This article explains how graph models are applied in Baidu's recommendation system, covering graph basics, common graph algorithms such as graph embeddings and GNNs, and the evolution and deployment of the Feed graph model across multiple product lines.
This article presents a comprehensive overview of multi‑task and multi‑scenario algorithms applied to recommendation systems, covering background challenges, algorithm taxonomy, recent research, detailed model architectures such as TAML, CausalInt and DFFM, experimental results on public and private datasets, and a Q&A discussion.
This article examines the fairness of recommendation systems from both consumer and provider viewpoints, discussing sources of bias, definitions of equality and equity, measurement metrics such as CGF and MMF, and proposes causal embedding models to mitigate unfairness while ensuring sustainable system performance.
This article provides a detailed overview of modern recommendation system technology, covering system architecture, user understanding layers, various recall and ranking techniques, additional algorithmic directions such as cold‑start and bias modeling, and the evolving evaluation metrics used in practice.
In this talk, senior researcher Dai Quanyu from Huawei Noah's Ark Lab presents his work on applying causal inference to identify and correct various biases in recommendation systems, detailing underlying theoretical frameworks, bias‑mitigation algorithms such as inverse propensity weighting and robust learning, and real‑world product deployments.
The 23rd CCF C³ Forum, organized by Alibaba’s Alimama and the CCF CTO Club, examined how large‑model AI is reshaping intelligent business technology, from data‑driven to knowledge‑driven approaches, enhancing e‑commerce with smarter search, personalized recommendations, content creation, and guiding merchants on future AI‑native strategies.
The article surveys how large language models can be integrated into recommendation systems, either as feature extractors or as end‑to‑end recommenders, showing that LLM‑derived semantics improve recall, ranking, diversity, and user experience, and outlining future multimodal, efficiency, and re‑ranking directions.
This article introduces the fundamentals of causal inference from observational data, explains confounding and the SUTVA assumptions, presents the do‑operator, and details four de‑confounding strategies—including RCT‑based resampling, feature‑decomposition, double machine learning, and back‑/front‑door adjustments—followed by real‑world applications in recommendation systems and resource allocation.
The DataFunSummit 2023 online summit (August 26‑27) will explore eight recommendation‑system topics—including core and engineering architecture, model training/inference, large models, graphs, cold start, and multi‑task scenarios—featuring Xiaohongshu leaders who will present on graph‑based business architecture, integrated training‑inference pipelines, and user/content cold‑start strategies.
This article explains the principles of collaborative filtering, outlines its three main types—user‑based, item‑based, and model‑based—and focuses on model‑based approaches such as matrix factorization, clustering, and deep‑learning techniques for recommending personalized game items to improve player experience and monetization.
The Multi‑Granularity Attention Model (MGAM) improves group recommendation by extracting subset, group, and superset preferences through hierarchical attention and graph neural networks, fusing them via self‑attention, and achieves state‑of‑the‑art offline results and a 1.2% online CTR lift in Alibaba’s local‑life services.
In this talk, Zhang Pengtao, a Sina Weibo technical expert with a Ph.D. in computer applications, explores how the strong memory capabilities of NLP large language models inspire the design of independent memory mechanisms for recommendation systems, covering model concepts, HCNet & MemoNet, experimental results, and practical takeaways for enhancing recommendation model performance.
BASM is a bottom‑up adaptive spatiotemporal model for online food ordering that uses hierarchical embedding, semantic transformation, and adaptive bias layers to dynamically modulate parameters according to time and location, thereby capturing multiple data distributions and achieving superior offline metrics and online A/B test performance.
The paper introduces StEN, a spatiotemporal-enhanced network for CTR prediction in location-based services, combining static spatiotemporal feature activation, dynamic preference activation, and target attention, achieving state-of-the-art offline results and a 1.6% CTR lift in online tests.
Deep Learningclick-through raterecommendation systems
The article showcases seven Meituan research papers accepted at KDD 2023—spanning feed‑stream, cross‑domain, takeaway, bonus allocation, contour‑based segmentation, living‑needs prediction, and multilingual recommendation—detailing their novel methods, real‑world deployments, and concluding with an invitation for academic collaboration.
The DataFun Decision Intelligence Summit brings together leading researchers and industry experts to present cutting‑edge reinforcement learning algorithms, safety considerations, distributional methods, and real‑world applications such as vehicle routing, recommender systems, and power‑grid scheduling, highlighting future directions and audience takeaways.
