腾讯高级研究员，专注于推荐系统与计算广告领域。在内容推荐（信息流、视频、动漫）和广告推荐（电商、游戏行业）上均有丰富的实战经验，覆盖召回、粗排、精排、特征建模全链路。研究兴趣：机器学习、深度学习、推荐算法、NLP、大语言模型（LLM）。2022年入选深圳创新中心专家人才库。2026年3月起转入腾讯基础研究中心，聚焦生成式模型与 Scaling Up 方向的研究。
I am a Senior Researcher at Tencent, working on recommendation systems and computational advertising. I have extensive hands-on experience in both content recommendation (feeds, video, comics) and ad recommendation (e-commerce, gaming), covering the full pipeline of retrieval, pre-ranking, ranking, and feature engineering. Research interests: Machine Learning, Deep Learning, Recommendation Algorithms, NLP, LLM. I was selected for the Shenzhen Innovation Center Expert Talent Pool in 2022. Since March 2026, I joined Tencent's Foundation Research Center, focusing on generative models and scaling up.
硕士毕业于国防科学技术大学（推荐算法方向），本科毕业于北京邮电大学。曾在华为担任高级算法工程师，负责大数据营销平台的开发。
I received my M.S. from National University of Defense Technology (NUDT) in Recommendation Algorithms, and B.S. from Beijing University of Posts and Telecommunications (BUPT). Before Tencent, I worked at Huawei as a Senior Algorithm Engineer on big data marketing platforms.
教育背景Education
国防科学技术大学 — 硕士National University of Defense Technology (NUDT) — M.S.2013 – 2016
信息系统与管理学院，研究方向：推荐算法 / 大数据 / 机器学习School of Information Systems and Management. Research: Recommendation Algorithms / Big Data / ML
北京邮电大学 — 本科Beijing University of Posts and Telecommunications (BUPT) — B.S.2009 – 2013
管理学院，电子商务专业School of Management, E-Commerce
News
Mar 2026转入腾讯基础研究中心，聚焦生成式模型与 Scaling Up 方向Joined Tencent Foundation Research Center, focusing on generative models and scaling up
Mar 2026新专利"用户行为序列分析和广告质量评估的多Agent构建"已提交New patent filed: "Multi-Agent Construction for User Behavior Sequence Analysis and Ad Quality Evaluation"
Jan 2026新专利"生成式推荐中的模型框架设计"已提交New patent filed: "Model Framework Design for Generative Recommendation"
2025两项新专利已提交（LLM重排、多模态+ID兴趣建模）Two new patents filed (LLM re-ranking, multimodal + ID interest modeling)
"对推荐，包括广告、信息流、电商等行业都有大量的经验，思维敏捷，快速给出基于过往经验的想法和解法，获益良多。""Extensive experience across recommendation, ads, feeds, and e-commerce. Quick-thinking, rapidly provides ideas and solutions. Greatly beneficial."
"在广告模型特征算法等方面经验非常丰富，对AI的理解也非常深刻，收获很大。""Very rich experience in ad model features and algorithms, with deep understanding of AI. Extremely rewarding."
2022入选深圳创新中心专家人才库Selected for Shenzhen Innovation Center Expert Talent Pool
个人项目Projects
为什么 Hadamard 积的计算效率比矩阵乘法低Why Hadamard Product Has Lower Compute Efficiency Than MatMul on GPUs技术分析Technical Analysis
从计算密度（Arithmetic Intensity）角度解释 Hadamard 积在 GPU 上效率低于矩阵乘法的反直觉现象。深入分析数据复用、Tensor Core 利用率、内存访问模式差异，以及推荐系统特征交叉中 DCN-V2/RankMixer 选择矩阵乘的工程原因。Explains the counter-intuitive phenomenon of Hadamard product being less efficient than MatMul on GPUs from the perspective of Arithmetic Intensity. Analyzes data reuse, Tensor Core utilization, memory access patterns, and why DCN-V2/RankMixer choose matrix multiplication for feature crossing in RecSys.
