Big Data Consultant & Data Scientist
chaocraig@gmail.com
User-based Recommendation
User-based / Item-based Recommendation
Latent Factors
Matrix = Associations
Things are associated Like people to colors
Associations have strengths Like preferences and dislikes
Can quantify associations Alice loves navy = +4, Carol dislikes olive = -2
We don’t know all associations Many implicit zeroes
In Terms of Few Features
Can explain associations by appealing to underlying features in common (e.g. “blue-ness”)
Relatively few (one “blue-ness”, but many shades)
Losing Information is Helpful
When k (= features) is small, information is lost
Factorization is approximate (Alice appears to like blue-ish periwinkle too)
ALS Algorithm
Optimizing X, Y simultaneously is non-convex, hard
Applications
Common Data Categories
Persona
Age, Gender, Birth date, City, …
Attributes
Phone brand/model, location, time, App, browser, banner…
Behavior
Click, Conversion(Installation, Cart, Purchase, …), Activation, Payment…
大數據分析找到更多潛在客群
In-database Processing(MPP)
Code Size Comparison
Performance Comparison
Integrated Framework
Data Sharing in Spark
Example: Logistic Regression
Example: Logistic Regression
val data = spark.textFile(...).map(readPoint).cache()
Valuation of Ads
How DSP Track & Optimize Bidding
Demand-side platform and its bidding engine in RTB
Bid Landscape Forecasting
Only reference market price by base price
Imp, UU, Click, Conv.
DSP, per campaign, per targeting criteria…
the advertisers’ targeting profiles the winning bid value
Pacing
Traffic forecasting
An impression on Jeremy Lin BBS post of MiuPTT
Two product ads
A: Linsanity T-Shirt
B: Baseketball shoes
Not optimized if only bid for highest price
B bid higher than A
Inventory A is much fewer than inventory B
Bid Landscape Forecasting
Bid Landscape Forecasting -- Bid Star Tree Expansion
Bidding Price Calculation
Bidding Price Model Building
Bidding Price Calibration
Forecasting
Sampling & learn
On-line adjustment
Feedback control & Re-learn
Loss reason
Prior Probability Shift
Budget, Freq. cap, …
Competition
Model Evaluation
ROC Curve
ROC Curve
4R: Reach, Richness, Representation, Range
Data Economy
Reach: The Value Funnel
Richness
Richness: Predictive Power
華盛頓郵報整理的 98 項 Facebook 廣告指標
Richness
Data Quality Richness
Attr. vs. behavior
Data Utilization Richness
Call taxi (short vs. long route)
Download times vs. Activation days
Data Model Richness
成功案例:優化廣告成效
成功案例:掌握 4R 成效更優異!
Range
Google trend, Viral install…
4R: Reach, Richness, Representation, Range
World, Model & Theory
謝謝大家!
16. 4V Solution Directions
Challenges Directions
Volume Scalability
Velocity Reactive / Streaming
Variety All data
Veracity Knowledge
engineering
Machine intelligence
17. The Revolution of Big Data
DATA
Hypotheses
Statistical Analysis
BIG DATA
Hypotheses
Machine Learning
Data Mining
Machine-generated
Sampling, Multi-variant… All, Hyper space, …
Volume, Velocity, Variety, Veracity
Human-explainable
18. Models ! Cases
Russ Merz, An Integrated Model of Media Satisfaction and Engagement: Theory, Empirical
Assessment and Managerial Implications, Journal of Applied Marketing Theory, Nov 2011
19. Models ! Cases
Russ Merz, An Integrated Model of Media Satisfaction and Engagement: Theory, Empirical
Assessment and Managerial Implications, Journal of Applied Marketing Theory, Nov 2011
BIG DATA
Hypotheses
Machine Learning
Data Mining
Machine-generated
All, Hyper space, …
Volume, Velocity, Variety, Veracity
deductive inductive
Cases
Models
Models
Cases
30. The Revolution of Big Data
DATA
Hypotheses
Statistical Analysis
BIG DATA
Hypotheses
Machine Learning
Data Mining
Machine-generated
Sampling, Multi-variant… All, Hyper space, …
Volume, Velocity, Variety, Veracity
Human-explainable
31. Models ! Cases
Russ Merz, An Integrated Model of Media Satisfaction and Engagement: Theory, Empirical
Assessment and Managerial Implications, Journal of Applied Marketing Theory, Nov 2011
32. Models ! Cases
Russ Merz, An Integrated Model of Media Satisfaction and Engagement: Theory, Empirical
Assessment and Managerial Implications, Journal of Applied Marketing Theory, Nov 2011
BIG DATA
Hypotheses
Machine Learning
Data Mining
Machine-generated
All, Hyper space, …
Volume, Velocity, Variety, Veracity
deductive inductive
Cases
Models
Models
Cases
36. Why Big Data?
• Internet Data
– 4 V
• Analysis Inference Action
– Data-driven with no people interception
• Building model programmatically
– Hyperparameters optimization
ALtoEatBoF? PeIsoFaDBPatBoF?