Manbang Group shares its practical experiences and insights on real-time computing, multi‑cloud platform architecture, data warehousing with Flink and Holo, real‑time decision and feature platforms, and future plans for scaling these systems to support logistics and recommendation algorithms.
This article details Tubi's multi‑stage recommendation system, focusing on the recall phase and describing how popularity metrics, embedding averaging, per‑video nearest‑neighbors, hierarchical clustering, real‑time ranking, and context‑aware sampling are combined to efficiently generate personalized video candidates at scale.
This article presents Fliggy's work on user travel demand prediction, outlining the unique challenges of travel scenarios, the evolution of recall and ranking algorithms—including multi‑task learning, graph‑based models, and intention‑capture mechanisms—and discusses future research directions such as long‑sequence modeling and cross‑domain learning.
This article introduces graph neural networks, explains graph representation learning, discusses their evolution from random walks to spectral and spatial convolutions, and details how OPPO applies GNNs to improve recommendation system recall and ranking, highlighting practical architecture, experimental gains, and future research directions.
This article presents Tencent TRS's industrial deployment of meta‑learning and cross‑domain recommendation, detailing problem definitions, solution architectures, challenges of industrialization, and practical implementations that achieve personalized modeling and cost‑effective multi‑scene recommendation across various online services.
This article examines the fairness of recommendation systems from both consumer and provider viewpoints, discussing sources of bias, definitions of equality and equity, measurement metrics such as CGF and MMF, causal embedding techniques, experimental results on MovieLens and Yelp, and future research directions.
In this expert roundtable, leading researchers and engineers discuss the current state of recommendation systems, how large language models can reshape the field, the technical and practical challenges involved, and practical advice for practitioners looking to adopt AI‑driven personalization solutions.
This article explains the fundamentals of recommendation systems, their digital representation, how ChatGPT and large language models are applied to enhance recommendation performance, and highlights emerging trends such as conversational recommendation and a recommended book on the subject.
The A2M Summit announcement details a 2023 AI and machine learning conference where senior algorithm architect Liu Lixi presents his talk on practical recommendation system techniques for sparse data, low‑frequency scenarios, and ad‑creative optimization within 58.com’s industry‑wide digital transformation.
This article presents a detailed survey of recent advances in recommendation system technology, covering system architecture, user understanding layers, various recall methods, ranking techniques, auxiliary algorithms such as cold-start and bias modeling, and evaluation metrics, with references to industry practices and academic research.
The paper introduces SASS, a scenario-adaptive self-supervised recommendation model that uses contrastive pre-training and multi-layer gating to expand global samples and transfer scene-aware parameters, enabling a single model to deliver personalized recommendations across diverse Taobao ‘SuoSuo’ scenarios while mitigating data sparsity and cross-domain challenges.
This article presents the Hyper‑Actor Critic (HAC) framework that splits reinforcement‑learning policies into continuous hyper‑actions and effective recommendation lists, introduces alignment and supervised losses, and demonstrates superior performance on an online simulator compared to existing RL and supervised methods.
This article reviews the evolution of Weibo's large‑scale recommendation technology, covering the system's business scenarios, technical roadmap, recent large model iterations, multi‑task and multi‑scenario modeling, feature engineering, consistency between recall and ranking, and emerging techniques such as causal inference and graph methods.
The article details how Kuaishou’s multimodal AI research, including its K7 trillion‑parameter model and VLUA algorithm, partners with Renmin University’s Gaoling AI Institute to launch a joint lab, produce cutting‑edge papers such as WebBrain and ChatImg, and advance recommendation and search technologies across the short‑video ecosystem.
This article provides algorithm engineers with a comprehensive career roadmap covering three key professional stages, practical planning advice, technical skill sharpening, industry trend analysis, and a detailed Q&A to help navigate the volatile tech job market.
This article provides a comprehensive introduction to user profiling, covering its definition, key elements, classification types, common dimensions, practical application scenarios, lifecycle considerations, development workflow, and validation methods for building effective data‑driven user models.
This article presents a comprehensive study of multi‑scenario personalized recommendation, introducing a scenario‑adaptive and self‑supervised model (SASS) that jointly addresses data sparsity, domain adaptation, and recall‑stage deployment through a two‑stage training pipeline and extensive experiments on Alibaba’s Taobao platform.
The CCF C³ event at Xiaohongshu gathered leading researchers and industry experts to dissect the latest advances, challenges, and future opportunities in intelligent recommendation and search, including multimodal content handling, decentralized distribution, cold‑start solutions, and the impact of large language models.
