Alibaba researchers propose UTTSI, a training-free, model-agnostic framework that estimates per-sample uncertainty at inference time using frequency priors and gradient-based attribution, then adaptively allocates compute — filtering noisy features and exploring multiple inference paths only for uncertain samples — achieving a 5.3% CTR lift in a 7-day online A/B test with just 2.8× average model calls.
Alibaba International Intelligent Technology presented ten CIKM 2026 papers that introduce multimodal large‑model techniques—such as SAM‑D2Q, C2P, PRO‑Bid, UTTSI, GRC, CDNet, CAIM and SORT—to overcome bottlenecks in e‑commerce search, recommendation, auto‑bidding and CTR prediction, delivering substantial offline metric lifts and significant online gains in CTR, CVR, GMV and revenue.
The article analyzes the evolution from traditional modular recommendation models to a generative Large User Model (LUM), detailing its three‑stage paradigm, tokenization, training objectives, scaling‑law findings, offline and online experiments, and the AI‑infra innovations that enabled a 4.5% CTR lift in production.
The DMF framework introduces a modality‑enriched Decoupled Target Attention (DTA) and a Complementary Modality Modeling (CMM) strategy that together close the semantic gap between ID and multimodal features, delivering up to 5.3% CTCVR lift, 7.43% GMV increase, and a three‑fold throughput gain in large‑scale e‑commerce recommendation scenarios.
This article explains a SIGIR 2025 paper that tackles cold‑start click‑through‑rate prediction in JD's ad system by using a Graph Isomorphism Network‑based cohort modeling framework, detailing its three‑module architecture, extensive experiments on public and industrial datasets, and a live deployment that achieved a 2.13% CTR lift.
This article explains Baidu's generative ranking model GRAB, detailing how scaling laws from large language models inspire a new recommendation paradigm, the model's architecture, custom attention mechanisms, training strategies, deployment optimizations, and the resulting business gains in CTR and revenue.
JD's retail technology team presents five SIGIR 2025 papers that introduce advanced graph neural, causal optimal transport, domain‑oriented relevance, multi‑objective bid‑word generation, and hierarchical user behavior models to dramatically improve precise matching in e‑commerce search, recommendation, and advertising.
This article showcases five JD Retail Technology research papers accepted at SIGIR 2025, covering graph‑based cohort modeling, causal optimal transport post‑event modeling, an autonomous domain‑oriented relevance engine, a multi‑objective bidword generation model, and hierarchical long‑term user behavior modeling, all advancing e‑commerce CTR prediction and advertising.
In 2024 Alibaba Mama outlines how large‑language models transform search advertising through a three‑line scaling roadmap—explicit inductive‑bias design, implicit compute growth, and auxiliary CV/NLP advances—implemented via a pre‑train/post‑train/CTR paradigm and the LUM user‑behavior model, promising gains in relevance, recall, and real‑time serving while highlighting inference efficiency challenges.
This paper proposes CAIG, a novel method for generating high-CTR advertising images using multimodal large language models, combining reinforcement learning and preference optimization to align generated content with product features.
CTR predictionMultimodal Large Language Modelsadvertising image generation
The paper presents a comprehensive multi‑scenario recommendation study introducing three models—SACN, SAINet, and DSWIN—that integrate scene‑aware attention, attribute‑level preferences, and contrastive disentanglement to capture distinct user interests, achieving consistent AUC gains and online CTR improvements across real‑world datasets.
This article highlights eight Meituan research papers accepted at CIKM 2024—spanning self‑supervised sequential recommendation, rating‑consistent explanation generation, CTR prediction via recommendation pre‑training, cross‑domain interest transfer, multimodal vector retrieval, design‑aware poster layout, order‑fulfillment cycle‑time forecasting, and delivery‑scope substitution—offering insights from both internal and university collaborations.
AI researchCTR predictionCross‑Domain Recommendation
This study investigates how adding an auxiliary ranking loss to click‑through‑rate (CTR) models not only improves ranking but also alleviates gradient‑vanishing for negative samples, thereby boosting the primary classification performance, especially under sparse positive‑feedback conditions.
This article presents a comprehensive study on integrating multimodal image‑text representations into large‑scale e‑commerce advertising CTR models, introducing a semantic‑aware contrastive pre‑training (SCL) method and two application algorithms (SimTier and MAKE) that together achieve over 1 % GAUC improvement and significant online gains.
