Kuaishou E-commerce's merchant team details their end-to-end AI delivery pipeline (Harness Engineering) combined with organizational restructuring, achieving 30% throughput gains and 40% delivery-time reduction by shifting from functional silos to full-stack 'super individuals' across requirements, spec, dev, test, and oncall phases.
AI PipelineFull-Stack DevelopmentHarness Engineering
The article examines how six Kuaishou engineering teams integrated AI across product, development, testing, e‑commerce, anti‑fraud and operations, using Agent‑First platforms and a Harness framework to convert individual coding speed gains into measurable organizational efficiency, merchant revenue growth, and faster risk detection.
Kuaishou’s analysis reveals that increasing developer count reduces AI efficiency gains, identifies three layers of human friction, three organizational gaps, and four AI‑native principles, and demonstrates a four‑quadrant framework that boosted AI code generation and demand delivery by up to fourfold.
This article presents Kuaishou’s 2026 H1 comprehensive review of its AI productivity system, detailing the three‑year evolution, the three critical bottlenecks that hindered scaling (skill polarization, workflow friction, and AI‑human coordination), the historical banking analogy that shaped a new AI‑productivity paradigm, a five‑stage roadmap with three parallel execution tracks, and concrete case studies that demonstrate a 10% rise in per‑person demand delivery, an 11%+ reduction in cycle time, and a doubling of AI code‑generation rates across the organization.
The article analyzes how AI‑driven test generation at Kuaishou evolved from an 8% to a 70% success rate, explains why testing is shifting from simple execution to system design, and offers concrete steps for test engineers to upgrade their capabilities and stay indispensable.
The Kuaishou technology team had 25 papers accepted at the prestigious KDD 2026 conference—including three oral presentations—covering generative recommendation, automated bidding, semantic ID learning, multi‑behavior modeling, and other AI‑driven advances that are already deployed at massive scale on the platform.
Amid exploding real‑time data demand, the traditional Lambda architecture suffers from high cost, data inconsistency and operational complexity, prompting a shift to a unified incremental computation engine that delivers minute‑level latency, sub‑hourly cost, and sub‑1% result divergence, as demonstrated by Kuaishou and Xiaohongshu production deployments.
Kuaishou’s over‑thousand‑engineer team discovered that merely adding AI tools boosted individual coding speed but left overall delivery cycles unchanged, prompting a three‑level AI‑native redesign (L1‑assist, L2‑collaborate, L3‑autonomous), new metrics, and a restructuring of information, workflow, and organization to truly capture AI’s productivity potential.
The talk reveals why, despite 89% of firms deploying AI, overall productivity only rose 0.29%, and explains Kuaishou's three‑layer AI‑native paradigm (L1‑L3), the hidden frictions between humans and AI, and the three‑tier restructuring of information, processes, and organization needed to turn AI capability into real engineering efficiency.
The Kuaishou technology team reports that several of its papers were accepted at the prestigious ICML 2026 conference—including a spotlight paper on metaphor video understanding, works on causal discovery for irregular time series, image super‑resolution, large‑scale notification dispatch, full‑order ranking, phase‑aware MoE for RL, end‑to‑end e‑commerce search, spatial‑reasoning rewards, a unified SWE benchmark, video temporal grounding, and interpretable transformers—while also inviting attendees to visit their booth B101 in Seoul.
The article recaps Kuaishou's June 13 technical salon, detailing the systematic evolution of generative recommendation—from scaling models to reasoning enhancements—through core projects OneReason, Pool‑Rec, OneSearch V2, and GR4AD, and announces the industry‑wide LLM‑Rec challenge for students.
The article outlines Kuaishou's seven technical sessions at HDC 2026, detailing solutions for HarmonyOS large‑scale deployment such as startup performance, HD streaming, memory‑leak mitigation, cross‑platform framework adaptation, KMP integration, ArkUI optimization, and AI‑native enhancements.