LLM/VLM 样本组织 vs 推荐系统：从 Megatron 到特征交叉LLM/VLM Data Organization vs RecSys: From Megatron to Feature Crossing技术分析Technical Analysis
从 Megatron-LM 数据组织出发，对比 LLM 预训练与推荐系统的样本组织差异，分析可迁移技术（mmap索引、Packing、多源混合、Token化）及个性化设计需求。深入探讨 Token 化序列模型中 Attention 与 FM/DCN 的特征交叉能力差异：乘法信号路径、数据效率、以及"双路并行"混合架构的工程方案。Starting from Megatron-LM's data pipeline, comparing sample organization between LLM pre-training and RecSys. Analyzes transferable techniques (mmap indexing, packing, blending, tokenization) and custom requirements. Deep dive into Attention vs FM/DCN feature crossing: multiplicative signal paths, data efficiency, and the dual-path hybrid architecture.
探索将大语言模型融入推荐系统全链路：利用LLM进行用户兴趣理解、候选集重排序和推荐理由生成，结合传统召回/精排模型构建混合推荐架构。Exploring LLM integration across the full recommendation pipeline: leveraging LLMs for user interest understanding, candidate re-ranking, and recommendation explanation generation.
把字节 2025–2026 年的四个工业 Ranking 架构放在同一张坐标系里对比：各自攻击 DLRM 两段式的哪一段、结构图并排、参数共享 / 长度瓶颈 / KV Cache / Scaling / 线上收益一张表对齐，最后给出选型决策树和下一代架构趋势判断。A side-by-side comparison of four ByteDance industrial Ranking architectures (2025-2026): which part of the DLRM two-stage paradigm each one attacks, architecture diagrams aligned, parameter sharing / length bottleneck / KV cache / scaling / online gains in one table, plus a selection decision tree and next-generation trajectory analysis.
RankMixer 的 Token 数 T 与 Scaling Law 的斜率 / 截距RankMixer's T and the Slope/Intercept of Scaling Law论文精读Paper Deep-Dive
围绕字节四篇工业 Ranking 架构（RankMixer / LONGER / OneTrans / MixFormer）的深度解读：Token 数 T 的五条约束、Td² 参数量辨析、scaling law 斜率 vs 截距的工业含义。附带 HTML 专题页（含 LaTeX + 论文原图）。A deep dive into four ByteDance industrial Ranking architectures (RankMixer / LONGER / OneTrans / MixFormer): the five constraints on T, parameter-count analysis of Td², and industrial meaning of scaling-law slope vs intercept. Includes an HTML companion page with LaTeX and paper figures.
RankMixerLONGEROneTransMixFormerScaling LawScaling Law
基于Capsule Network和Transformer的用户多兴趣抽取方法，结合行为序列时序建模，捕获用户动态兴趣演化。Multi-interest extraction via Capsule Networks and Transformers, combined with temporal sequential modeling to capture dynamic user interest evolution.
Transformer多兴趣建模Multi-Interest序列推荐Sequential Rec
广告电商行业模型E-commerce Ad Industry Model已上线Shipped
从0搭建行业模型，使用MoE/MMoE/CGC多任务融合建模，引入STAR/DIN-attention/FwFFM等模型优化。上线效果：电商行业消耗+3.48%，GMV+5.93%，带动大盘GMV+1.48%。Built industry-specific models from scratch using MoE/MMoE/CGC multi-task learning, with STAR/DIN-attention/FwFFM optimizations. Results: +3.48% ad spend, +5.93% GMV, +1.48% overall platform GMV.
