值得一提的是，公司通过使用多模态大模型，提升了意图识别和风险判定的精度，能精准捕捉"还款困难"等语义特征。另外智能预警模型使早期风险识别效 率提升10%，为信贷业务安全、稳定开展提供了坚实保障。 今年，海尔消金"特征英雄"活动共吸引32名员工踊跃参与，从海量数据里挖掘出了2023个高质量特征！这可不是小数目——相当于给风控系统，添了2023 个"新武器"、"新视角"和"新战场"。 （2025年度特征英雄大赛颁奖仪式，图片来源：海尔消费金融） 本届"特征英雄"大赛一等奖获得者来自风险管理中心，其聚焦创新信息维度，一方面通过采用大模型来替代人工对语音数据做批量处理，提取有效信息用于 信贷模型研发，助力信贷风控策略有效施行，进一步降低信用风险；另一方面充分挖掘信贷场景中时序特征的潜在价值，突破传统"静态快照"局限，利用 RNN、Transformer等序列模型解析用户信息时序数据，精准识别"以贷养贷"行为，阻断风险传导。 近日，海尔消费金融有限公司（以下简称"海尔消金"）2025年"特征英雄"活动圆满收官。据海尔消金风控管理中心相关负责人介绍，"特征英雄"活动旨在激 发公司全员坚持数据科学驱动的价值理念，深度挖掘金 ...

（原标题：港理工成立人工智能高等研究院 聚焦去中心化AI与成果转化） 南方财经 21世纪经济报道记者 张伟泽 实习生 金颖 张艺宁 香港报道 12月10日，香港理工大学人工智能高等研究院（PAAI）正式成立。香港理工大学校长滕锦光在成立仪 式上表示，该研究院是理大未来重点布局的五大高等研究院之一，致力于开展跨学科基础研究并推动成 果转化。研究院将充分发挥理大在内地的科研网络优势，着力将前沿科研成果转化为现实的产品、技术 与标准，服务社会发展。 针对人工智能发展中存在的"数据孤岛"与"隐私安全"核心挑战，PAAI创院院长杨强指出，未来AI发展 的方向不应是依赖单一的巨型模型，而是构建连接的"垂直领域专家模型"。"联邦学习"（Federated Learning）应运而生，这是一种安全、去中心化的人工智能开发方式，能够在不交换原始数据的前提下 有效打破医疗与金融领域的数据流通壁垒及破解隐私保护难题。 杨强认为，作为国际金融中心与科创中心，香港背靠粤港澳大湾区丰富的临床网络与产业需求，未来将 着力构建去中心化AI基础设施，推动各类机构在安全可控的前提下运用先进AI技术。 对于AI领域未来的挑战，杨强表示，PAAI ...

Core Viewpoint - The article highlights the evolution and significance of the GAIR (Global Artificial Intelligence and Robotics Conference) as a platform for Chinese AI scholars, particularly those associated with Microsoft, to connect and collaborate, marking a shift in China's position in the global AI landscape [5][9]. Group 1: Historical Context - In 1996, Wu Feng, a doctoral student at Harbin Institute of Technology, reached out to Zhang Yaqin, a prominent scientist, to advocate for China's inclusion in the MPEG committee, aiming to enhance the international recognition of local scholars [2][4]. - Zhang Yaqin, alongside Li Kaifu, co-founded the Microsoft Research Asia, which became a pivotal institution for AI development in China, fostering connections between academia and industry [5][6]. Group 2: GAIR Development - The first GAIR conference was held in Shenzhen, initiated by prominent figures like Zhu Xiaorui and Lin Jun, bringing together top overseas scientists to discuss AI and robotics [7][8]. - Over the years, GAIR has become a gathering point for over 40 Microsoft-affiliated scientists, facilitating discussions on various AI topics and fostering collaboration between academia, industry, and investment sectors [9][10]. Group 3: Notable Contributions and Events - The GAIR conferences have featured significant contributions from Microsoft scientists, addressing critical issues in AI, such as deep learning challenges and interdisciplinary integration [9]. - The upcoming eighth GAIR conference is scheduled for December 12-13, 2025, in Shenzhen, continuing the tradition of fostering innovative ideas and collaborations in the AI field [10].

