This article presents a comprehensive overview of how causal inference techniques are applied to identify and correct various biases in Kuaishou's recommendation pipeline, covering background theory, recent research, practical implementations such as popularity debias, causal embedding decoupling, and video completion‑rate debias, along with experimental results and future challenges.
This talk explains how recommendation bias arises from popularity and position effects, introduces causal inference concepts and three inference levels, reviews recent research such as DICE and Huawei’s causal embedding, and details Kuaishou’s practical applications—including popularity debias, causal representation decoupling, and video completion‑rate debias—along with experimental results and future challenges.
This article outlines Kuaishou’s approach to establishing a standardized data governance evaluation framework, detailing the challenges of large‑scale data management, the design of assessment metrics across model, quality, and cost dimensions, and the practical strategies and operational mechanisms used to improve data asset health and business value.
This article details the technical evolution of Kuaishou's short‑video recommendation pipeline, focusing on sequence re‑ranking, multi‑content mixing, and on‑device re‑ranking, and explains how transformer‑based models, generator‑evaluator frameworks, and reinforcement‑learning strategies are employed to maximize overall sequence value, user engagement, and revenue.
This article presents Kuaishou's mobile mixed reality (MR) system, detailing its integration of deep learning, SLAM, and scene reconstruction for real‑time spatial computing, the design of a monocular depth‑estimation model, a lightweight 3D rendering engine, and its deployment across iOS and Android devices with various user‑facing effects.
The Kuaishou audio team achieved a world‑runner‑up in the ICASSP 2022 Acoustic Echo Cancellation Challenge and had two deep noise suppression papers accepted, showcasing a hybrid traditional‑signal‑processing and deep‑learning system that excels in echo removal, noise reduction, and speech recognition across multiple real‑time communication scenarios.
This article explains the principles, challenges, and implementation details of Kuaishou’s proprietary green‑screen matting algorithm, covering fine‑detail handling, color‑spill reduction, green‑reflection removal, and its real‑time deployment across mobile video‑editing and live‑streaming products.
This article explains Kuaishou's comprehensive approach to data governance, detailing the background challenges, a standardized evaluation framework, practical solutions for model, quality, and cost issues, scoring strategies, operational mechanisms, measurable benefits, and future plans for continuous improvement.
Kuaishou’s audio‑video team built a dedicated high‑quality transcoding tier for Jackie Chan’s debut on the platform, deploying AI‑driven analysis, enhancement and compression techniques to deliver ultra‑clear short videos while keeping cost and latency low, and describing the end‑to‑end processing pipeline and trigger mechanisms.
Kuashou’s audio‑video team built a custom short‑video transcoding tier that combines AI‑driven enhancement, aggressive compression and a specialized delivery strategy to meet billions of viewers’ high‑quality expectations for Jackie Chan’s debut, showcasing rapid cross‑team collaboration and advanced video processing techniques.
Kuaishou reduced half of its massive object storage expenses by redesigning its architecture to use HBase indexing, HDFS large‑file storage, MemoryCache, and a cross‑IDC LRC erasure‑coding warm layer that maintains disaster‑recovery while dynamically moving data from hot to warm to cold tiers.
At ACM MM2021 in Chengdu, Kuaishou presented two accepted papers on recommendation and outfit compatibility, won the Grand Challenge with its DAP congestion‑control system, and forged a strategic partnership with the CCF‑MM committee to deepen multimedia research collaboration across academia and industry.
The interview with Kuaishou senior architect Zhao Jianbo details the three‑phase evolution of its trillion‑scale big data platform, covering foundational Hadoop services, real‑time and OLAP extensions, deep customizations, Spring Festival Gala challenges, scheduling innovations, Hadoop usage, and the relationship between big data and cloud architectures.
This article explores Kuaishou’s server‑side video pipeline, detailing why videos are re‑encoded, how AI‑based enhancement (De‑Art, SR, Deblur) and perceptual encoding techniques (CAQ, CARDO) improve subjective quality, reduce bitrate, and lower processing costs, while presenting algorithmic logic and real‑world results.
This article explains why a robust data metric system is essential for business decision‑making, defines the characteristics of a good metric system, introduces the OSM framework for constructing such systems, and illustrates the approach with Kuaishou's Spring Festival campaign, highlighting practical outcomes and lessons learned.
This article explains how Kuaishou applies causal inference frameworks, such as Rubin's potential outcomes and Pearl's causal graphs, together with machine‑learning techniques like double‑machine learning, causal forests, and meta‑learners to evaluate product features, recommendation strategies, and user behavior under complex network effects in live streaming.
This article analyzes Kuaishou's live‑streaming ecosystem, detailing causal‑inference frameworks, observational and experimental techniques such as DID, double machine learning, causal forests, uplift meta‑learners, and complex experiment designs like dual‑sided and time‑slice rotation to evaluate product and recommendation strategies.
This article presents Dr. Jin Yaran’s comprehensive overview of causal inference challenges, frameworks, and practical case studies—including DID, double machine learning, causal forests, and meta‑learners—applied to Kuaishou’s live‑streaming product, and discusses complex experimental designs such as bilateral and time‑slice experiments.
This article details Kuaishou's adoption of Apache Hudi to address data scheduling, synchronization, and massive update inefficiencies, describing the pain points, evaluation of alternatives, architectural integration with Spark/Flink, implementation challenges, and the performance improvements achieved.
The article details Kuaishou's Y-tech AI SDK (YKit) architecture, covering its design for computer vision capabilities, performance optimization strategies for mobile devices, and real-world case studies such as GAN-based effects and intelligent matting, outlining challenges and future directions.