TencentMoE/MMoE/CGCSTARDIN-attentionFwFFM
信息流推荐系统Feed Recommendation System已上线Shipped
负责波洞App（从0到1）、手Q动漫、企鹅电竞视频Feeds的推荐算法，从LR到NN模型演进，经历冷启动、策略迭代等各阶段。对比产品部算法中台：波洞视频观看时长+68%，点击率+41%。Led recommendation systems for Boodo App (from scratch), QQ Comics, Penguin Esports video feeds. Evolved from LR to deep NN models. vs. product algorithm team: +68% watch time, +41% CTR.
TencentDeep NN冷启动Cold StartFeeds
UDM大数据营销平台UDM Big Data Marketing Platform已上线Shipped
负责Lookalike用户挖掘模块和实时营销时机事件引擎开发，基于Hadoop/Spark/Kafka构建实时敏捷数字化营销解决方案。Developed Lookalike user mining module and real-time marketing event engine using Hadoop/Spark/Kafka for agile digital marketing solutions.
HuaweiHadoopSparkKafkaLookalike
💡 深耕推荐系统8年+，持续关注LLM与推荐的深度融合，以及多目标建模、因果推断在广告推荐中的前沿应用。💡 8+ years in RecSys, continuously exploring deep integration of LLMs with recommendation, and frontier applications of multi-objective modeling and causal inference in ad recommendation.
多模态广告内容理解与匹配Multimodal Ad Content Understanding & Matching进行中In Progress
融合广告素材的图像、视频、文本等多模态信息，通过CLIP/BLIP类模型提取语义特征，构建跨模态用户-广告匹配模型。Fusing image, video, and text from ad creatives, extracting semantic features via CLIP/BLIP-family models, building cross-modal user-ad matching models.
CLIPBLIP跨模态匹配Cross-Modal
多模态+ID融合的用户生成式兴趣建模Multimodal + ID Fusion for Generative User Interest Modeling进行中In Progress
将多模态内容表征与传统ID Embedding进行深度融合，通过生成式建模捕获用户多维度兴趣，解决冷启动和跨域推荐问题。已申请相关专利。Deep fusion of multimodal content representations with traditional ID embeddings, capturing multi-dimensional user interests via generative modeling. Related patent filed.
构建DSSM及自监督Embedding特征，泛娱乐特征及电商embedding特征应用于召回/粗排/精排全链路。上线效果：电商行业消耗+0.88%，GMV+1.38%。Built DSSM and self-supervised embedding features applied across retrieval/pre-ranking/ranking pipeline. Results: +0.88% ad spend, +1.38% GMV.
覆盖A股/港股/美股的技术分析平台，支持K线形态识别、MA20偏离度/MACD/RSI等多指标分析、日线与周线双周期信号生成，以及基本面数据和新闻聚合展示。Technical analysis platform covering A-shares, HK, and US stocks. Features candlestick pattern recognition, MA20/MACD/RSI multi-indicator analysis, dual-timeframe signal generation, and fundamental data with news aggregation.
基于强化学习的A股量化交易策略RL-Based Quantitative Trading Strategy for A-Shares进行中In Progress
利用PPO/DQN强化学习算法，结合技术指标和市场情绪因子，构建自动化交易策略。Building automated trading strategies using PPO/DQN reinforcement learning with technical indicators and market sentiment factors.
搭建基于ROS2的家用机器人原型，集成SLAM导航、语音交互和视觉识别能力。Building a home robot prototype based on ROS2, integrating SLAM navigation, voice interaction, and visual recognition.
探索用大语言模型作为机器人高层任务规划器，通过自然语言指令分解复杂任务。Exploring LLMs as high-level task planners for robots, decomposing complex tasks from natural language instructions.