At NVIDIA GTC 2023, Kuaishou engineers unveiled cutting‑edge solutions ranging from video quality assessment and enhancement, 3D digital‑human live streaming, a custom TensorRT‑based performance framework, large‑scale recommendation model acceleration, to multimodal massive‑model deployment for short‑video scenarios.
In a March 27 livestream, Alibaba senior algorithm engineer Chen Zhihong will explain how popularity bias affects recommendation pipelines, review existing mitigation techniques, and introduce a decoupled domain‑adaptive unbiased dual‑tower recall model that leverages bias while preserving recommendation fairness.
At the GTC conference, the talk presents Alibaba Cloud’s heterogeneous computing platform and introduces the Open Deep Learning API (ODLA), then details how CUTLASS‑based operator fusion dramatically accelerates attention and MLP layers in large‑scale recommendation models, achieving multi‑fold performance gains in production.
NetEase’s Fuxi team developed the multimodal ‘YuZhi’ model, a large‑scale image‑text dual‑tower system optimized with the EET inference framework, which powers personalized recommendations in NetEase News and Cloud Music, while a partnership with Huawei Ascend AI and MindSpore enables further model acceleration, compression, and the new ‘YuZhi‑Wukong’ model that improves video recommendation metrics by about 5%.
The article explains JD’s Explore & Exploit (EE) module, its bias‑related challenges, the iterative optimization loop, model debiasing techniques for position and popularity bias, personalized bias modeling, causal inference methods, online AB results, and offline evaluation metrics, highlighting significant improvements in search diversity and efficiency.
This article explains what recommendation systems are, their core tasks, evaluation metrics, popular algorithms such as collaborative filtering and latent factor models, how to handle cold‑start and contextual challenges, the role of social networks, and typical system architecture, providing a comprehensive overview for beginners and practitioners.
This article reviews the author’s six‑year hands‑on experience with deep learning, covering breakthroughs in speech recognition, computer vision, language modeling, reinforcement learning, privacy protection, model compression, recommendation systems, and future research directions, while summarizing technical lessons and practical insights.
This article presents a comprehensive overview of the background, current multi‑path recall methods, and a series of practical optimizations—including dual‑tower models, enhanced vectors, an unbiased IPW‑based framework, and a travel‑state‑aware deep recall model—applied to Feizhu's homepage recommendation system, with both offline and online experimental results demonstrating click‑through rate improvements.
bias mitigationdual-tower modelrecommendation systems
The article argues that algorithm engineers should move beyond a narrow focus on deep‑learning models, emphasizing the importance of system architecture, data quality, and thoughtful problem framing to break through performance plateaus in advertising and recommendation systems.
This interview with three recommendation‑system experts explores the technical architecture, data sources, feature engineering, recall and ranking strategies, evaluation metrics, cold‑start solutions, and practical difficulties, offering actionable insights to avoid common pitfalls in real‑world recommender deployments.
This article presents a detailed study on applying recommendation‑system techniques to test case prioritization for Bilibili's mobile apps, describing the problem definition, evaluation metrics, data processing, FM model selection, experimental results, practical deployment, and future research directions.
BilibiliTest Case Prioritizationfactorization machine
Vivo’s AI recommendation team leveraged Alibaba’s DeepRec engine—introducing dynamic Embedding Variables, feature admission/elimination, Parquet datasets, and advanced CPU/GPU inference optimizations such as SessionGroup, device placement, multi‑stream and BladeDISC compilation—resulting in notable gains in model accuracy, latency reduction, and resource efficiency.
In this interview, Li Benyang, head of the intelligent recommendation content understanding team at Autohome, shares his AI background, the evolution of recommendation systems, knowledge‑graph construction for automotive content, career choices between big firms and startups, and practical advice for technologists navigating fast‑changing industries.
This article explains why and how sample weighting is used in machine learning, illustrates YouTube's conversion of video watch time into sample weights to align with its commercial goals, and describes practical weighted‑logistic‑regression techniques applied to improve game recommendation systems.
This article explains why traditional CPU inference and naïve GPU usage are inefficient for recommendation workloads, introduces NVIDIA Multi‑Process Service (MPS) technology, describes VIVO's custom Rust‑based inference engine and deployment strategies, and presents performance and cost benefits along with practical deployment considerations.
