invited talk on recommendation and advertising in alibaba, recsys'13
DESCRIPTION
Invited talk on Recommendation and Advertising in Alibaba, RecSys'13TRANSCRIPT
1
Recommendation &
Advertising in e-commerce
Quan Yuan
Agenda
• Recommendation – Item-based
– User-based
• Recommendation in Ads Targeting – Look-alike Targeting
– Ctx-weather Targeting
– User Profiling • FM for Purchasing power prediction
• Life-stage modeling
Alibaba in China
1st B2C
1st C2C
1st Group-buy site
1st Price-comparison
1st Shopping App
$30B usd 11.11,2012
100+ online scenarios for Recommendation
10% of GMV (daily trading volume) comes from Recommender System Alibaba is a platform, to help sellers’ selling.
Item-based Rec
• Item-Item Relevance Model – 10+ relevance model based on: click, buy, search, bookmark, content, etc.
– Different granularity: session/day/week/month/year
– Offline computation, PB size, 19min
• Similar VS. Complementary – P( buy | user, item, ctx)
– Similar only is good for
CTR but harm for CVR
机密文档
Similar Reco
Complementary Reco
User-based Rec
• “Products you may like”
• Brand Rec @Tmall
40+%CVR; 50%+GMV Lift VS. Popular
MostPopular : Red BrandRec: Blue
Rec is good for long-tail brands
Agenda
• Recommendation – Item-based
– User-based
• Recommendation in Ads Targeting – Look-alike Targeting
– Ctx-weather Targeting
– User Profiling • FM 4 Purchasing power prediction
• Life-stage modeling
Ads Platform
1st B2C
1st C2C
1st Group-buy site
1st Price-comparison
1st Shopping App
$30B usd 11.11,2012
1st Ads Platform
Taobao Ads Network & Exchange (TANX)
External Publisher Video/News/Apps...
Display Ads Targeting
Coverage
Performance
Behavioral Targeting
Contextual Targeting
Re-targeting
Demographic Targeting
Look-Alike Targeting
After the targeting(matching), we adopt the ranking to optimize CTR/CVR
Rec for Ads
Coverage
Performance
Behavioral Targeting
Contextual Targeting
Re-targeting
Demographic Targeting
Look-Alike Targeting
ItemCF
ItemCF
UserCF
ContentRec
WeatherRec
FM for Purchase Power Modeling
Life-Stage Modeling
Look-alike Targeting
Contextual Engine
Weather-based Ads
PM2.5
Temp.
Climate
Wind
Ads Trigger
CTR & click value both improved than popular! In mobile ads, it is hard to track users behavior
User-Cat Consumption Matrix
I 1 I2 I3 I4
U1 ? 0 21 1
U2 9 ? 1 0
U3 0 ? 6 11
U4 19 0 2 ?
User Consumption Matrix
• User-Category Amount Matrix
Sample: Purchasing Prediction Model
Global Bias
0.9574
User Bias
0.9132
Cat Bias
0.8868 Cross-Cat
0.8729
Indicator-Cat
0.8681 100M x 100K Matrix Factorization on a single server
Life-stage Modeling
• Using multiple behavioral data (buy, view, search, etc.) and split the data as sequence
Current
The baby will be
one year old at
6-12 month 1-6 month 1-6 month pregnancy
Life-stage Modeling cont.
• Given an observed behavioral sequence x=(x1...xl), our task is to predicts a life stage sequence y=(y1...yl)
Sequence of behavioral
Sequence of Life-stage
y1 y2 y3
x1 x2 x3
y4 y5 y6
x4 x5 x6
16
Thanks &
We are Hiring~! “A-Star” for Top Talents http://campus.alibaba.com/
Sina weibo: 袁全V