This article reviews eight Kuaishou‑authored papers accepted at ICLR 2026, summarizing their problem statements, novel methods such as front‑door causal attribution, visual table retrieval, denoising rerankers, difficulty‑adaptive reasoning, diffusion code infilling, generative ordinal regression, multimodal video retrieval, e‑commerce dialogue benchmarks, and a new LLM creativity evaluator, together with reported experimental gains.
Within a 60‑90 minute window on Dec 22, hackers breached Kuaishou’s systems, exploiting nighttime staffing fatigue, high‑privilege token leaks, AI moderation tricks, direct video uploads, massive account overload, and microservice architecture flaws, highlighting critical security gaps for large platforms.
This article explores Kuashou's e‑commerce data metric system, detailing why a metric framework is essential, the business context, the challenges faced across data production, querying and usage, and the practical solutions implemented to improve decision‑making and operational efficiency.
This article explains why Kuaishou e‑commerce needed a metric system, outlines its business context, describes the challenges faced by data product managers, engineers, and operators, and shares the practical steps and management practices used to construct and maintain the system.
The article examines Kuaishou’s e‑commerce data metric system, detailing why a metric framework is essential, how it was built, the product practice, management methods, and the challenges faced by data product managers, engineers, and operators across production, querying, and usage stages.
This article explores Kuaishou e‑commerce's journey in constructing a comprehensive data metric system, detailing its business context, the necessity of metrics, challenges faced by data product managers and engineers, practical implementation steps, management practices, and a concluding Q&A.
The article reviews eight AI developments—including Tencent's task‑mode‑enabled ima 2.0, DeepSeek's visual‑compression OCR model, Quark's Qwen3‑Max assistant, Kuaishou's three‑layer AI coding ecosystem, Opera Neon’s ODRA agent, USTC‑ByteDance MoGA video generator, real‑time voice deepfakes, and OpenAI’s upcoming music tool—while providing analytical commentary on their broader impact.
This article examines Kuaishou's e‑commerce data metric system, explaining why such a framework is essential, how it was constructed, the practical product implementation, management practices, and lessons learned, while also addressing common data production, query, and usage challenges faced by product and engineering teams.
This article explores Kuaishou E‑Commerce’s journey in constructing a comprehensive data metric system, detailing its business context, the necessity of metrics, challenges across data production, querying and usage, practical implementation steps, management practices, and a concluding Q&A.
This article reviews Kuaishou's two‑year exploration of multimodal large‑model techniques for advertising, outlining challenges in content‑domain ad estimation, the COPE unified product representation framework, and the LEARN LLM knowledge‑transfer approach that together improve ad system performance.
This article reviews Kuaishou's two‑year exploration of large‑model techniques in advertising, outlining challenges in content‑domain ad estimation, introducing the COPE unified content representation framework and the LEARN LLM knowledge‑transfer approach, and showing how these innovations delivered tangible business gains.
This article reviews Kuaishou's two‑year exploration of multimodal large‑model techniques for advertising, detailing challenges of fragmented user behavior, the COPE unified product representation framework, and the LEARN LLM knowledge‑transfer approach that together delivered measurable business gains.
This article reviews Kuaishou’s two‑year exploration of large‑model techniques in advertising, detailing the content‑domain estimation challenges, how multimodal and LLM approaches improve full‑domain behavior utilization and external knowledge integration, and introducing the COPE product‑content representation framework and the LEARN LLM knowledge‑transfer system.
Over the past two years, Kuaishou has leveraged multimodal large‑model techniques to overcome sparse advertising data, integrating full‑domain user behavior and external knowledge via the COPE unified product representation framework and the LEARN LLM knowledge‑transfer system, achieving measurable business gains.
The article reviews the evolution of e‑commerce search, explains why traditional multi‑stage pipelines struggle with relevance and personalization, and details OneSearch’s end‑to‑end large‑model design—hierarchical quantization encoding, integrated user profiling, and preference‑aware ranking—that achieved up to 27% model‑utilization and double‑digit CTR, CVR, and order growth in production.