Alibaba’s advertising team introduces semantic‑aware contrastive learning to pre‑train multimodal image‑text embeddings, integrates them via SimTier and MAKE into ID‑based CTR models, achieving up to 6.9% lift in Taobao display ad click‑through rates and improving long‑tail item performance.
This paper investigates the embedding dimensional collapse problem that hinders scaling of recommendation models and proposes a Multi-Embedding paradigm that learns multiple embeddings per feature with independent expert networks, demonstrating consistent performance gains across major CTR benchmarks and real‑world ad systems.
The article describes Alibaba's approach to distilling privileged features for CTR prediction using a calibration-compatible listwise distillation loss (CLID) that normalizes teacher and student outputs within sessions to align top‑ranking probabilities, improving both accuracy and ranking while preserving calibration.
AI in advertisingCTR predictionListwise distillation
The paper introduces FIN (Fragment and Integrate Network), a novel spatial‑temporal model that extracts multiple sub‑sequences from ultra‑long user behavior logs, applies simplified and multi‑head attention, and fuses them with physically meaningful set operations, achieving up to 5.7% CTR lift and 7.3% RPM improvement in real‑world food‑delivery advertising.
This article explains the fundamentals of graph neural networks and graph representation learning, outlines how graphs enhance recommendation systems, and details OPPO's practical implementation of a hybrid dual‑tower and graph sub‑network model to improve recall and ranking 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.
Eight Alibaba Mama team papers accepted at CIKM 2023 present advances such as task‑specific bottom‑representation networks for recommendation, a unified GNN for multi‑scenario e‑commerce search, multi‑slot bid shading, consistency‑oriented pre‑ranking, bias‑mitigating CTR prediction, efficient progressive‑sampling self‑attention, delayed‑feedback conversion modeling, and hybrid contrastive multi‑scenario ad ranking.
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.
Maria is a multi‑scenario ranking framework that adaptively learns features across heterogeneous e‑commerce query types—visual search, similar‑product search, and interest search—by employing Feature Scaling, Feature Refinement, and Feature Correlation Modeling modules, achieving superior performance and reducing the seesaw effect on the Ali‑CCP and Alimama datasets.
To eliminate data fragmentation, feature inconsistencies, and multi‑language implementation challenges, Bilibili built a unified real‑time advertising feature platform that aligns offline, hourly, and online pipelines via a shared C++ library and JNI, boosting CTR prediction accuracy, cutting training costs, and increasing ad revenue by over 1 %.
The Joint Optimization of Ranking and Calibration (JRC) model introduces a two‑logit generative‑discriminative architecture that jointly minimizes LogLoss for calibration and a listwise ranking loss, delivering superior GAUC and CTR performance across Alibaba’s display‑ad system, especially for sparse long‑tail users, while remaining simple to train and deploy.
AdaSparse introduces an adaptive sparse network that learns a dedicated sub‑network for each advertising scenario, balancing shared and specific knowledge while keeping computational cost low, and achieves +4.63% CTR and -3.82% CPC improvements in Alibaba’s external ad system, as validated on both public and massive production datasets.
Adaptive Parameter Generation Network (APG) dynamically creates sample‑specific model parameters for click‑through‑rate prediction using low‑rank factorization, parameter sharing, and over‑parameterization, achieving up to 0.2% AUC improvement, 3% CTR lift, and up to 96.6% storage reduction with faster inference.
The article proposes a three-stage contrastive learning framework—pre‑training, fine‑tuning, and debiasing—to generate unbiased, fine‑grained image embeddings for mobile Taobao CTR prediction, achieving higher accuracy, fairness, and a 4‑5% CTR lift in large‑scale offline and online evaluations.
Decoupled Graph Neural Networks (DC‑GNN) improve large‑scale e-commerce ad recall by separating graph processing from CTR prediction, using multi‑task pretraining (edge prediction + contrastive learning), efficient deep linear aggregation, and a dual‑tower CTR model, achieving higher efficiency and performance on billions‑scale data.
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 article details a step‑by‑step evolution of an e‑commerce platform’s top‑slot recommendation system, moving from a DCN‑mix single‑objective model through BST‑based dynamic features, position‑bias debiasing, multi‑task MMoE learning, and finally BST with target‑attention, each yielding measurable CTR, conversion, and user‑value gains.
This article presents Hello's Intelligent Creative project, detailing its business motivations, system architecture, algorithmic choices such as seq2seq, VAE, GAN, and pre‑trained models, the implementation of material libraries, tagging, recall strategies, a creative racing model, performance gains, and future challenges.