Core Insights - The article discusses the introduction of a new framework called FedVLR, which addresses the challenges of multimodal integration in federated learning environments while ensuring data privacy [2][3][19]. Multimodal Integration Challenges - Current recommendation systems utilize multimodal information, such as images and text, but face difficulties in federated learning due to privacy concerns [2][5]. - Existing federated recommendation methods either sacrifice multimodal processing for privacy or apply a one-size-fits-all approach, which does not account for individual user preferences [2][5]. FedVLR Framework - The FedVLR framework redefines the decision-making flow for multimodal integration by offloading heavy computation to the server while allowing users to control how they view the data through a lightweight routing mechanism [3][19]. - It employs a two-layer fusion mechanism that decouples feature extraction from preference integration [8][19]. Server-Side Processing - The first layer involves server-side "multi-view pre-fusion," where the server processes data using powerful pre-trained models to create a set of candidate fusion views without burdening client devices [9][10]. - This approach ensures that the server prepares various "semi-finished" views that contain high-quality content understanding [10]. Client-Side Personalization - The second layer focuses on client-side "personalized refinement," utilizing a lightweight local mixture of experts (MoE) routing mechanism to dynamically compute personalized weights based on user interaction history [11][12]. - This process occurs entirely on the client side, ensuring that user preference data remains on the device [12]. Performance and Versatility - FedVLR is designed to be a pluggable layer that can integrate seamlessly with existing federated recommendation frameworks like FedAvg and FedNCF, without increasing communication overhead [16]. - The framework demonstrates model-agnostic capabilities, allowing it to enhance various baseline models significantly [26]. Experimental Results - The framework has been rigorously tested on public datasets across e-commerce and multimedia domains, showing substantial and stable improvements in core recommendation metrics like NDCG and HR [26]. - Notably, FedVLR performs exceptionally well in sparse data scenarios, effectively leveraging limited local data to understand item content [26]. Conclusion - FedVLR not only enhances recommendation systems but also provides a valuable paradigm for implementing federated foundational models, addressing the challenge of utilizing large cloud models while maintaining data privacy [19].

Core Insights - The EMNLP 2025 conference will be held in Suzhou, China, from November 4-9, 2025, focusing on breakthroughs in data-driven NLP technologies and attracting thousands of researchers and engineers globally [1] - WeBank has had three papers accepted at EMNLP, focusing on federated large model frameworks, model compression, and knowledge transfer in inference [1] - WeBank is committed to becoming a leading AI-native bank, emphasizing technological self-reliance and innovation in AI, blockchain, big data, and cloud computing [3] Group 1: Research and Development - WeBank has published a total of nine papers in top conferences such as EMNLP, NeurIPS, CVPR, KDD, TIFS, and TKDE in 2025, marking a record high for the bank [3] - The bank's AI research focuses on key technologies such as federated model compression, large model distillation, and customized large model services [3][4] - WeBank has published over 50 papers in international conferences and journals, applying various technologies to core business areas like financial risk control, intelligent customer service, and fraud prevention [4] Group 2: Standards and Recognition - WeBank led the drafting of the national standard GB/T 46284-2025 for federated learning technology, which came into effect on October 5, 2025 [3] - The bank, in collaboration with Tongji University and Microsoft Research Asia, won the third prize in the CCF Technology Achievement Award for Natural Sciences in 2025, highlighting its technological innovation capabilities [3] Group 3: Future Directions - WeBank aims to deepen its AI core technology research, focusing on "financial technology" and "smart living" to convert research outcomes into industrial value [5] - The bank is dedicated to providing leading technological solutions for building a "safe, efficient, and fair" AI application ecosystem [5]

