2019 DevDay
How LINE Ads Platform Is
Constantly Evolving
> Jinsoo Oh
> LINE Plus AD Performance Software Engineer
Slide 2
Slide 2 text
Two and a Half Years Ago…
Slide 3
Slide 3 text
The Direction of Evolution
Become an Advertiser-Friendly Platform
Support Smart Features Using Machine Learning & Algorithmic Techniques
Comply with Industry Standards
Slide 4
Slide 4 text
Takeaways
Gather Professionals in the Ad Industry!
Don’t Focus on the Beauty of your Logic! Then What?
Make your System the Industry Standard!
Take a Look at the Features that are Related to Machine Learning & Logic
Online
Account
2019.11
JP Migration
Completed
2019.03
JP Migration
Started
2018. 08
TH OPEN
2017.10
Two and a Half Years Ago…
Slide 8
Slide 8 text
Two and a Half Years Ago…
Conversion
Events Added
CTR/CVR Conversion
Auto Bidding
Click
Auto Bidding
Slide 9
Slide 9 text
Understanding Some Terms
Conversion Event
App Install
Purchase
CVR
Conversion Rate
=
# of conversions / # of
clicks
Bidding Amount
1,000 yen per app install
2,000 yen per purchase
Slide 10
Slide 10 text
Two and a Half Years Ago…
Conversion
Events Added
CTR/CVR Conversion
Auto Bidding
Click
Auto Bidding
Slide 11
Slide 11 text
Machine Learning &
Algorithms
System Architecture
Two and a Half Years Ago…
Slide 12
Slide 12 text
Takeaways
Gather Professionals in the Ad Industry!
Don’t Focus on the Beauty of your Logic! Then What?
Make your System the Industry Standard!
Take a Look at the Features that are Related to Machine Learning & Logic
Slide 13
Slide 13 text
Takeaways
Gather Professionals in the Ad Industry!
Don’t Focus on the Beauty of your Logic! Then What?
Make your System the Industry Standard!
Take a Look at the Features that are Related to Machine Learning & Logic
Slide 14
Slide 14 text
System Architecture
> Standard Protocols such as OpenRTB and VAST
> Scalability
> Components that Follow the Industry Standard
Slide 15
Slide 15 text
System Architecture
A Brief View of the AD System
Slide 16
Slide 16 text
System Architecture
A Brief View of the AD System
Supply Side Platform
SSP
Demand Side Platform
DSP
Slide 17
Slide 17 text
System Architecture
Supply Side Platform
SSP
Demand Side Platform
DSP
A Brief View of the AD System
Slide 18
Slide 18 text
Supply Side Platform
SSP
Demand Side Platform
DSP
System Architecture
A Brief View of the AD System
Ad Serving Ad Manager
Slide 19
Slide 19 text
System Architecture
A Brief View of the AD System
Supply Side Platform
SSP
Demand Side Platform
DSP
Data Management Platform
DMP
Data Pipeline
Pipeline
Slide 20
Slide 20 text
System Architecture
Some Examples
Supply Side Platform
SSP
Demand Side Platform
DSP
OpenRTB
Slide 21
Slide 21 text
Some Examples
System Architecture
Supply Side Platform
FIVE SSP
Demand Side Platform
DSP
OpenRTB
Slide 22
Slide 22 text
Takeaways
Gather Professionals in the Ad Industry!
Don’t Focus on the Beauty of your Logic! Then What?
Make your System the Industry Standard!