This article presents JD's comprehensive technical solution for boosting recommendation ad ranking by addressing cold‑start, shallow user interest extraction, and insufficient global data through a variational feature learning framework, enhanced user‑interest networks, and full‑domain collaborative modeling, achieving over 1% AUC gain and notable revenue growth.
This article presents JD's end‑to‑end improvements for advertising click‑through‑rate prediction, addressing cold‑start, deep user‑interest mining, and full‑domain collaborative information through a variational feature learning framework, enhanced interest networks (PPNet+, NeNet, Weighted‑MMoE) and exposure‑sequence modeling, achieving over 1% cumulative AUC gain and publication in top conferences.
This article reviews the evolution of recommendation models from early collaborative filtering to modern deep learning approaches, discusses core challenges such as CTR prediction, outlines user‑behavior and combination‑feature modeling techniques, introduces large‑embedding training and continuous‑feature embedding methods like AutoDis, and presents distributed training frameworks such as ScaleFreeCTR, concluding with future research directions.
This article explains how MaFengWo's recommendation ranking system tackles position bias by incorporating position features, using inverse propensity weighting, and adjusting click metrics, resulting in measurable improvements in click‑through rate, content exposure, and overall recommendation accuracy.
The article provides a thorough guide to ranking in recommendation systems, detailing the pipeline architecture, sample handling challenges, extensive feature engineering categories, the evolution from collaborative filtering to deep and attention‑based models, and key optimization trade‑offs between memorization, generalization, and efficient user‑interest modeling.
This article presents a comprehensive overview of NetEase Cloud Music's advertising system, detailing its architecture, core challenges, CTR and CVR prediction models, feature engineering, model evolution from LR to deep learning, user vector modeling, and practical recommendations for improving ad performance.
The article presents JD's technical evolution for advertising ranking, covering recommendation ad sorting, context‑aware CTR re‑estimation, reinforcement‑learning‑based sequence optimization, and session‑level auction mechanisms, and includes a Q&A that highlights practical gains and implementation challenges.
This report details how the Search Recommendation team at 58.com upgraded their deep learning ranking model for recruitment by adding multi-valued and semantic vector features, integrating conversion sequences, employing feature‑crossing techniques, optimizing offline data pipelines, and planning future multi‑scene improvements to boost CTR and relevance.
The article details Shopee Chatbot’s end‑to‑end development of a dropdown recommendation feature, describing the retrieve‑then‑rank architecture with BM25 and vector recalls, multilingual pre‑training and distillation, DeepFM‑based ranking, experimental gains in CTR and conversion, deployment infrastructure, business impact, and future enhancements.
The article presents ZEUS, a self‑supervised multi‑scenario ranking model that leverages user‑initiated behavior pre‑training to break feedback loops and improve query recommendation efficiency across diverse e‑commerce search scenarios, achieving significant gains in CTR, CVR, and GMV.
This article presents a three‑part technical guide covering the fundamentals of computational advertising and real‑time bidding, detailed offline pCTR model training pipelines with feature engineering, calibration and model structure improvements, and advanced online learning techniques such as parameter freezing, sample replay and knowledge distillation, all aimed at boosting CTR performance and reducing bias in large‑scale ad platforms.
Meituan's ad team uses deep learning to handle LBS distance constraints and long‑term periodic behavior, introducing DPIN for position/context bias, an ultra‑long sequence encoder with spatiotemporal activator, dynamic candidate generation, and memory‑augmented continual learning, boosting RPM 2‑20% and enabling sub‑millisecond inference.
The 58.com AI algorithm competition showcased over 210 teams competing to improve job recommendation click‑through and conversion rates, featured an award ceremony with speeches, highlighted the ten winning teams, and presented detailed solution shares—including tree models, feature‑engineering techniques, and deep‑learning approaches—while offering GPU resources on the WPAI platform for continued participation.
This article reviews the characteristics of e‑commerce personalized prediction, traces Alibaba’s advertising CTR model evolution from large‑scale logistic regression through deep learning architectures such as DIN and CrossMedia, and discusses future research directions like representation learning and white‑box modeling.
The article presents a hybrid bandit framework combined with a visual‑aware ranking model to efficiently select and dynamically optimize advertising creatives, addressing cold‑start challenges, element‑level personalization, and production‑parameter search, and validates the approach with extensive offline and online experiments.