This article outlines Kuaishou Data Platform's mission to boost data decision efficiency, describes its three‑layer architecture, explains the BI process from data ingestion to application, and shares practical experiences and future outlook for intelligent BI powered by AI and big data.
This article outlines Kuaishou's Data Platform team's mission to boost data‑driven decision making through advanced compute engines, high‑performance services, and AI‑enhanced BI, detailing its architecture, challenges, solutions, and future outlook for large‑scale intelligent analytics.
This article outlines how Kuaishou’s Data Platform Department enhances decision‑making efficiency by building advanced compute engines and high‑performance services, detailing the platform’s architecture, challenges of intelligent BI, AI‑driven solutions, and the end‑to‑end BI workflow from data ingestion to analysis.
The KuaiMM Conversation project introduces a multimodal large‑model‑driven dialogue system for Kuaishou, featuring the world‑first short‑video mixed‑dialogue dataset, a Chain‑of‑Thought interaction framework, and large‑scale industrial deployments that dramatically improve live‑stream comments and intelligent customer service.
The article details Kuaishou architect Zhao Qingjie’s analysis of B‑end business challenges versus C‑end, outlines the multi‑dimensional governance covering service, business, storage and disaster‑recovery layers, and explains how systematic quality‑assurance, monitoring, error‑code design and organizational mechanisms dramatically improved system stability.
OneRec, an end-to-end generative recommendation system from Kuaishou, uses an encoder-decoder architecture, reward-based preference alignment, and reinforcement learning to dramatically improve video recommendation efficiency, boosting user engagement and reducing operational costs while achieving scaling-law performance comparable to large language models.
The OneRec system from Kuaishou replaces traditional cascade recommendation pipelines with an encoder‑decoder architecture, leverages reward‑based preference alignment via reinforcement learning, achieves ten‑fold FLOPs gains, cuts operational costs by 90%, and delivers significant user‑engagement improvements across short‑video and local‑service scenarios.
This article details how Kuaishou dramatically cut storage costs and improved cache efficiency for its core Photo data object by cleaning up redundant JSON fields, applying selective serialization, and performing large‑scale data cleaning, achieving a 25% size reduction, a 2% cache‑hit increase, and multi‑hundred‑TB savings.
Facing slow startup and low first‑contentful‑paint rates, Kuaishou’s development team overhauled the mini‑program simulator by consolidating its dual‑process architecture, introducing container caching, implementing on‑demand compilation, and refining debugger interactions, ultimately boosting instant‑open rates from 18% to 64% and reducing FCP P90 from 4.4 s to 1.9 s.
Kuaishou has achieved a significant milestone with 7 papers accepted at AAAI 2025, covering diverse AI research areas including video processing, recommendation systems, and image restoration, demonstrating the company's strong research capabilities in artificial intelligence.
ByteDance and Kuaishou have launched an “anti‑jargon” campaign, posting restroom cartoons that urge employees to abandon buzzwords like “赋能” and “抓手,” arguing that such jargon hampers clear communication, reduces efficiency, and fuels overtime, while also serving as a status symbol within the tech industry.
The NeurIPS 2024 Auto‑Bidding competition attracted over 15,000 submissions and 1,500 teams, featuring two tracks—General and AI‑Generated Bidding—where Kuaishou’s commercial algorithm team secured first place in both by leveraging reinforcement‑learning‑based online exploration and a decision‑transformer‑driven generative approach, achieving more than a 5% lift in ad revenue.
This article presents a comprehensive analysis of page performance issues across Kuaishou's commercial front‑end projects, outlines the challenges of unified governance, B‑end experience measurement, and C‑end web‑native integration, and details the systematic optimization strategies and measurable results that significantly improved user experience and business metrics.