2022入选深圳创新中心专家人才库Selected for Shenzhen Innovation Center Expert Talent Pool
2018–2024腾讯连续多年四/五星员工Tencent 4/5-Star Employee for multiple consecutive years
2020《特征挖掘工程》获 TEG 2020H2 SEVP业务突破奖Feature Mining Engineering — TEG 2020H2 SEVP Business Breakthrough Award
2019《开放推荐引擎》获 TEG 2019H2 SEVP业务突破奖Open Recommendation Engine — TEG 2019H2 SEVP Business Breakthrough Award
2016华为明日之星、两次部门软件之星Huawei Star of Tomorrow, twice Department Software Star
2015Twitter数据分析平台获军事创新二等奖Twitter Analytics Platform — Military Innovation 2nd Prize
2015
Kaggle — San Francisco Crime Classification, Top 10% (2335 teams)
2012全国大学生"创意、创新、创业"电子商务比赛北京赛区三等奖National E-Commerce Innovation Competition, Beijing Region 3rd Prize
简历Curriculum Vitae
基本信息General Information
姓名Name郭亮 (Liang Guo)Liang Guo (郭亮)
语言Languages中文（母语）、英语（流利）Mandarin (native), English (fluent)
研究方向Research推荐系统、计算广告、深度学习、NLP、LLMRecommendation Systems, Computational Advertising, Deep Learning, NLP, LLM
工作经历Experience
2018 – Present
高级研究员Senior Researcher
腾讯，深圳Tencent, Shenzhen
广告电商行业模型（2020–2022，负责人）：从0搭建行业模型，使用MoE/MMoE/CGC多任务融合建模，引入STAR/DIN-attention/FwFFM等模型优化。上线效果：电商行业消耗+3.48%，GMV+5.93%，带动大盘GMV+1.48%。E-commerce Ad Industry Model (2020–2022, Lead): Built industry-specific models from scratch using MoE/MMoE/CGC multi-task learning, with STAR/DIN-attention/FwFFM optimizations. Results: +3.48% ad spend, +5.93% GMV, +1.48% overall platform GMV.
电商特征图谱构建（2019–2020，负责人）：构建DSSM及自监督Embedding特征，泛娱乐特征及电商embedding特征应用于召回/粗排/精排全链路。上线效果：电商行业消耗+0.88%，GMV+1.38%。E-commerce Feature Graph (2019–2020, Lead): Built DSSM and self-supervised embedding features applied across retrieval/pre-ranking/ranking pipeline. Results: +0.88% ad spend, +1.38% GMV.
信息流推荐（2018–2020，负责人）：负责波洞App（从0到1）、手Q动漫、企鹅电竞视频Feeds的推荐算法，从LR到NN模型演进，经历冷启动、策略迭代等各阶段。对比产品部算法中台：波洞视频观看时长+68%，点击率+41%。Feed Recommendation (2018–2020, Lead): Led recommendation systems for Boodo App (from scratch), QQ Comics, Penguin Esports video feeds. Evolved from LR to deep NN models. vs. product algorithm team: +68% watch time, +41% CTR.
内部行家咨询：在腾讯"行家"平台提供专题咨询，累计46次咨询，评分98.55/100，2023年度优秀行家。Internal Expert Consulting: Provided consulting on Tencent's Expert platform. 46 consultations, rated 98.55/100, named 2023 Annual Outstanding Expert.
2016 – 2018
高级算法工程师Senior Algorithm Engineer
华为，深圳Huawei, Shenzhen
UDM大数据营销平台：负责Lookalike用户挖掘模块和实时营销时机事件引擎开发，基于Hadoop/Spark/Kafka构建实时敏捷数字化营销解决方案。UDM Big Data Marketing Platform: Developed Lookalike user mining module and real-time marketing event engine using Hadoop/Spark/Kafka for agile digital marketing solutions.
教育背景Education
2013 – 2016
硕士M.S.
国防科学技术大学，信息系统与管理学院National University of Defense Technology (NUDT)
研究方向：推荐算法 / 大数据 / 机器学习Research: Recommendation Algorithms / Big Data / Machine Learning
2009 – 2013
本科B.S.
北京邮电大学，管理学院，电子商务专业Beijing University of Posts and Telecommunications (BUPT), E-Commerce