Take a Look at the Features that are Related to Machine Learning & Logic
Advertiser
Target Event: App Install
Willingness To Pay: 5,000 Yen per App Install
Auto Bidding With Bid Cap Supporting App Install Conversion
:
Couldn’t Be Happier
If You Can Lower the Price Under 5,000 Yen
Slide 35
Slide 35 text
Auto Bidding With Bid Cap Supporting App Install Conversion
:
CTR (Click Through Rate)
CVR (Conversion Rate)
BA (Bidding Amount)
Building Block
Slide 36
Slide 36 text
Before This,
We Prioritize Other Issues
Slide 37
Slide 37 text
3 Important Things Before Logic
Real-time Monitoring
Real-time Serving
Experimental Environment
Slide 38
Slide 38 text
Real-Time
Monitoring
Slide 39
Slide 39 text
Real-Time Monitoring
Slide 40
Slide 40 text
Prometheus & Grafana
Real-Time Monitoring
Slide 41
Slide 41 text
Druid / Imply
OLAP
Slide 42
Slide 42 text
Experiment Environment
Slide 43
Slide 43 text
Experimental Environment
Offline Test Online Test
Data Pipeline
Slide 44
Slide 44 text
Online Test
Offline Test Multi A/B
Test
Experimental Environment
Slide 45
Slide 45 text
Real-Time Serving
Slide 46
Slide 46 text
Latency
Data
Pipeline
Model
Performance
Real-Time Serving
Slide 47
Slide 47 text
Latency
Joiner Model
Performance
Real-Time Serving
QA Tools
Various Logs
Slide 48
Slide 48 text
3 Important Things Before Logic
Real-time Monitoring
Real-time Serving
Experimental Environment
Slide 49
Slide 49 text
Real-time Monitoring
Real-time Serving
Experimental Environment
Interest
Consideratio
n
Convers
ion
Retentio
n
Brand
Awareness
Slide 50
Slide 50 text
No content
Slide 51
Slide 51 text
Takeaways
Gather Professionals in the Ad Industry!
Don’t Focus on the Beauty of your Logic! Then What?
Make your System the Industry Standard!
Take a Look at the Features that are Related to Machine Learning & Logic
Slide 52
Slide 52 text
Latency
Real-Time Model
Performance
CTR Prediction
Slide 53
Slide 53 text
Latency
Real-Time Model
Performance
CTR Prediction
Logistic
Regression
Factorized
Machine
Field-Aware
Factorized
Machine
FTRL-
Proximal
Gradient
Boosting Tree
Clustering
Slide 54
Slide 54 text
CVR Prediction
Impression
Click Conversion Conversion
30 Days Later…
Slide 55
Slide 55 text
CVR Prediction
Data
Cleansing
Rarity Delay Multi-Label
Slide 56
Slide 56 text
CVR Prediction
Data
Cleansing
Rarity Delay Multi-Label
Logistic
Regression
Delayed
Feedback Model
L-BFGS FTRL-
Proximal
Exponential
Distribution
Poisson
Distribution
Slide 57
Slide 57 text
Given Total Budget
Auto Bidding
small
Bidding Amount 3
Time
Winning Price
Budget1 Budget2
Bidding Amount 2
Bidding Amount 1 Total Budget
Slide 58
Slide 58 text
About 2 and a Half Years Ago…
Conversion
Events Added
CTR/CVR Conversion
Auto Bidding
Click
Auto Bidding
Slide 59
Slide 59 text
About 2 and a Half Years Ago…
Conversion
Events Added
CTR/CVR Conversion
Auto Bidding
Click
Auto Bidding
Slide 60
Slide 60 text
About 2 and a Half Years Ago…
Conversion
Events Added
CTR/CVR Conversion
Auto Bidding
Click
Auto Bidding
Slide 61
Slide 61 text
About 2 and a Half Years Ago…
Conversion
Events Added
CTR/CVR Conversion
Auto Bidding
Click
Auto Bidding
About 2 and a Half Years Ago…
Conversion
Events Added
CTR/CVR Conversion
Auto Bidding
Click
Auto Bidding
New
Look-Alike Engine R&F Simulator Optimization Status
?
Slide 71
Slide 71 text
Auto
Delivery
Auto
Bidding
Auto
Experiment
Auto Series
Slide 72
Slide 72 text
Takeaways
Gather Professionals in the Ad Industry!
Don’t Focus on the Beauty of your Logic! Then What?
Make your System the Industry Standard!
Take a Look at the Features that are Related to Machine Learning & Logic