This article presents Kuaishou's data platform and BI system, describing its evolution from traditional reporting to AI‑driven intelligent analytics, the challenges of diverse user needs and data quality, and the controllable, trustworthy, and feasible solutions that enable large‑scale smart BI deployment.
Kuaishou and Tsinghua University were honored with the first‑place Qian Weichang Chinese Information Processing Science and Technology Award for their collaborative content recommendation project, which achieved international‑level innovations in explainable recommendation, bias correction, and edge intelligence, and has been applied widely in Kuaishou's platform and top academic conferences.
This article examines how Kuaishou’s e‑commerce platform leverages AI to analyze and improve B‑side user experience, detailing current challenges, the CPS metric model, a self‑built intelligent inspection platform, workflow automation, and future prospects for AI‑driven UX optimization.
This article presents Kuaishou's big‑data analytics system, describing its current capabilities, the pain points of traditional ETL workflows, the NoETL concept, the implementation of a metric‑center platform, and practical features such as custom fields, automated modeling and acceleration, followed by future plans and a Q&A session.
This article presents Kuaishou's metric middle platform, detailing its background, design principles, architecture, metric management, data modeling, unified analysis language OAX, federated query engine OCTO, acceleration strategies, and future directions, illustrating how it improves data quality, development efficiency, and analytical capabilities at scale.
The article details Kuaishou's adoption of a data lake powered by Apache Hudi, covering the challenges of growing data warehouses, the migration from Hive to Hudi, concrete business case studies, promotion strategies, and key takeaways for large‑scale data engineering.
At the China Multimedia 2024 conference in Yinchuan, Kuaishou presented its latest AI‑driven large‑model technologies—including text‑to‑image, text‑to‑video, and audio models—alongside advances in intelligent video coding, a new research‑fund initiative, and recent industry awards.
This article presents Kuaishou's metric middle platform, detailing its background, design principles, metric management and service architecture, including headless BI concepts, unified analysis language OAX, query engine OCTO, data modeling layers, acceleration strategies, and future directions toward intelligence and high performance.
This article details Kuashou's journey of adopting the Hudi data lake, covering business challenges, migration from Hive to Hudi, architectural redesign, promotion strategy, real‑world use cases such as CDC sync and batch‑stream integration, and key lessons learned for large‑scale data engineering.
Kuaishou invites attendees to its CVPR 2024 presence in Seattle, featuring an elite dinner, booth talks, tech talks, and a showcase of eight accepted papers covering video quality assessment, image generation, and multimodal AI, with detailed schedules, authors, and presentation locations, plus recruitment opportunities.
This article shares Kuaishou's practical experience with data lake technology (Hudi), detailing the challenges of growing data warehouses, the migration from Hive to Hudi, the promotion strategy, real-world use cases such as CDC sync and batch‑stream integration, and key takeaways for future deployments.
The article reviews the evolution of large‑model technology, details Kuaishou’s self‑developed “Kuaishou Yi” model architecture, data pipeline, evaluation benchmarks, and explains how the AI XiaoKuai assistant achieves multimodal, personable interactions while also announcing related recruitment opportunities.
This article investigates why the storage size shown inside iOS apps often differs from the total size displayed in the iPhone Settings, analyzes four main causes—including app bundle size, binary vs decimal units, measurement scope, and cache path handling—and demonstrates how Kuaishou aligned its storage reporting with the system through experiments and code changes.
This article shares Kuaishou’s real‑time, massive‑scale model data processing pipeline, covering model scenarios, recommendation workflow complexity, large‑scale data storage, streaming joins, feature computation, NVM‑based storage solutions, strong consistency mechanisms, and future outlook for AI recommendation systems.
This article details Kuaishou's large‑scale data security platform, covering its background, architectural layers, authentication and permission models, full‑link audit, data classification and protection mechanisms, operational results, future road‑maps, and a Q&A session on practical challenges.