The article surveys six Alibaba Mama papers accepted at CIKM 2021, presenting novel AI methods—including a heterogeneous graph neural network for keyword matching, a star‑topology multi‑domain CTR model, a compact hash embedding technique, adaptive masked twins layers, automated hierarchical conversion prediction, and a scalable multi‑view ad retrieval system—each demonstrating substantial online performance improvements and large‑scale deployment.
The article describes Baidu’s end-to-end consistency testing framework for advertising click-through-rate models, which uses a five-stream verification pipeline and six implementation phases to compare Q-values across feature extraction, table conversions, and DNN computation, enabling precise detection and localization of data and model inconsistencies in production.
This article provides an extensive retrospective and forward‑looking analysis of the evolution of click‑through‑rate prediction technologies in online advertising, covering shallow‑learning era challenges, the rise of industrial‑scale deep learning, system‑level innovations such as recall, coarse‑ranking, fine‑ranking, bidding, and the emerging co‑design of algorithms, compute, and architecture.
This article describes how Alibaba's advertising team tackled the challenges of modeling long‑term user interests for CTR prediction by co‑designing incremental computation services, introducing memory‑network‑based models (MIMN and HPMN), and achieving significant offline and online performance gains.
This article describes how Alibaba's search advertising team transformed trillion‑parameter CTR models into lightweight, high‑precision systems by compressing embedding layers through feature‑space reduction, dimension quantization, and multi‑hash techniques, while also introducing graph‑based pre‑training and dropout‑driven feature selection to maintain accuracy.
The Meituan team introduces DPIN, a three‑module deep network that jointly models ads and their positions to mitigate position bias in CTR prediction, achieving up to 2.98% AUC improvement, 2.25% higher CTR and 2.15% RPM gains while keeping latency modest, and is applicable to broader ranking tasks.
This article presents the Search‑Based Interest Model (SIM), a two‑stage retrieval framework that indexes a user's entire behavior history to enable long‑term interest modeling for click‑through‑rate prediction, demonstrating practical deployment and improved recommendation of long‑term interests in e‑commerce.
This article systematically reviews the evolution of click‑through‑rate (CTR) prediction models—from early distributed linear models like logistic regression, through automated feature engineering with GBDT+LR, various factorization‑machine variants, embedding‑MLP shallow modifications, dual‑tower combinations, and advanced explicit feature‑cross networks—highlighting each model’s structure, advantages, limitations, and comparative insights.
This article details the development and deployment of CAN (Co‑Action Net), a novel click‑through‑rate prediction model that captures item‑item co‑action via attention‑based slot embeddings, offering superior performance to Cartesian‑product methods while reducing parameter and serving costs.
This article details how 58.com leverages deep learning models such as DNN, Wide&Deep, DeepFM, DIN and DIEN, combined with extensive user‑behavior feature engineering, offline vectorization, and online TensorFlow‑Serving pipelines to improve home‑page recommendation click‑through rates and overall platform efficiency.
This article introduces a Kalman Filtering Attention (KFAtt) framework that enhances click‑through‑rate (CTR) prediction by modeling user behavior with a Kalman‑filter‑based attention mechanism and a frequency‑capped variant, addressing new‑interest coverage and frequency bias in e‑commerce scenarios.
This article introduces the Deep Sparse Network (NON), a new deep neural architecture for recommendation systems that combines field‑wise networks, across‑field interaction networks, and an operation‑fusion network, and demonstrates its superior performance through extensive experiments and ablation studies.
This article reviews the MiNet model, which leverages cross‑domain information by modeling long‑term, source‑domain short‑term, and target‑domain short‑term user interests with hierarchical attention and an auxiliary task to improve CTR prediction and alleviate cold‑start issues.
This article presents a comprehensive study of the Wide & Deep architecture applied to user purchase‑intent quantification in the real‑estate domain, detailing feature engineering, model design, training procedures, experimental results, and extensions with GRU‑based sequential modeling to improve accuracy.
The paper presents an optimized CTR prediction model for Tencent’s App Store that merges multi‑behavior shared embeddings, long‑term DeepWalk graph embeddings, BERT‑derived app description vectors, and attention‑based fusion, reducing parameters while improving bias, AUC, and recommendation performance for sparse, long‑tail data.
This article describes how 58.com applied deep‑learning techniques—including feature engineering, sample construction, model evolution from Wide&Deep to DIN/DIEN and multi‑task learning—and system‑level optimizations to improve CTR/CPM performance in its large‑scale commercial ranking platform.
