This article details Tencent Search's exploration of large‑model‑driven ranking, covering the evolution from traditional keyword retrieval to RAG‑based AI search, the multi‑stage AI ranking architecture (L0‑L5), model training pipelines, distillation, synthetic data generation, and future research directions.
The author recounts how, by modeling user intent with a multi‑layer Gaussian‑based PODM‑MI framework and addressing a novel ‘sand‑glass’ bottleneck in RQ‑VAE semantic identifiers, JD’s search ranking achieved significant UCVR gains, annual order increases of over ten million, and a SIGIR 2024 paper acceptance.
This article recounts a JD engineer’s journey from theory to practice, detailing the development of the PODM‑MI re‑ranking framework, its three‑layer distribution‑based design, the discovery of a novel SID bottleneck, and the resulting multi‑million‑order impact validated at SIGIR 2024.
JD’s algorithm engineer recounts how his team transformed e‑commerce search by developing the PODM‑MI re‑ranking framework, uncovering a novel “hourglass” bottleneck in generative retrieval, and implementing lightweight solutions that boosted diversity, relevance, and order volume, culminating in a SIGIR publication.
JD’s algorithm engineer describes how a three‑layer PODM‑MI re‑ranking framework, combining Gaussian preference modeling, mutual‑information optimization, and utility‑matrix fusion, overcame the hourglass bottleneck in generative retrieval, dramatically improving search diversity, user experience, and generating over ten million additional orders.
This article introduces PODM‑MI, a preference‑oriented diversity model that uses mutual information and variational Gaussian representations to jointly optimize accuracy and diversity in e‑commerce search re‑ranking, and reports significant online A/B test improvements on JD.com.
The article explains the purpose of coarse ranking in industrial search pipelines, outlines key evaluation metrics, discusses sample construction and model architecture choices, and highlights trade‑offs between consistency with downstream ranking and overall system performance.
This article examines the role of reranking in modern recommendation pipelines, explains why context‑aware listwise models are needed, surveys the evolution from pointwise to generative and diversity‑aware approaches, and reviews real‑world deployments at companies such as Kuaishou, Alibaba, WeChat, iQIYI, and Meituan, highlighting key challenges, evaluation metrics, and business‑rule integrations.
The Baidu Research Institute’s 2023 Paper Sharing Session featured eight cutting‑edge papers—from semi‑supervised web‑search ranking and hierarchical reinforcement learning for autonomous intersections to spatial‑heterophily graph networks, a unified XAI benchmark, differentiable neuro‑symbolic KG reasoning, and novel stochastic‑gradient and neural‑field loss analyses—showcasing advances across AI, data mining, and computer vision.
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.
Airbnb upgraded its neural‑network ranking system by adding a similarity network that penalizes duplicate‑like listings, enabling the algorithm to present a more diverse set of options, which boosted booking rates, value, and five‑star ratings, demonstrating that reduced result similarity improves overall search quality.
This article details the multi‑stage evolution of JD's search Explore‑Exploit (EE) system—covering an adaptive dynamic detection model, scenario‑modeling upgrades, end‑to‑end scoring and insertion consistency, and context‑aware brand/store dimension detection—demonstrating how each iteration improves result diversity, user experience, and key online metrics while maintaining search efficiency.
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 presents a comprehensive overview of industry search, covering its background, core retrieval and ranking technologies—including sparse and dense retrieval, pre‑trained language models, tokenization, NER, adaptive multi‑task training, and re‑ranking models—followed by detailed case studies such as address analysis, family‑ID unification, emergency call handling, education photo‑search, and power‑knowledge‑base integration.
The paper presents a two‑stage e‑commerce search framework where the coarse‑ranking stage is redesigned with multi‑objective optimization, expanded negative sampling, and listwise distillation—guided by a new global transaction hitrate metric—enabling it to surpass fine‑ranking on large candidate sets and boost overall GMV by about one percent.