This talk presents Kuaishou's research on combining reinforcement learning with multi‑task recommendation, detailing a two‑stage constrained actor‑critic method for short‑video ranking, a multi‑task RL framework, experimental results on offline and online systems, and practical Q&A insights.
This article presents Kuaishou's end‑to‑end analytics platform, detailing the evolution from the early tool‑based stage through Service 1.0 and 2.0 to the unified Service 3.0 architecture, its unified analysis and query engines, data acceleration techniques, performance gains, and future intelligent analytics roadmap.
On March 1, Kuaishou hosted the launch of the Future High-Tech Video Intelligence Innovation Center in Beijing, highlighting collaborations with leading universities, breakthroughs in video codecs, a new transmission protocol, 6DoF video, digital humans, and large AI models that aim to drive industry-wide digital and intelligent transformation.
This article explains how Kuaishou tackled the growing complexity of its recommendation system by developing the Dragonfly strategy engine framework, detailing the challenges, architectural abstractions, DSL-based workflow composition, data handling, ecosystem tools, and future development plans.
This article presents Kuaishou's end‑to‑end engineering solutions for handling massive, real‑time recommendation model data, covering scenario description, complex business pipelines, trillion‑parameter model storage, high‑throughput processing with Flink and NVM, and future directions for cloud‑native scalability.
This article presents how causal inference methods are applied to Kuaishou’s single‑column short‑video recommendation, covering the platform’s recommendation scenario, model representations, duration bias mitigation, viewing‑time prediction techniques such as D2Q and TPM, experimental results, and future research directions.
Kuaishou’s Dragonfly framework replaces tightly‑coupled C++ recommendation code with a Python‑centric DSL that compiles to high‑performance JSON, offering DAG‑based operators, schema‑free dataframes, and a full ecosystem of debugging, tracing, and governance tools, dramatically cutting code volume while enabling fast, scalable scenario development.
This article describes how Kuaishou built the Dragonfly strategy engine framework—covering problem analysis, architecture design, DSL-based workflow orchestration, process and data abstractions, ecosystem tools, and future plans—to solve the scalability, coupling, and maintenance issues of its rapidly expanding recommendation services.
This article describes Kuaishou's approach to cold-start content recommendation, outlining the problems addressed, challenges in modeling sparse new videos, and solutions including graph neural networks, I2U retrieval, TDM hierarchical retrieval, bias correction, and future research directions.
The Kuaishou Academic Forum held on November 23 in Beijing presented cutting‑edge research on short‑video recommendation algorithms, featuring talks on reinforcement learning, user interest modeling, graph neural networks, and a comprehensive recommender‑system simulator, while also offering registration details and a brief company overview.
To handle billions of daily users and massive flash‑sale spikes in its live‑ecommerce streams, Kuaishou built a large‑scale elastic container cloud, integrated with Alibaba Cloud in a hybrid‑cloud setup, employing load balancing, caching, message queues, rate‑limiting, and intelligent resource scheduling to achieve million‑request‑per‑second throughput and high availability.
Kuaishou tackles massive live‑e‑commerce flash‑sale traffic by combining an elastic Docker‑based container cloud with a hybrid private‑public cloud architecture, enabling rapid scaling, high concurrency handling, and cost‑effective resource management.
The article describes how Kuaishou handles massive live‑streaming flash‑sale traffic by employing a distributed system, elastic Docker‑based container cloud, and a hybrid IDC‑Alibaba Cloud architecture that automatically scales resources to sustain millions of requests per second with high availability.
The article highlights five CIKM 2023 papers covering a lightweight model‑compression framework for recommender systems, a query‑dominant user‑interest network for large‑scale search ranking, a causal watch‑time labeling approach for short‑video recommendation, implicit negative‑feedback optimization for short‑video feeds, and the KuaiSAR unified search‑and‑recommendation dataset, each with download links, author lists, and key findings.