This article presents CSCNN, a category‑specific convolutional neural network that integrates visual priors into click‑through‑rate (CTR) models for JD.com’s e‑commerce advertising, detailing its motivation, architecture, engineering optimizations, offline and online training strategies, and empirical performance gains on both public and industrial datasets.
This article details the end‑to‑end deep‑learning workflow for click‑through‑rate (CTR) prediction in 58.com’s commercial ranking system, covering system architecture, feature engineering, sample construction, model evolution from Wide&Deep to DIN/DIEN, and engineering optimizations that together yielded significant CPM and CVR improvements.
This article presents Alibaba's comprehensive approach to user modeling for search CTR/CVR ranking, detailing the abstraction of user information, multi‑scale behavior processing, enhanced transformer‑based model structures, client‑side click and exposure modeling, and experimental results showing significant AUC improvements.
This article details the end‑to‑end technical architecture of WeChat's "Look" personalized recommendation service, covering data collection, recall, multi‑stage ranking, various CTR and multi‑objective models, reinforcement‑learning based mixing, diversity optimization, and the engineering hurdles overcome to deploy these solutions at massive scale.
The paper introduces a Category‑Specific Convolutional Neural Network (CSCNN) that jointly leverages product category information and visual features of e‑commerce images to improve click‑through‑rate (CTR) prediction, detailing its architecture, training strategy, large‑scale experiments, and significant performance gains in JD.com’s advertising system.
The article reviews recent research on representing massive sparse features in click‑through‑rate (CTR) models, introducing Alibaba's Res‑embedding method and Google's Neural Input Search (NIS) approach, and discusses how these techniques improve embedding efficiency and model generalization in large‑scale recommendation systems.
This article reviews recent innovations in CTR prediction for an intelligent marketing platform, covering deep spatio‑temporal networks, deep memory networks, and a feature‑expression‑assisted learning framework, with system architecture details, experimental results, and references to KDD and IJCAI papers.
This article surveys the application of attention mechanisms in deep learning recommendation systems, reviewing models such as AFM, DIN, DIEN, DSIN, Behavior Sequence Transformer, Deep Spatio‑Temporal Networks, and ATRank, and discusses their architectures, attention types, advantages, and limitations.
This article presents the design, challenges, and experimental evaluation of DSTN (with pooling, self‑attention, and interactive‑attention variants) and MA‑DNN models for CTR prediction, highlighting how temporal and contextual ad information improves accuracy and yields significant online gains in large‑scale advertising systems.
This talk presents a dynamic creative optimization framework that combines style‑material selection with DSA modeling to address the combinatorial explosion in CTR prediction, and introduces DeepMCP, an auxiliary network that improves feature embeddings through user‑ad and ad‑ad relationships, achieving superior performance in large‑scale advertising systems.
The article presents the Deep Match to Rank (DMR) model, which integrates collaborative‑filtering inspired user‑to‑item relevance modeling into the ranking stage of recommendation systems, achieving significant offline and online improvements in click‑through rate and revenue metrics for e‑commerce platforms.
This article describes the end‑to‑end recommendation pipeline of the Province Money Fast Report app, covering business background, data collection, model training and evaluation, the evolution from FM to DeepFM, DIN, DCN, xDeepFM, ESMM and custom networks, as well as serving strategies and practical lessons learned.
The paper proposes the Deep Match to Rank (DMR) model, which integrates matching‑stage collaborative‑filtering ideas into the ranking stage to explicitly represent user‑to‑item relevance, thereby enhancing personalization and achieving significant CTR and DPV improvements in e‑commerce recommendation scenarios.
This article describes iQIYI’s dual‑DNN ranking architecture that combines a high‑capacity teacher network with a lightweight student network via online knowledge distillation, addressing the trade‑off between model effectiveness and inference efficiency in large‑scale recommendation systems.
iQIYI’s dual‑DNN ranking model uses an online teacher‑student knowledge‑distillation framework where a complex teacher DNN shares representations with a lightweight student DNN, enabling end‑to‑end training, large‑scale feature crossing, and substantially higher recommendation accuracy while cutting inference latency and model size.
This article presents the Deep Match to Rank (DMR) model, which integrates matching and ranking stages in recommendation systems by jointly learning user‑to‑item and item‑to‑item representations with attention mechanisms, achieving significant CTR and DPV improvements in both offline experiments and large‑scale online deployments.