This report details the multi‑stage architecture of Taobao’s main search, introduces a new global‑transaction hitrate metric, analyzes offline and online evaluation gaps, and presents a series of model, loss‑function, and sampling improvements that together lifted overall conversion by about one percent.
This article explains the architecture of video search engine ranking, covering optimization objectives such as relevance, click‑through rate and watch time, and detailing pointwise, pairwise and listwise learning approaches, model training pipelines, and online serving strategies.
This article describes Meituan Search's pre‑ranking (coarse‑ranking) system evolution and presents two major optimization strategies—leveraging knowledge distillation to align coarse‑ranking with fine‑ranking and employing neural architecture search to jointly improve effectiveness and latency—demonstrating significant offline and online performance gains.
Meituan’s search team upgraded its pre‑ranking layer from simple linear models to end‑to‑end neural networks, boosting effectiveness by applying three knowledge‑distillation techniques—including result‑list, score, and contrastive representation transfer—and by using latency‑aware neural architecture search to automatically select features and network structures, achieving significant recall and CTR gains without added latency.
The talk presents JD Search's Explore‑and‑Exploit (EE) module, detailing its bias‑mitigation pipeline—including position, popularity, and exposure debiasing—model architecture upgrades with SVGP and causal inference, online AB metrics, offline evaluation methods, and future research directions to improve search diversity and long‑term value.
This article details Meituan’s implementation of on‑device deep learning models for search re‑ranking, covering the motivations for edge intelligence, feature engineering, feedback sequence modeling, model architecture, deployment optimizations, experimental results, and future directions, offering practical insights for developers building large‑scale AI on mobile.
This presentation details how JD applied reinforcement learning—using RNN‑based user state modeling and a DDPG framework—to improve e‑commerce search ranking by optimizing long‑term cumulative value, showing significant offline and online gains in conversion and GMV.
This article presents the DIRN model, which leverages product inclusion graphs and graph‑based embeddings to address the challenges of ranking both single‑item and complex travel products on Fliggy, demonstrating significant CTR, CVR, and GMV improvements through offline experiments and online A/B testing.
This article presents Meituan's search ranking system, detailing its multi‑business sorting architecture, layered ranking pipeline, quota and fine‑ranking models, aggregation modeling techniques, and supporting platforms such as Lego and Poker, while also sharing practical insights and recruitment information.
Meituan Search tackles the difficulty of ranking items from diverse business lines by introducing a five‑tower mixed architecture, group‑lasso and feature‑gate selection, a probabilistic graph model, and a joint block‑order/size predictor, achieving notable offline NDCG gains and online CTR and purchase‑rate improvements.
This article presents a multi‑category Mixture‑of‑Experts (MoE) approach for e‑commerce search ranking, addressing category‑specific behavior and small‑category learning by introducing hierarchical soft constraints and adversarial regularization, and demonstrates significant AUC and NDCG gains on Amazon and JD in‑house datasets.
At iQIYI’s iTech Conference, Zhang Zhigang detailed a full‑stack search ranking system that combines NLP‑driven query analysis, hierarchical indexing, multi‑stage coarse‑to‑fine ranking, Transformer‑based re‑ranking, sparse‑feature DNN enhancements and LIME/SE‑Block explainability, delivering measurable gains in CTR and NDCG for the platform’s video search.
Meituan Search tackles the multi‑business ranking challenge by introducing a quota‑allocation model (MQM) and a series of precise ranking models (MBN) that progressively incorporate sub‑networks, multi‑task learning and transformer‑based behavior sequences, delivering consistent CTR and purchase‑rate gains across food, hotel, travel and other services while outlining future work on feature utilization, sample‑imbalance mitigation and multi‑objective optimization.
This article presents the challenges of travel‑related product search, explains why traditional rule‑based sorting is insufficient, and describes how Alibaba Flypig’s team built a deep‑learning based personalized ranking system—including architecture, model variants, experimental results, and future optimization directions—to improve conversion rates for flight and ticket searches.
