X How Artificial Intelligence is Changing Advertising in China Bessie Lee Ching Law Founder & CEO - Withinlink General Manager - Tencent Social Ads
X About Tencent
X Tencent Has Developed into A Leading Fortune Global 500 2015.01 Companies 2017 Provider of Internet Services in China Tencent included "Internet+" concept proposed by Pony Ma was included in the National Strategy Report 2017.05 2013.08 Market Mobile QQ and capitalization Weixin launched over US$ 300 game centers billion 2013.09 2011.01 Market Launched Weixin capitalization over US$ 100 2007.06 2011.06 billion Established Launched open Tencent platform strategy Charity Fund 2009.12 2004.06 Annual revenue Listed on Hong over RMB 10 Kong Stock billion 2005.05 Exchange 2003.08 Launched social Entered online networking service Qzone games market 1998.11 2003.12 Launched web Tencent was portal – founded www.QQ.com 1999.02 Launched instant messaging service- QQ
X 6 Business Groups Provide A Range of Internet Products CSIG IEG CDG TEG WXG PCG Cloud & Smart Interactive Technical Engineering Corporate Development W eixin Group Platform & Content Group Industries Group Entertainment Group Group Group Tencent Big Data Tencent Cloud QQ Browser Wechat Tencent Games Advertising & Tencent Data Center YingYongBao Marketing Service Qzone Tencent Auto lntelligence China Literature Security WeChat Work PC Manager Financial Tech QQ Map Tencent Video QQ Mail Tencent Pictures Mobile Manager Tencent Internet Plus Tencent News Tencent eSports Tencent W iFi Manager Tencent Open Platform Tencent animation&comics
X “Connection” Strategy Connecting People with Connecting People with Connecting People with People Services Devices
X AI @ Tencent
X AI In All Reinforcement Decision Computer Learning Vision Making Creation Generation Machine Learning Natural Voice Language Comprehension Recognition Cognition Processing
X Tencent Al Labs Wechat AI Tencent Youtu AI.QQ.COM AI Lab
X CASE 1 Smart Retail Face Payment Face recognition payment solution that empowers the retail industry Precise identification Extremely efficient Safe and reliable 99% 1s 3D Live detection 应用场景 Self-service cash Small screen cash Mobile self-service cash register register register
X CASE 2 Medical Imaging Lung cancer monitoring Locate suspicious nodules to improve the discovery rate and accuracy of early lung cancer It has been applied to more than 60 hospitals Pulmonary nodule detection Lung cancer diagnosis AI electronic diagnosis report
X AI in Tencent Ads
X Ads Serving – Machine Gun for a Moving Target AD2 AD1 • Tens of billions of requests AD3 • Hundreds of thousands of ads • Real-time matching of the most relevant User profile ads to users • Supported by machine learning models trained on extremely large historical data Best Ad Ad request
X Intelligent Profile Understanding & Accurate User Insights Demographics Interest • inherit • purchasing • location intent • education • hobby • marriage • wealth • employment Behavior Device • travel • carrier UID • network • App • eCommerce • brand • type • O2O • Ad engagement • OS • purchase • price Custom Vertical • travel • seed • auto • 3C • expansion • education • finance • 1-party labels • eCommerce • FMCG • game
X Traditional User Definition - Manual Targeting Gender, Interests, age, geo, False keywords education. Cons Gender: F Age • Prior knowledge Followers, 30~60 Marriage, APP installs, • True Information loss employment, payments, Exclude: wealth, culture engagements previous • Coarse graininess buyers • Manual optimization Interest: APP, designer Context, False eCommerce, purse device, Search, and weather Social behaviors Large number of Advertisers manually specify the Boolean match during targetable attributes targeting criteria ads retrieval
X Auto Targeting: Automatic User Identification by AI 0.56 Existing leads from 1st-party data x 1 A change of mindset x 2 0.81 Previously • Targeting by user-ad relevance Converted users x 3 +1 User/Ad features +1 0.79 Multiple ways to define Targeting by machine Matching based on seed user learning models relevance
X Auto-Targeting : Representation Learning & ANN Representation learning Approximate Nearest Neighbor Search Learns embeddings for users and ads; Efficient ad retrieval based on user/ad can be used during ad retrieval embeddings under scalability constraints User embedding Ad embedding User Context Channel Ad features features features features ” Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs ” , 2016
X Lookalike Expansion for Targeting: Identify Potential New Customers Custom audience based on known converters Lookalike 1st-party data Tencent data MORE User Profiles Understand the commonality of the seed Find lookalike Mined user audience audience features 2015 DNN - Deep embedding for each seed user, 2014 expansion done online Logistic Regression - One model per seed audience, 2013 expansion done offline Peacock - Large scale distributed LDA
X Static Creative is Hard to Optimize for Advertisers with Large Inventory User Static Creative : Impression Challenges Creative Hit-or-miss Coarse-grained targeting Serving Static content with no personalization Creation Advertiser
X DPA: Automatic Creation, Retrieval, and Rendering of Personalized Ad Creatives Advantages User Automatic creative generation Impression Online creative • No need for trial-and-error generation optimization from advertisers Creative Fine-grained targeting Product retrieval Serving • 100m targetable products • More detailed and accurate than ad targeting Product targeting Truly personalized Creation Product DB • Different experience even for the same ad Advertiser
X AI Innovations by Chinese Startups
X Collaboration AI Backend
X Collaboration Marketing Solution
X Collaboration Offline Data Platform
X Collaboration KOL Management
X Thank you
X Contact Us Bessie Lee Ching Law Bessie.lee@withinlink.com chinglaw@tencent.com
Recommend
More recommend