmentation is then performed using a feature-threshold method. Specifically, we utilize simple combinations of single or dual features to identify concise, explainable, and sustainable user cohorts that enhance both business outcomes and user experience over the long term. Treatments and the reward function for these segments are refined through ParallAX, a novel methodology at Meta that uses offline estimates to optimize online metrics, ensuring impactful and precisely tailored interventions for diverse user groups. The deployment of GAS across various Meta productions has demonstrated its efficiency and effectiveness.
Most recommendation systems typically follow a product-based paradigm utilizing user-product interactions to identify the most engaging items for users. However, this product-based paradigm has notable drawbacks for Xianyu~. Xianyu is China's largest online C2C e-commerce platform where a large portion of the product are post by individual sellers. Most of the product on Xianyu posted from individual sellers often have limited stock available for distribution, and once the product is sold, it's no longer available for distribution. This result in most items distributed product on Xianyu having relatively few interactions, affecting the effectiveness of traditional recommendation depending on accumulating user-item interactions. To address these issues, we introduce IU4Rec, an Interest Unit-based two-stage Recommendation system framework. We first group products into clusters based on attributes such as category, image, and semantics. These IUs are then integrated into the Recommendation system, delivering both product and technological innovations. IU4Rec begins by grouping products into clusters based on attributes such as category, image, and semantics, forming Interest Units (IUs). Then we redesign the recommendation process into two stages. In the first stage, the focus is on recommend these Interest Units, capturing broad-level interests. In the second stage, it guides users to find the best option among similar products within the selected Interest Unit. User-IU interactions are incorporated into our ranking models, offering the advantage of more persistent IU behaviors compared to item-specific interactions. This interest unit based recommendation can be beneficial from mitigating the side effect of limited-stock problem, since most interaction can be gathered on interest units which can persist and accumulate over time. Experimental results on the production dataset and online A/B testing demonstrate the effectiveness and superiority of our proposed IU-centric recommendation approach. This study not only advances recommendation technologies but also emphasizes the potential for co-evolution between product innovations and the technologies involved in item supply and distribution.
Large-scale recommendation systems mainly rely on sparse ID features, struggling with data sparsity. It's important to use multimodal information to assist ID learning for better performance. However, there exists two challenges: (1) distribution discrepancy between multimodal and ID makes direct integration prone to user-item mismatch; (2) slower convergence of multimodal representations compared to ID, causing optimization imbalance under a unified objective, which limits the potential of multimodal representations. In this paper, we comprehensively investigate the two problems and proposes a framework named AB-Rec to align and balance ID and multimodal representations learning for recommendation. We design three alignment tasks to fine-tune a pre-trained multimodal large language model (MLLM), which is then utilized to generate a unified multimodal representation for each item. AB-Rec aligns the distributions of ID and multimodal representations by minimizing the in-batch Wasserstein distance, and maximizes the distance between the two types of representations for the same item to avoid representation collapse. To solve the optimization imbalance, we propose a gradient modulation method that adaptively controls the optimization process by monitoring the contribution differences between ID and multimodal representations. Finally, we conduct extensive offline experiments on four datasets and an A/B test on an online video platform, demonstrating the effectiveness and scalability of our proposed method.
Personalized Query Auto-Completion for Long and Short-Term Interests with Adaptive Detoxification Generation
PDF ↗Query auto-completion (QAC) plays a crucial role in modern search systems. However, in real-world applications, there are two pressing challenges that still need to be addressed. First, there is a need for hierarchical personalized representations for users. Previous approaches have typically used users' search behavior as a single, overall representation, which proves inadequate in more nuanced generative scenarios. Additionally, query prefixes are typically short and may contain typos or sensitive information, increasing the likelihood of generating toxic content compared to traditional text generation tasks. Such toxic content can degrade user experience and lead to public relations issues. Therefore, the second critical challenge is detoxifying QAC systems. Recent efforts to mitigate toxicity have involved generating queries unrelated to the given prefix, leading this approach still negatively impacts user experience. To address these two limitations, we propose a novel model (LaD) that captures personalized information from both long-term and short-term interests, incorporating adaptive detoxification. In LaD, personalized information is captured hierarchically at both coarse-grained and fine-grained levels. This approach preserves as much personalized information as possible while enabling online generation within time constraints. To move a further step, we propose an online training method based on Reject Preference Optimization (RPO). By incorporating a special token [Reject] during both the training and inference processes, the model achieves adaptive detoxification. Consequently, the generated text presented to users is both non-toxic and relevant to the given prefix. We conduct comprehensive experiments on industrial-scale datasets and perform online A/B tests, delivering the largest single-experiment metric improvement in nearly two years of our product. Our model has been deployed on Kuaishou search, driving the primary traffic for hundreds of millions of active users. The code is available at https://github.com/JXZe/LaD.