This article presents the practical evolution of NetEase Yanxuan's recommendation pipeline, covering background, multi‑objective and cross‑domain modeling, bias correction, loss function enhancements, long‑term value strategies, and multi‑scene modeling, with experimental results and a Q&A session.
The study reveals that industrial deep click‑through‑rate models often overfit dramatically after the first training epoch—a “one‑epoch phenomenon” caused by the embedding‑plus‑MLP architecture, fast optimizers, and highly sparse features, with performance dropping sharply unless sparsity is reduced or training is limited to a single pass.
SimRank, a graph‑theoretic recommendation algorithm, models users and items as a bipartite graph and computes similarity through iterative matrix operations, with extensions like SimRank++ incorporating edge weights and evidence, while scalable solutions use big‑data frameworks or Monte‑Carlo simulations.
This article presents a comprehensive overview of feature embedding modeling in recommendation systems, discussing the necessity of feature modeling, three technical directions (gate threshold, variable‑length embeddings, and enrichment), detailed descriptions of models such as FiBiNet, FiBiNet++, ContextNet, and MaskNet, experimental findings, and a Q&A session that addresses practical challenges and future work.
This article explores the evolution of recommendation systems for interactive live‑streaming scenarios, covering common system traits, user cold‑start solutions, prior knowledge modeling, scene‑specific modeling, and practical Q&A insights drawn from Tencent Music’s real‑world deployments.
Alibaba’s research team secured three NeurIPS 2022 papers—introducing an Adaptive Parameter Generation network that boosts click‑through rates and revenue, a tuning‑free Global Batch Gradient Aggregation method that speeds recommendation model training by 2.4×, and a Sustainable Online Reinforcement Learning framework that outperforms existing auto‑bidding strategies.
This talk explores the evolution, objectives, and core challenges of pretraining models, their application in recommendation scenarios, service modes, and detailed case studies of graph neural network pretraining, illustrating how self‑supervised learning and multi‑domain data integration enhance user and item embeddings for improved recommendation performance.
In the era of deep learning, feature engineering remains crucial for recommendation and search advertising because it bridges raw relational data and models, improves performance, reduces complexity, and handles high‑cardinality, large‑scale, and time‑sensitive scenarios with robust transformations and statistical encoding.
This article reviews how graph neural networks are applied across the three stages of recommendation systems—recall, ranking, and re‑ranking—detailing novel models such as NIA‑GCN, GraphSAIL, and DGENN, their experimental improvements, and future research directions.
The article details Meituan Waimai's use of graph neural network techniques—feature‑graph crossing, subgraph expansion, and metapath‑based scene graphs—to model user‑restaurant interactions across location, time, and context, describing the engineering pipeline, online serving optimizations, and offline AUC improvements of up to 2.5 ‰ for high‑ and low‑frequency scenarios.
The team enhanced real‑time recommendation by redesigning TensorFlow graphs—using constant‑folding, a custom CallGraphOP cache, a simplified dense layer, and CUDA‑Graph compatibility—boosting single‑machine throughput ~40%, raising GPU utilization from 30% to 43%, cutting latency and saving roughly 30% of hardware resources.
CTNet is a continual transfer learning framework that uses a lightweight Adapter to map source‑domain features onto evolving target‑domain recommendation tasks, preserving all model parameters to avoid catastrophic forgetting and delivering substantial gains in click‑through rate, conversion, and overall business performance in Taobao’s cross‑domain e‑commerce scenario.
This article explains why meticulous feature engineering remains crucial in recommendation and search advertising, outlines what constitutes good features, describes common transformation techniques such as scaling, binning, and encoding, and provides practical examples and Q&A for practitioners.
The authors introduce AGE, an adversarial‑gradient‑driven exploration framework that injects uncertainty‑scaled perturbations into ad embeddings to approximate the downstream learning effect, combines Monte‑Carlo dropout uncertainty, a dynamic gating unit, and achieves up to 15 % offline gains and 6 % online CTR improvement over strong baselines.
The article presents a comprehensive technical overview of a live‑streaming recommendation system, covering common and specific characteristics, user cold‑start strategies using unbiased clustering, prior knowledge integration, multi‑task modeling, and scene‑aware routing to improve relevance and engagement in interactive environments.
Disney’s Content Discovery team introduces M5, a multi‑modal, multi‑interest, multi‑scenario recall model that enhances VOD and live streaming recommendations by leveraging rich metadata, user behavior, and contextual features, outperforming baseline methods with significant hit‑ratio gains across Hulu and Disney+.