This article presents Kuaishou's approach to solving the content cold-start problem by analyzing its impact on video growth, detailing the challenges of sparse and biased training data, and describing a suite of graph‑neural‑network, I2U/U2I, TDM, and debiasing techniques that improve early video exposure and long‑term ecosystem health.
This paper presents TWIN, a two-stage interest network that aligns the similarity metrics of coarse‑grained and fine‑grained modules to improve lifelong user behavior modeling for CTR prediction in large‑scale online recommendation systems.
This article presents Kuaishou’s journey in building a data lake using Apache Hudi, detailing the lake architecture, key challenges such as ingestion bottlenecks and update inefficiencies, the solutions implemented, practical case studies, and the roadmap for future enhancements.
This article explains why Kuaishou built a data lake, describes its Hudi‑based architecture, outlines five major challenges encountered during implementation, and presents the solutions and future development plans, illustrating performance improvements and practical use cases across various business scenarios.
This paper presents RLUR, a reinforcement‑learning algorithm that models user‑retention optimization as an infinite‑horizon request‑based Markov Decision Process, addressing uncertainty, bias, and delayed reward challenges to directly improve retention, DAU, and engagement in short‑video recommendation platforms.
This article summarizes the Kuaishou Data Platform’s metric middle‑platform sharing from the 2022 DataFun forum, detailing its three‑year evolution, key concepts, architectural design, implementation challenges, and practical lessons for building an enterprise‑grade metric platform that unifies data definition, production, and consumption across the company.
This article presents Kuaishou's metric middle platform, detailing its three‑year evolution, unified semantic layer design, metric standardization workflow, core technologies such as Headless BI, automated modeling, and a unified metric service, and outlines future directions and a Q&A session.
On January 25, Kuaishou’s community science team announced that seven of its papers were accepted at the ACM Web Conference 2023 (WWW’23), covering reinforcement‑learning‑based user retention, constrained actor‑critic recommendation, divide‑and‑conquer embedding retrieval, causal embedding with contrastive learning, latent action space exploration, dual‑interest factorization attention, and multi‑task reinforcement learning for recommendation.
This article details Kuaishou’s three‑year evolution of its metric middle platform, covering the data infrastructure, key challenges of data inconsistency and low analysis efficiency, the enterprise‑level OneMetric solution, architectural design, development phases, practical lessons, system implementation, and real‑world applications.
Kuaishou's container cloud platform tackled the challenge of large‑scale, stable image distribution by integrating Dragonfly and its Nydus subproject, achieving over 70% bandwidth relief, 90% pull‑time reduction, and seamless migration to containerd, thereby supporting billions of monthly users with high efficiency.
This article presents a comprehensive technical overview of Kuaishou's Virtual World Interaction Platform (KMIP), detailing its background, platform and software solutions—including KVS and KFC—its layered architecture, key features, real‑world case studies, and future outlook for immersive virtual interactions.
This article introduces Kuaishou's exploration of uplift modeling for estimating heterogeneous treatment effects, discusses practical challenges such as continuous treatment variables and statistical inference for nonlinear models, presents a dual‑neural‑network solution with evaluation metrics, and showcases applications in fan growth and push notifications.
The article explores the rapid evolution of virtual digital humans, highlighting recent innovations, industry applications like Kuaishou's V‑Star program, academic research breakthroughs, and the technical and commercial challenges that must be overcome for these AI‑driven avatars to become mainstream.
This article explains the fundamentals of graph networks, their use in Kuaishou’s short‑video ecosystem for community detection, trend‑driven content creation, and causal inference, and describes how key network nodes and frequency‑illusion effects are leveraged to boost user engagement and content virality.
This article details Kuaishou's practical advances in short‑video recommendation, covering the CTR PPNet model, multi‑domain multi‑task learning, short‑ and long‑term user behavior sequence modeling, and the resulting performance gains in user engagement and app usage.