This article details Alibaba’s Taobao search CTR/CVR user modeling approach, covering background, model architecture with self‑attention and attention pooling, handling short‑term, long‑term, and on‑device behavior sequences, experimental results showing AUC improvements, and future directions.
This article details a novel Multi‑Layer Multi‑Frequency streaming training approach that enables minute‑level real‑time updates of massive CTR/CVR models by partitioning weights into freezing embeddings, changing embeddings, and changing weights, demonstrating significant offline and online AUC gains, especially during high‑traffic events like Double 11.
This article describes a standardized workflow for feature collection, configuration, processing, and model training/prediction in large‑scale recommendation systems, using CSV‑based definitions and code generation to ensure consistency between offline training and online serving while reducing manual coding effort.
This article details Kuaikan Comics' personalized recommendation pipeline, covering business context, diverse content formats, technical challenges, content‑based and collaborative‑filtering methods, ranking models, system architecture, A/B testing, and future directions for improving recommendation quality.
This article explores the evolution of recommendation algorithms at Xueqiu, highlighting the limitations of traditional point-wise click-through rate prediction models and detailing the ongoing transition to a pair-wise TF-Ranking framework designed to mitigate user and content biases while significantly enhancing overall recommendation accuracy and user experience.
This article presents a co-designed algorithm‑system solution—MIMN and an independent UIC module—that enables ultra‑long user behavior modeling for click‑through rate prediction, delivering significant offline AUC gains and online CTR/RPM improvements in Alibaba's display advertising platform.
This article explains how Alibaba's App recommendation pipeline integrates marketing scenario cards using weak personalization and machine‑learning models, detailing the metrics, feature engineering, recall and ranking strategies that together raise exposure revenue and click‑through performance.
The article reviews the characteristics of e‑commerce personalized prediction, outlines Alibaba's model iteration from large‑scale linear regression to deep learning architectures such as DIN, CrossMedia, and Deep Interest Evolution, and discusses future directions like disentangled representation and white‑box modeling.
Alibaba's X‑DeepLearning (XDL) is an open‑source deep‑learning framework optimized for high‑dimensional sparse data, offering industrial‑grade distributed training, built‑in CTR/recommendation algorithms, structured compression, and online learning capabilities, with benchmark results demonstrating superior scalability and performance.
The article explains how large‑scale machine learning and AutoML are applied to search advertising click‑through‑rate (CTR) prediction, covering problem definition, feature generation, model training, optimization methods, distributed systems, and recent advances in AutoML with practical case studies.
Rocket Launching introduces a novel co‑training framework that jointly trains a lightweight network and a more powerful booster network, sharing parameters and using gradient‑blocking and hint loss to improve click‑through‑rate prediction accuracy while keeping online inference latency unchanged, validated on public datasets and Alibaba’s ad system.
Online click‑through‑rate (CTR) prediction demands millisecond‑level response times, yet deep models are too slow; this paper introduces a “Rocket Launching” framework that jointly trains a lightweight net and a powerful booster net, sharing parameters and using gradient‑blocking and hint loss to improve accuracy without increasing inference latency.
This article details the design, architecture, and key components of 360 Mobile Assistant's recommendation system, covering business scenarios, data warehouse and computing layers, feature engineering, model selection, and online deployment strategies to improve app discovery and user engagement.
CTR predictionMachine LearningRecommendation System
The article details how 58.com’s Cupid push control system leverages machine‑learning models, especially XGBoost‑based CTR prediction, to prioritize and filter billions of daily push notifications, improving click‑through rates, reducing user annoyance, and providing a scalable, data‑driven architecture for diverse business services.
This article introduces Wide&Deep, PNN, DeepFM, and a novel Wide‑ResNet model for Alibaba’s 1688 “Guess You Like” recommendation, explains the underlying feature services and real‑time scoring pipeline, presents offline experiments showing AUC gains with batch normalization, and shares practical tuning insights.
This article presents DeepInsight, an industrial‑grade visual analytics platform that reveals the inner workings of large‑scale deep learning CTR prediction models, demonstrating how visualization can assess generalization, feature influence, and hidden‑layer representations to improve advertising performance.
This article recounts Alibaba‑Mama researcher Jing Shi’s presentation on the evolution of deep learning for click‑through‑rate (CTR) estimation, covering the shift from handcrafted features and linear models to piecewise linear MLR, end‑to‑end neural networks, multi‑interest user modeling, and large‑scale distributed training challenges.