Core Insights - The EMNLP 2025 conference will be held in Suzhou, China, from November 4-9, 2025, focusing on breakthroughs in data-driven NLP technologies [1] - WeBank has three papers accepted at EMNLP, exploring federated large model frameworks, model compression, and knowledge transfer in reasoning [1] - WeBank is a pioneer in digital banking in China, committed to making financial services accessible to the public through technological innovation [2] Group 1 - WeBank is accelerating its transition from a "digital native" to an "AI native" bank, focusing on AI infrastructure, applications, and governance [2] - The bank has achieved a record high of nine papers accepted at top conferences in 2025, including EMNLP, NeurIPS, and CVPR [2] - WeBank led the development of the national standard GB/T 46284-2025 for federated learning technology, effective from October 5, 2025 [2] Group 2 - WeBank's AI research emphasizes a problem-oriented approach, aiming to address real industry pain points [3] - The bank has published over 50 papers in top international conferences, with technologies applied in financial risk control, intelligent customer service, and fraud prevention [3] - Future research will focus on transforming AI research outcomes into industrial value, contributing to a safe, efficient, and fair AI application ecosystem [3]

Core Insights - The G20 GPU financial AI ecosystem addresses challenges in traditional risk control systems, such as insufficient computing power and data fragmentation, by integrating hardware, software, and data collaboration for enhanced security in digital finance [1][2] - The ecosystem enables real-time risk identification, achieving risk assessment within 0.3 seconds for each transaction, significantly improving the efficiency and accuracy of fraud detection [1][2] Group 1: Risk Control System Enhancements - The G20 ecosystem allows for real-time sharing of risk characteristics among algorithm vendors, reducing the model update cycle from 1-2 weeks to 24 hours, resulting in a 22% increase in fraud interception rates and an 18% decrease in false positives [2] - The system incorporates 18 detection measures, including transaction behavior analysis and device security checks, to generate risk scores and determine transaction approval [1] Group 2: Data Sharing and Collaboration - The ecosystem has established a cross-institution risk data sharing mechanism with banks and insurance companies, utilizing federated learning to ensure data privacy while optimizing risk control models [2] - A participating bank reported a 15% improvement in credit card default risk prediction accuracy and a 0.8 percentage point reduction in non-performing loan rates after joining the data sharing initiative [2] Group 3: Ecosystem Impact - The G20 financial AI ecosystem has served over 20 financial institutions across various sectors, intercepting suspicious transactions worth over 1.5 billion and handling more than 300,000 risk events, thereby supporting the stable operation of digital finance [3]

Core Insights - The Chinese federated learning industry is experiencing steady growth driven by policy support, technological advancements, and market demand, with a projected market size of 254 million yuan in 2024, representing a year-on-year increase of 11.89% [1][8] - Federated learning effectively addresses the challenges of data silos and privacy security, enhancing model accuracy by over 20% in various applications such as financial risk control, medical joint diagnosis, and urban traffic optimization [1][8] Industry Overview - Federated learning (FL) is a distributed machine learning method aimed at enabling collaborative model training while protecting data privacy. It allows participants to train models locally using their own data and share encrypted model parameters with a central server, thus avoiding data sharing across institutions and complying with privacy regulations like GDPR [2] - The industry has evolved through four stages since the concept was introduced by Google in 2017: exploration, application, ecosystem building, and mature expansion [3] Market Size - The market size of the Chinese federated learning industry is expected to reach 254 million yuan in 2024, with a growth rate of 11.89% year-on-year [1][8] - The industry is supported by continuous iterations of technological frameworks, such as WeBank's FATE and Ant Group's shared intelligence platform, which incorporate privacy protection technologies like homomorphic encryption and secure multi-party computation [1][8] Key Companies - Leading companies in the federated learning sector include Ant Group and WeBank, with Ant Group holding a 36.7% market share in the privacy computing market for three consecutive years [8] - WeBank has pioneered the application of federated learning technology in the financial sector, with its open-source FATE framework becoming an industry benchmark [8] Industry Development Trends - The integration of federated learning with AI large models, edge computing, and 5G/6G technologies is expected to create a new paradigm of distributed AI collaboration [10] - Applications of federated learning are expanding beyond finance and healthcare into industrial internet, autonomous driving, and energy management, enhancing the technology's role in digital transformation [11][12] - The establishment of standards and the improvement of domestic policies are expected to strengthen the industry's foundation, with initiatives like the IEEE P3652.1 standard and the implementation of data security laws providing compliance support [13]