Feature interaction modeling is crucial for deep recommendation models. A common and effective approach is to construct explicit feature combinations to enhance model performance. However, in practice, only a small fraction of these combinations are truly informative. Thus it is essential to select useful feature combinations to reduce noise and manage memory consumption. While feature selection methods have been extensively studied, they are typically limited to selecting individual features. Extending these methods for high-order feature combination selection presents a significant challenge due to the exponential growth in time complexity when evaluating feature combinations one by one. In this paper, we propose TayFCS, a lightweight feature combination selection method that significantly improves model performance. Specifically, we propose the Taylor Expansion Scorer (TayScorer) module for field-wise Taylor expansion on the base model. Instead of evaluating all potential feature combinations' importance by repeatedly running experiments with feature adding and removal, this scorer only needs to approximate them based on their sub-components' gradients. They can be simply computed with one backward pass based on a trained recommendation model. To further reduce information redundancy between feature combinations and their sub-components, we introduce Logistic Regression Elimination (LRE) that estimates the information gain of feature combinations over their sub-components based on the above importance scores. Experimental results on three benchmark datasets validate both the effectiveness and efficiency of our approach. Furthermore, online A/B test results demonstrate its practical applicability and commercial value.
Many real-world applications contain data in the form of multivariate time series (TS) with the hierarchical structure, where classic methods forecasting each TS independently are inadequate for coherency (i.e., satisfying the hierarchical aggregation constraints). Furthermore, the discrepancies between statistical properties of different levels can be huge, exacerbated by non-Gaussian distributions and non-linear correlations. In this paper, we propose a novel end-to-end hierarchical TS forecasting model, i.e., a Flow-based Reconcile Transformer (FRT). FRT employs a conditional normalizing flow-based autoregressive transformer, to represent complex data distribution, while simultaneously reconciling the forecasts to ensure coherency. Go beyond other state-of-the-art methods, FRT accomplishes forecasting and reconciliation simultaneously, while avoiding any post-processing steps. Moreover, FRT is a deep model that does not rely on any strong assumptions such as unbiased estimates or Gaussian distribution. Our experiments are conducted on four real-world hierarchical datasets from different industrial domains (three public ones and a dataset from the application servers of our company's data center) and the results demonstrate the efficacy of our proposed method. Our method has been implemented extensively within the production environments of a prominent global payment company. It has emerged as a cornerstone for workload forecasting within their data center and plays a critical role in the optimization of cloud computing resource allocation across the entire cluster. This successful deployment of the application demonstrates that our approach achieves precise hierarchical workload prediction, which is of great significance for efficient resource scheduling, enhancing resource utilization, and reducing server resource use and energy consumption in data centers.
The profound psychological interplay between music and human affect underscores the critical role of emotion-aware modeling in music recommendation systems. While existing approaches predominantly employ emotion classification or vector representation methods, their capacity to capture nuanced emotional semantics remains limited. This paper presents the Large-Language-Model-based emotion-aware music recommendation model(LEMON), a novel framework that harnesses Large Language Models to address two fundamental challenges: extracting multi-faceted emotional semantics from music content and dynamically aligning users' transient emotional states with long-term preferences. Our framework introduces three key innovations: First, a hierarchical LLM-driven emotion extraction pipeline that decodes complex emotional narratives from lyrics through multi-stage prompt engineering, generating semantically rich song embeddings. Second, a dual-temporal encoder that separately models users' short-term emotion representation derived from recent engagement patterns and long-term preferences extracted from historical interactions. Third, an adaptive fusion coefficient that dynamically balances temporal features using user profile attributes to enable personalized recommendation adaptation. Extensive evaluations on industrial and public datasets demonstrate LEMON's superiority, achieving 4.22% AUC improvements over state-of-the-art baselines. Online A/B tests on NetEase Cloud Music, one of China's largest music streaming platforms validate a 1% increase in user engagement metrics. Ablation studies confirm the necessity of each component, particularly highlighting LLMs' performance gain in capturing context-dependent emotions compared to conventional methods. This work advances the integration of LLMs' semantic reasoning capabilities into emotion-aware recommendation systems, offering practical insights for online music personalization.
Put Teacher in Student's Shoes: Cross-Distillation for Ultra-compact Model Compression Framework
PDF ↗In the era of mobile computing, deploying efficient Natural Language Processing (NLP) models in resource-restricted edge settings presents significant challenges, particularly in environments requiring strict privacy compliance, real-time responsiveness, and diverse multi-tasking capabilities. These challenges create a fundamental need for ultra-compact models that maintain strong performance across various NLP tasks while adhering to stringent memory constraints. To this end, we introduce Edge ultra-lIte BERT framework (EI-BERT) with a novel cross-distillation method. EI-BERT efficiently compresses models through a comprehensive pipeline including hard token pruning, cross-distillation, parameter quantization, and plugin-and-play deployment. Specifically, the cross-distillation method uniquely positions the teacher model to understand the student model's perspective, ensuring efficient knowledge transfer through parameter integration and the mutual interplay between models. Through extensive experiments, we achieve a remarkably compact BERT-based model of only 1.91 MB - the smallest to date for Natural Language Understanding (NLU) tasks. This ultra-compact model has been successfully deployed across multiple scenarios within the Alipay ecosystem, demonstrating significant improvements in real-world applications. For example, it has been integrated into Alipay's live Edge Recommendation system since January 2024, currently serving the app's recommendation traffic across 8.4 million daily active devices.