This article explains how Kuaishou detects and predicts viral short‑video素材 by defining content types, outlining essential viral elements, describing a two‑stage coarse‑recall and fine‑ranking model that combines speed‑based features, Gaussian mixture modeling, and a lightweight DNN, and showcases real‑world case studies and Q&A.
This article details Kuaishou's real‑time data warehouse architecture, the business challenges of massive traffic and diverse requirements, and the forward‑ and reverse‑assurance strategies—including lifecycle standards, monitoring, fault‑injection testing, and a Spring Festival case study—that together ensure high stability, low latency, and sub‑0.5% accuracy for billion‑scale streaming workloads.
To ensure consistent user experience across the wide range of Android and iOS devices, Kuaishou’s Y‑tech team designed a dynamic model‑grading framework that evaluates CPU, GPU, NPU, memory and other hardware metrics, then dispatches appropriately sized AI effect models and configurations in real time.
This article presents a comprehensive overview of how contrastive learning can alleviate data sparsity and distribution bias in recommendation systems, detailing its theoretical advantages, recent research progress in computer vision and NLP, and a multi‑task self‑supervised framework applied to Kuaishou's short‑video ranking pipeline with significant offline and online performance gains.
BodyDance is Kuaishou Y‑Tech’s self‑developed human motion transfer system that converts a single static portrait into a dancing video using 3D pose reconstruction, semantic parsing, and style‑based generation, achieving high‑quality results without online finetuning and outperforming existing methods in realism, hand fidelity, and speed.
3D pose reconstructionKuaishouhuman motion transfer
This article presents Kuaishou's large‑scale big‑data task scheduling system, describing its evolution from Airflow to the self‑developed Kwaiflow, the performance and reliability challenges of handling hundreds of thousands of tasks, and the design decisions that achieve low latency, high availability, and strong open capabilities.
Kuaishou's audio‑video technology team presents a self‑developed custom video matting system that combines foreground, interactive, and video object segmentation to let creators extract arbitrary subjects without green screens, featuring adaptive cropping, multi‑stage training, and deployment across Android and iOS devices.
This article presents Kuaishou's graph‑based item‑to‑item (I2I) recall pipeline for short‑video recommendation, detailing the business challenges, pipeline architecture, optimization techniques such as similarity‑measure tricks, graph structure learning, edge‑weight learning, and future research directions.
This article explains how Kuaishou leverages graph neural networks for item‑to‑item (I2I) recall in short‑video recommendation, detailing the system background, pipeline architecture, optimization techniques such as similarity measurement, graph structure learning, edge‑weight learning, and future research directions.
This article presents Kuaishou's graph‑based item‑to‑item (I2I) recall pipeline for short‑video recommendation, detailing the business challenges, the advantages of graph neural networks, the system architecture, optimization techniques such as similarity metric refinement, graph structure learning, edge‑weight learning, and future research directions.
This article details Kuaishou's short‑video recommendation pipeline, explaining the challenges of large‑scale sequencing, the development of sequence re‑ranking, multi‑content mixing, on‑device re‑ranking, and reinforcement‑learning‑based strategies, and demonstrates how these innovations improve user engagement and business metrics.
The Kuaishou audio‑video team designed a high‑precision, low‑latency intelligent danmu‑blocking system that uses advanced image‑segmentation and temporal‑stability techniques to generate accurate masks, improve scene robustness, eliminate mask delay, and enhance user experience across diverse video content.
This talk explores how causal inference techniques are used to identify and mitigate various biases in Kuaishou's recommendation pipeline, covering background theory, recent research advances, practical implementations for popularity and video completion debiasing, and reflections on challenges and future directions.
This article explains the adaptive 3D face cutout technology behind Kuaishou's "3D Zoom Face" effect, detailing its problem‑solving approach, implementation workflow, camera‑control optimizations, and how it expands creative possibilities while lowering production costs for both creators and users.