Core Viewpoint - The article discusses the FedPall algorithm, which addresses the feature drift problem in federated learning by combining prototype-based adversarial and collaborative learning techniques, achieving state-of-the-art performance across various datasets [2][10]. Methodology - The FedPall framework introduces an adversarial learning mechanism between clients and the server, enhancing feature representation alignment in a unified feature space through collaborative learning [3]. - A hierarchical integration strategy is developed to combine global prototypes with local features, facilitating client-server collaboration [5]. - The server trains a shared global amplifier and utilizes KL divergence to enhance heterogeneous information from different clients, mapping raw data to a unified feature space [5]. - The global classifier is distributed to each client, replacing the local classifiers to improve generalization and mitigate feature drift [6]. Performance Evaluation - FedPall was evaluated on three publicly available feature drift datasets: Digits, Office-10, and PACS, demonstrating superior accuracy compared to classical methods and state-of-the-art baselines [8][10]. - In the Office-10 dataset, FedPall achieved an overall accuracy approximately 3 percentage points higher than the second-best method, ADCOL [10]. - The Digits dataset results showed FedPall outperforming all other models, with an accuracy exceeding the second-best model, FedBN, by about 1.1 percentage points [10]. - FedPall consistently maintained higher average accuracy across all three datasets compared to ADCOL, with improvements ranging from 1.1 to 3 percentage points [12]. Future Directions - The research aims to validate the FedPall framework's generalization capabilities across other data modalities and task types in future studies [13].

Core Insights - Short video platforms have become a crucial battleground for brand promotion, with Douyin leveraging its large user base and precise algorithmic recommendations to attract businesses [1] - The "Giant Advertising" system launched by Douyin helps advertisers achieve efficient exposure and conversion through a bidding model, leading many companies to outsource their advertising accounts to professional operation companies [1] Group 1: Industry Challenges and Needs - Many small and medium-sized enterprises face multiple challenges when managing Douyin ads, including budget wastage due to lack of experience with billing models like OCPM and CPC [2] - Creative content production is often weak, requiring cross-disciplinary collaboration that non-professional teams struggle to achieve [2] - Weak data analysis capabilities hinder effective interpretation of vast data metrics, making it difficult to adjust advertising strategies [2] Group 2: Core Competencies of Operation Companies - Professional operation companies excel in building precise user profiles by deeply analyzing target customer segments and utilizing Douyin's tagging system [5] - They employ a data-driven dynamic adjustment mechanism, using intelligent bidding systems combined with manual intervention to optimize key performance indicators [5] - Established teams have standardized content production processes, enhancing the replicability of successful content, with professionally produced ads retaining viewer attention 60% longer than self-made content [6] Group 3: Industry Issues and Partner Selection - The surge in market demand has led to some negative phenomena, such as studios charging high fees while using deceptive methods to inflate performance metrics [7] - Companies should verify the authenticity of case studies and request recent advertising data reports to avoid misleading claims [9] - It is essential to confirm the certification status of operation partners and ensure transparency in contract terms to prevent disputes [9] Group 4: Future Trends - The operation industry is evolving towards automation and precision, with automated tools taking over basic tasks, allowing human resources to focus on strategic decision-making [10] - The ability to integrate across multiple platforms will become a new competitive barrier, with service providers managing resources across Douyin, Kuaishou, and WeChat ecosystems being more favored [10] - Breakthroughs in privacy computing may lead to new models for collaborative advertising targeting while protecting user privacy [10]