At the 16th OSDI conference, Tencent’s WeChat team presented the award‑winning Ekko system—a groundbreaking, ultra‑low‑latency model‑update solution for massive recommendation workloads that dramatically speeds up updates, supports over a trillion‑scale models, and has already boosted user engagement across billions of daily users.
The article introduces collaborative filtering, derives the matrix‑factorization model R≈X·Yᵀ with L2‑regularized ALS updates, demonstrates a full Python example on a small rating matrix, then shows how to implement and scale Spark’s ALS for massive user‑item data, ending with production tips and references.
This article presents a comprehensive overview of conversion‑rate (CVR) optimization for Alitrip’s travel recommendation platform, detailing the challenges of extremely sparse user feedback, the design of item, user, query and context features, and a series of model‑level and loss‑function techniques—including generic‑label modeling, global‑transaction modeling, ESMM, rank‑loss approximations, and multi‑task CTR auxiliary training—to improve both CTR and CVR performance in high‑ticket‑price scenarios.
This article presents an in‑depth overview of Tencent's Wuliang deep learning platform for recommendation systems, detailing the real‑time data challenges, high‑throughput requirements, parameter‑server architecture, model compression techniques, multi‑level caching, and answers to common technical questions.
This article reviews several mature knowledge‑graph applications, describing Meituan’s large‑scale “Meituan Brain” for lifestyle services, the Fourth Paradigm’s Sage Knowledge Base platform with various representation‑learning models, and additional use cases in recommendation, QA, drug discovery, and power‑grid domains.
AI applicationsGraph Neural NetworkIndustry Case Study
This article presents a comprehensive overview of search term recommendation in QQ Browser, covering various recommendation scenarios, challenges, query library architecture, multi‑task ranking models, coarse‑to‑fine ranking pipelines, auto‑completion strategies, and future research directions.
The paper proposes CDAN, a dual‑tower retrieval model that separates item attribute and popularity representations via a Feature Decoupling Module with orthogonal embeddings, aligns head‑tail attribute distributions using MMD and contrastive learning, and jointly trains biased and unbiased towers, achieving higher tail recall, lower exposure concentration, and measurable online click‑through improvements.
Contrastive Learningdomain adaptationdual-tower model
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.
The GoodFuture International Algorithm Team, together with Jinan University Guangdong Smart Education Research Institute, distinguished themselves among 95 global teams in the 5th Educational Data Mining in Computer Science Education Workshop, securing a champion title in one task and a runner‑up in another, showcasing advanced AI‑driven predictive and recommendation techniques for intelligent student assessment.
Educational Data Miningmachine learningpredictive modeling
Hello’s automated C‑side growth algorithm loop integrates diverse traffic sources, semi‑supervised PU‑learning, graph‑based look‑alike targeting, causal uplift models for smart subsidies, and adaptive copy and external ad optimization, dramatically boosting ride‑hailing and lifestyle service revenue while minimizing engineering duplication.
This article explains the collaborative denoising autoencoder (CDAE) for recommendation, walks through data preparation with MovieLens, shows a full PyTorch Lightning implementation, tunes hyper‑parameters using Ray Tune and CometML, and reports detailed evaluation metrics.
The paper introduces Sliding Spectrum Decomposition (SSD), a tensor‑based method that quantifies feed diversity through singular‑value volume within sliding windows, integrates quality‑exploration trade‑offs, and employs a hybrid CB2CF model for item embeddings, achieving superior offline and online performance versus DPP in Xiaohongshu’s feed.
The paper demonstrates how self‑supervised models SwAV and M6 generate high‑quality image and multimodal embeddings for Meiping Meiwu’s recommendation system, delivering notable gains in scene/style consistency, ranking AUC, classification and retrieval performance, especially for cold‑start items, and achieving measurable production lifts.
The article reviews a TheWebConf 2022 paper that introduces a two‑stage framework combining differential‑privacy‑based random subspace publishing (using Johnson‑Lindenstrauss and sparse‑aware transforms) with asymmetric deep models to achieve accurate, privacy‑preserving cross‑domain recommendation, and discusses broader differential‑privacy applications.
This article reviews the development of CTR prediction models from early collaborative‑filtering methods to modern deep‑learning approaches, discusses core challenges such as feature interaction and continuous‑feature embedding, introduces recent Huawei solutions like AutoDis and ScaleFreeCTR for efficient large‑embedding training, and outlines future research directions.