The article describes JD's end‑to‑end online learning pipeline for retail search, covering the background, system architecture, real‑time feature collection, sample stitching, Flink‑based incremental training, parameter updates, and full‑link monitoring to achieve low‑latency, high‑accuracy model serving.
This article describes how 58.com’s rental platform leverages deep learning models such as Wide&Deep, DeepFM, DCN, DIN, and DIEN to improve search ranking, detailing data pipelines, feature engineering, model iteration, multi‑task training, prediction optimizations, and resulting online performance gains.
This article reviews the development of Zhihu's search system, describing the transition from early GBDT ranking to deep neural networks, the introduction of multi‑objective and position‑bias‑aware learning‑to‑rank methods, context‑aware techniques, end‑to‑end training, personalization, and future research directions.
This article presents the end‑to‑end design, feature engineering, model evolution (Wide&Deep, DeepFM, DCN, DIN, DIEN), multi‑task training, and deployment optimizations that 58.com applied to improve search ranking for its rental business, demonstrating significant gains in click‑through and conversion rates.
Elasticsearch calculates a relevance score for each document, but using the function_score query you can customize this scoring by combining the original query_score with a user-defined func_score through various functions such as weight, random_score, field_value_factor, decay_function, and script_score, allowing flexible ranking based on business needs.
This article explains how Yanxuan built a comprehensive vector system—from product embeddings and graph models to large‑scale similarity computation—and applied it across search, recommendation, and purchase prediction tasks, highlighting practical algorithms, infrastructure, and future directions.
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.
The article explains how to systematically improve Elasticsearch search relevance by fine‑tuning Query DSL with filters, phrase matching, and boosts, incorporating static scoring via function_score, adjusting BM25 similarity parameters, and using diagnostics like _explain to iteratively achieve higher ranking quality.
The Meituan Search team boosted ranking relevance by training a domain‑specific BERT, applying data augmentation, brand‑sample optimization, knowledge‑graph fusion, multi‑task and pairwise fine‑tuning, joint end‑to‑end training with LambdaLoss ranking models, and compressing the model for low‑latency inference, delivering up to +925 BP offline accuracy gains and measurable CTR and NDCG improvements in production.
Alibaba’s Youku video search platform combines six-layer architecture—data extraction, technology integration, recall, relevance, ranking, and intent understanding—leveraging CV, NLP, knowledge graphs, and multi‑modal cues such as face, OCR, and audio recognition to overcome title‑mismatch, entity, and semantic challenges and deliver precise, diverse video retrieval.
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 how Ctrip's search team leverages XGBoost for learning-to-rank, covering L2R concepts, feature engineering, data preparation, model training, hyper‑parameter tuning, offline evaluation, and deployment insights for large‑scale search ranking systems.
This article reviews the Airbnb search ranking paper, detailing offline and online performance gains, the progression from SimpleNN to LambdaRankNN, GBDT/FM NN, and Deep NN models, failed embedding attempts, extensive feature engineering practices, and the production system architecture that enabled large‑scale deep learning deployment.
The article summarizes the 58 Group technical salon where experts presented the microservice‑based recommendation system architecture, data and strategy layers, and the internally built search ranking platform covering sampling, feature engineering, and model training, highlighting practical implementations and lessons learned.
This article presents a comprehensive overview of 58.com’s search efficiency optimization, detailing the business background, ranking framework, data, algorithm, and engineering components, describing the three-stage ranking process, strategy and platform optimizations, feature engineering, model upgrades, and the resulting performance improvements.
The Meituan‑Dianping search team progressed its core ranking from linear, FM and GBDT models to a knowledge‑graph‑enhanced deep‑learning architecture, culminating in the listwise LambdaDNN network that directly optimizes NDCG, supported by extensive feature engineering, distributed TensorFlow training, and the Athena diagnostic system.