The article introduces Kuaishou's Havok data‑service platform—a one‑stop, configuration‑driven solution that lowers development barriers—and details the comprehensive high‑availability architecture, including hierarchical isolation, elastic scaling, link grading, disaster recovery, and rate‑limiting mechanisms that enable zero‑failure support for large‑scale events.
This article introduces Kuaishou's data platform serviceification, outlining the background challenges for data engineers, the platform's architecture and key technologies such as configuration‑driven development, multi‑mode APIs, data acceleration, and high‑availability mechanisms, and concludes with a summary of achievements and future directions.
This article details Kuaishou's Flutter audio-video implementation, optimization practices, and engineering support, covering video playback via external textures, data communication strategies, FFI integration, and performance tuning.
Kuaishou becomes the first company to deploy a deep‑learning‑based real‑time voice‑conversion system on PC clients, delivering stable, natural‑sounding transformed speech with sub‑200 ms latency, and the article analyzes industry methods, technical challenges, and the four‑module architecture that made it possible.
This article details Kuaishou's award‑winning acoustic echo cancellation (AEC) solutions, explains the physics of echo formation, reviews traditional linear and nonlinear AEC methods, and introduces their Deep AEC approach that combines cross‑network architecture and advanced loss functions to achieve superior performance in both single‑talk and double‑talk scenarios.
This article details Kuaishou Y‑Tech's practical deployment of GAN‑based high‑precision face attribute editing—covering gender, age, hair, and expression transformations—for short‑video effects, discussing background, business applications, technical challenges, and solutions across data preparation, model training, and mobile deployment.
This article presents an in‑depth technical overview of Kuaishou's massive HDFS deployment, detailing its architecture, petabyte‑scale data and thousands‑of‑node clusters, the key scalability challenges faced, and the custom solutions—including FixedOrder, RBF balancer, observer read, slow‑node mitigation, and tiered protection—implemented to keep the system performant and reliable.
Hammer is an integrated hardware-aware model compression tool developed by Kuaishou in collaboration with universities, combining pruning, quantization, search, and distillation to achieve efficient and accurate neural network models tailored to specific hardware.
At Kuaishou, the Hive MetaStore service, which stores metadata for Hive, faced scalability and performance challenges due to massive dynamic partitions and high query volume, leading to a series of architectural optimizations—including read‑write separation, API enhancements, traffic control, and federation—to improve stability and efficiency.
This article details Kuaishou's multi‑objective ranking pipeline for short‑video recommendation, covering manual score fusion, GBDT ensemble, Learn‑to‑Rank, online auto‑tuning, ensemble sorting, reinforcement‑learning rerank, and on‑device rerank, and reports their impact on DAU, watch time and user interaction.
This article chronicles Kuaishou’s ten‑year journey from a humble GIF‑making app created by Cheng Yixiao, through its partnership with Su Hua, strategic pivots, IPO success and subsequent controversies, illustrating how focused product design and relentless R&D management built China’s premier short‑video platform.
This article chronicles how Kuaishou’s founders, from humble beginnings and early coding experiences, navigated careers at Google and Baidu, merged technical teams, leveraged AI-driven recommendation algorithms, and grew the short‑video platform into a multibillion‑dollar public company.
Zhao Jianbo, head of Kuaishou's big data architecture team, presents an in‑depth overview of Apache Flink's adoption at Kuaishou, covering reasons for selection, development history, business data flows, technical innovations such as the Slimbase state engine, stability improvements, and future roadmap.
This article chronicles Kuaishou's evolution from a modest GIF‑sharing tool to a market‑leading short‑video platform that listed on the Hong Kong Stock Exchange with a valuation exceeding HK$1.3 trillion, highlighting its funding rounds, product strategy, competition with Douyin, and aggressive growth campaigns.
This article introduces the development history, current architecture, abstraction methods, and key application scenarios of Kuaishou's metadata platform, highlighting challenges such as heterogeneous data integration, large-scale asset management, and the platform's role in data search, lineage, governance, and future enhancements.
This article details Kuaishou's Flink‑based real‑time computing architecture, its massive cluster scale, and the comprehensive strategies—including overload protection, system stability, pressure testing, and resource guarantees—implemented to ensure reliable streaming for the 2020 Spring Festival Gala and its real‑time dashboard.
This article describes how Kuaishou leverages Apache Flink for large‑scale real‑time multi‑dimensional analytics, details the architecture of its analytics platform using Kudu storage and KwaiBI, and introduces SlimBase—a lightweight, embedded shared state backend that replaces RocksDB to reduce I/O, latency, and CPU overhead.
In a mobile‑first world where limited device resources and unstable networks threaten user retention, Kuaishou’s performance team built an APM monitoring platform and applied systematic memory, startup, and jank optimizations that cut startup time by 40%, reduced package size by 23 MB, and significantly improved key product metrics.
This article presents Kuaishou's extensive use of Apache Flink for real-time multi-dimensional analytics, detailing the platform's architecture, cluster scale, data processing pipelines, the design of a shared state storage engine called SlimBase, and performance improvements achieved through replacing RocksDB with a customized HBase‑based solution.
During Kuaishou's 2020 Spring Festival red‑envelope campaign, Tencent Cloud's Cloud File Storage (CFS) provided a high‑throughput, read‑heavy NFS solution that handled up to 5 GB/s across 1 700 pods, delivering 100% availability, low latency, and massive interaction volumes for billions of users.
Jiaoxiao Fang’s talk details Kuaishou’s YARN deployment, covering its architecture, support for offline, real‑time and ML workloads, and recent enhancements such as event‑handling stability, refined preemption, high‑throughput parallel scheduling, shuffle‑caching for small I/O, plus plans for job protection and multi‑cluster resource utilization.