71. Matrix = Associations
Rose Navy Olive
Alice 0 +4 0
Bob 0 0 +2
Carol -1 0 -2
Dave +3 0 0
• Things are associated
Like people to colors
• Associations have
strengths
Like preferences and
dislikes
• Can quantify associations
Alice loves navy = +4,
Carol dislikes olive = -2
• We don’t know all
associations
Many implicit zeroes
Source: Sean Owen(2012), Cloudera
72. In Terms of Few Features
• Can explain associations by appealing to underlying features in
common (e.g. “blue-ness”)
• Relatively few (one “blue-ness”, but many shades)
(Alice)
(Blue)
(Navy)
Source: Sean Owen(2012), Cloudera
73. Losing Information is Helpful
• When k (= features) is small, information is lost
• Factorization is approximate
(Alice appears to like blue-ish periwinkle too)
(Alice)
(Blue)
(Navy)
(Periwinkle)
Source: Sean Owen(2012), Cloudera
139. Internet as a mass media
“Half the money I spend on advertising is
wasted; the trouble is I don‘t know which half.”
-- John Wanamaker, ~ 1875a pioneer in marketing
140. Current Challenges
Find the best match between a given user in
a given context and a suitable advertisement.
-- Dr. Andrei Broder and Dr. Vanja Josifovski, Standford University
How to identify?
Limited Info.
Budget? Creative?...
Bid Price?
141. Channels play different roles
in the customer journey
Source: http://www.thinkwithgoogle.com/
148. Bid Landscape Forecasting
• Only reference market price by base price
• Imp, UU, Click, Conv.
• DSP, per campaign, per targeting criteria…
• the advertisers’ targeting profiles the
winning bid value
• Pacing
Source: Ying Cui, Ruofei Zhang, Wei Li, Jianchang Mao(2011), Bid Landscape Forecasting in Online Ad Exchange Marketplace, Yahoo! Labs
149. Traffic forecasting
• An impression on Jeremy Lin BBS post of
MiuPTT
• wo product ads
– A: Linsanity T-Shirt
– B: Baseketball shoes
• Not optimized if only bid for highest price
– B bid higher than A
– Inventory A is much fewer than inventory B
151. Source: Ying Cui, Ruofei Zhang, Wei Li, Jianchang Mao(2011), Bid Landscape Forecasting in Online Ad Exchange Marketplace, Yahoo! Labs
*
*
C1
*
*
C2
*
B1
C1
*
B1
C2
*
*
*
* *
*
* *
C1
* *
C2
*
B1
*
• Remove few-imp path for
not too sparse
• Easily to target all
Target Attribute
Bid Landscape Forecasting
-- Bid Star Tree Expansion
152. Bidding Price Calculation
Bidding Price = F (base price, CVR )• In the same campaign, the conversion value is
the same
= base price * φ • φ = CVR / avg CVR
φ = p(c|u, i) / Ej [ p(c|u, j) ]
Ej [ p(c|u, j) ] = Σj p(c|u, j) p(j) = p(c|u)
• I, j : inventory (on Web/App)
φ = p(c|u, i) / p(c|u)
φ = p(c|s, i) / p(c|s)
• All inventories in the same segment are the same,
• i could also be inventory cluster
153. Bidding Price Model Building
• Cold start
• Training feature of segment and inventory
respectively
• Do not train combined feature for preventing over
fitting on few training data
φ = p(c|s, i) / p(c|s)
• Cold start
• Training feature of segment
AUC: the area under the ROC curve(TP/FP)
Lift: target response divided by average response.
bid = BasePrice(s, a) * p(c|s, I, a) / p(c|s, a)
DSP cross-campaigns:
154. Bidding Price Calibration
• Forecasting
– Sampling learn
• On-line adjustment
– Feedback control Re-learn
• Loss reason
– Prior Probability Shift
• Budget, Freq. cap, …
– Competition
158. 4R: Reach, Richness,
Representation, Range
Reach
Richness
High
High
Low
(DAU)
(Behavioral data)
Range
High
( Affiliate of
whole context)
Representation
(Format
Content)
159. Data Economy
Traditional - Internet Economy
HighREACH
RICHNESS
High
Low
Traditional
Economy
Internet Economy
(quality)
(quantity)
160. Reach: The Value Funnel
CPM campaign:
Revenue = N/1000 ⋅CPM
CPC campaign:
Revenue = N ⋅ CTR ⋅ CPC
CPA campaign:
Revenue = N ⋅ CTR ⋅
CVR⋅ CPA
UU Reach (DAU)
ARPU = Life-time Value
166. Richness
• Data Quality Richness
– Attr. vs. behavior
• Data Utilization Richness
– Call taxi (short vs. long route)
– Download times vs. Activation days
• Data Model Richness
171. 4R
Cross-screen synergy
Big data synergy ith Cross-screen effect
+TV
0.00%
2.00%
4.00%
6.00%
8.00%
10.00%
12.00%
14.00%
16.00%
0
5000
10000
15000
20000
25000
30000
35000
40000
Thu Fri Sat Sun Mon Tue Wed Thu Fri Sat Sun
APP
App Download Rate Optimized Conversion Rate
172. Range
- Roger Martin
Rothman School of Management, Toronto
If only attach importance to quantify the business
model, it will not have the ability to find a potential
growth opportunities: The pursuit of quantifying
the biggest problem is that people ignore the
context of the behavior generated, detached from
the context of the event, and have not been
included in the model ignores variables
effectiveness.
175. High
4R: Reach, Richness,
Representation, Range
Reach
Richness
High
High
Low (DAU)
(Behavioral data)
Range
( Affiliate of
whole context)
Representation
(Format
Content)