This article details iQIYI's practical experience in building and iterating its video search system, covering basic relevance, semantic matching via translation and click models, deep‑learning approaches, and ranking model evolution from heuristic rules to learning‑to‑rank, highlighting challenges, solutions, and performance gains.
This article details how 58.com improved the efficiency of its search list ranking by moving from simple time‑based ordering to a comprehensive ranking framework that incorporates feedback strategies, basic machine‑learning models, feature upgrades, and advanced model upgrades, achieving significant gains in click‑through, conversion, and revenue across multiple business lines.
Xianyu achieves second‑level real‑time search ranking adjustments by using Alibaba Cloud HybridDB for PostgreSQL to normalize heterogeneous data into JSONB, merge attributes with timestamp‑based logic, and trigger PostgreSQL NOTIFY events that instantly recalculate scores, boosting transaction volume by ~30% and feedback by ~28%.
The JD.com Rental platform, launched in early 2018, combines e‑commerce and service models to offer a full‑cycle online rental experience across eight cities, leveraging custom front‑end channels, cloud‑based infrastructure, big‑data ranking algorithms, and the "Zhouzi" integration to sync millions of listings and streamline merchant operations.
On August 25, iQIYI’s Chengdu R&D Center hosted its second Technical Salon, featuring talks on AI-driven content understanding for short‑video feeds, speech synthesis and editing, industry‑standard speech recognition, semantic search ranking, anti‑spam UGC text analysis, and concluding with recruitment invites and a preview of the upcoming Shanghai salon.
Learning to rank reframes search as a machine‑learning problem that optimizes document ordering rather than numeric prediction, using relevance metrics such as NDCG and feature‑based scoring functions, and comparing point‑wise, pair‑wise (RankSVM) and list‑wise (ListNet) approaches while stressing that proper error definition and feature selection matter more than the specific algorithm.
This article examines the challenges of high‑dimensional sparse features in search ranking, explains why traditional linear models struggle, and describes how deep neural networks with novel encoding schemes and online updates can dramatically improve CTR prediction and real‑time performance.
This article explores how Taobao leverages reinforcement learning, multi‑armed bandits, and reward‑shaping techniques to improve large‑scale e‑commerce search ranking and recommendation, detailing problem modeling, algorithm designs such as Tabular Q‑learning and DDPG, experimental results from Double‑11, and advanced models like GBDT+FTRL and Wide‑&‑Deep.
Alibaba leveraged reinforcement learning, highlighted by MIT Technology Review’s 2017 breakthrough list, to transform its e‑commerce search and recommendation systems during Double 11, deploying large‑scale online and batch training pipelines, dynamic market segmentation, and real‑time decision models that boosted click‑through rates by up to 20 %.
This article explores how Alibaba's Taobao leverages reinforcement learning, Markov decision processes, and reward shaping to improve large‑scale product search ranking and recommendation, detailing problem modeling, algorithm designs such as Tabular Q‑learning and DDPG, experimental results, and advanced recommendation models like GBDT‑FTRL and Wide‑Deep.
This article explains Alibaba’s dual‑link real‑time computing framework, detailing its micro‑ and macro‑level pipelines, key components such as Pora, iGraph and SP, online learning architectures, pointwise and pairwise ranking models, bandit‑based strategy optimization, PID‑controlled traffic balancing, and the impressive performance gains achieved during the Double 11 shopping festival.
The article presents Alibaba's real‑time computing platform for search, detailing its micro‑ and macro‑level architectures, online learning frameworks, point‑wise and pair‑wise ranking models, bandit‑based strategy optimization, and PID‑controlled traffic regulation, and reports significant performance gains during the Double‑11 shopping festival.
This article describes Meituan's online search ranking architecture for O2O services, covering data pipelines, feature loading, ranking service workflow, A/B testing, model choices, cold‑start handling, and position bias mitigation, all tailored for mobile‑centric personalized ranking.