This article explains how JD's intelligent recommendation selling point system leverages NLP, BERT, Transformer and pointer‑generator models to automatically create short, personalized product highlights, describing the technical background, system architecture, model training pipeline, online/offline monitoring, and the resulting business benefits.
This article introduces the ESMM (Entire Space Multi‑Task Model) proposed by Alibaba, explaining how it tackles sample selection bias and data sparsity in post‑click conversion rate (CVR) prediction through shared embeddings and implicit pCVR learning, and provides a detailed implementation using the EasyRec framework with code examples.
Coarse ranking bridges recall and fine ranking by trimming tens of thousands of candidates to a few hundred or thousand using a three‑part framework—sample construction, ordinary and cross‑feature engineering, and evolving deep models—from rule‑based to lightweight MLPs, while employing distillation, feature crossing, pruning, quantization, and bias mitigation to balance accuracy with strict latency constraints.
The article examines cold‑start challenges for rich‑media recommendations, outlines detection via calibration and lifecycle monitoring, and proposes two remedies—the multi‑stage “rise channel” for promoting fresh content and cross‑modal understanding using CLIP, CB2CF and dual‑tower models—demonstrating NetEase Cloud Music’s 25% distribution boost, over 20% CTR rise, and 40% review‑work reduction.
This article presents a comprehensive overview of QQ Music's recall pipeline, detailing business characteristics, challenges such as noisy user behavior and cold‑start, and four major solutions—including knowledge‑graph‑enhanced recall, sequence‑based and multi‑interest modeling, audio‑based recall, and federated learning—along with practical insights and Q&A.
To tackle the massive, multimodal tagging challenge of short‑video platforms—characterized by a huge long‑tail tag set, sparse annotations, and uneven modality contributions—the authors propose a two‑stage recall‑ranking system that first retrieves candidates via text, visual, audio and classification cues, then refines them with contrastive learning and extensive hard‑negative sampling, achieving 0.884 tag accuracy in a real‑world news video recommender.
Alibaba’s research team had five papers accepted at the prestigious SIGIR 2022 conference in Madrid, covering innovations such as joint ad‑ranking and creative selection, personalized bundle generation, calibrated neural predictions, disentangled counterfactual regression, and cold‑start user recommendation, showcasing strong expertise in information retrieval and online advertising.
The paper tackles Taobao Live’s cold‑start problem for new users by introducing a multi‑channel lifelong product‑sequence network that enriches purchase histories with side information, extracts relevance‑focused subsequences across five channels, and integrates them via target‑attention DIN, achieving substantial offline and online performance gains, especially for low‑activity users.
This article details NetEase Cloud Music's technical approach to building implicit user relationship chains—using SimHash, Item2Vec, and MetaPath2Vec embeddings, large‑scale vector search, and a unified service architecture—to address cold‑start challenges across multiple business scenarios.
This article explores how combining causal inference with machine learning can uncover subtle correlations in large datasets, detailing user growth metrics, propensity‑score matching, causal recommendation models, heterogeneous treatment effect analysis, and practical strategies for improving retention and activity in recommendation systems.
The article introduces KuaiRec, a densely observed short‑video recommendation dataset with 99.6% density covering 1,411 users and 3,327 videos, discusses its structure, advantages over sparse public datasets, and its applicability to unbiased, interactive, conversational and reinforcement‑learning based recommendation studies.
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.
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 traces the development of embedding methods—from the early word2vec model through item2vec, DeepWalk, Node2vec, EGES, HERec, GraphRT, and target‑fitting approaches like DSSM and YouTube recommendation—highlighting how sequence‑construction and target‑fitting paradigms have shaped modern recommendation systems and AI applications.
The Co‑Action Network replaces costly Cartesian‑product feature crossing with lightweight micro‑net‑based interaction units that share parameters across feature pairs, delivering comparable CTR prediction accuracy while cutting parameters to one‑tenth and boosting online latency, as proven in large‑scale advertising deployments.
Co-Action Networkfeature interactionrecommendation systems
This article presents a comprehensive study of graph‑based recall techniques for content recommendation, detailing how knowledge‑graph‑augmented user‑behavior graphs and novel attention‑driven models such as GADM, SGGA, and SGGGA improve performance for users with sparse interaction histories.
The paper proposes graph‑based deep recall models that enrich sparse user behavior sequences in video recommendation by integrating content knowledge graphs and adaptive attention mechanisms, demonstrating that variants such as GADM, SGGA, and SGGGA significantly boost click‑through rates in online experiments.