Naoki Watanabe (hackmylife)
LINE / Development Team A
Mainstream ad platforms today employ RTB (Real-time Bidding) where an auction is carried out for one impression in real-time. Faster ad load yields to better performance and the process of auction and ad determination needs to be completed in the matter of hundredth of a second. This is not as simple as returning high bid price - an ad has to be relevant to the audience and to achieve this, audiences' characteristics need to be inferred and have an optimal ad selected via machine learning.
Challenges (and fun) with ad platform lies in processing these complex task within a short span of time.
The delivery system not only has to be high-speed, but also requires accumulating and analyzing a vast amount of data in real-time.
In order to solve this complex problem, LINE has developed various systems on its own such as ad delivery system (SSP and DSP), Ad Ranker (for auction), Data Pipeline (for real-time logging and analysis), DMP (for managing inferred audience attributes and expanding targeting (look-a-like)), and many more.
This session outlines reasons LINE has decided to develop most of its ad platform on its own, and the system that LINE has been envisioning.
20. ● Id
● Advertiser
● Frequency
● Etc..
● Category
● Place
● Time
● Etc..
Audience AD Site
CTR Prediction Features
● Gender
● Age-range
● Interest
● Etc..
21. Feature Dimensions
gender × agerange × time × ad
2 × 9 × 24 × 1,000 = 432,000
For more details, see poster session
22. CTR Prediction on LAP
Feature Millions of feature dimensions
Main Model LR, FM, FFM
Learning Method FTRL online learning
Multi-armed Bandit Contextual Thompson Sampling
Prediction time Real-time (under few msec)
29. ● Collect as much information as possible
● Real-time machine learning (within 50ms)
● Enhance speed of PDCA (iteration) cycle
Conclusion
30. • Store and analyze huge data
• Stable servers and networks
• Fast and stable data pipeline
Essentials
31. ● Private cloud platform (Verda)
● Fully managed
● Server, LB, DNS, Storage, etc. can be create by UI
Stable Servers and Networks
For more details, see Hall A 13:20
32. ● Kafka cluster
● Message transmission between services
● Queue and job scheduling
Fast and Stable Data Pipeline
For more details, see Hall A 15:00
33. ● Robust Hadoop cluster called Datalake
● Huge database that collects all LINE data
● OASIS and Yanagishima make analysis easy
Store and Analyze Huge Data
For more Datalake details, see Hall A 15:50
For more Oasis details, see Hall B 16:40
37. ● Online meeting with simultaneous language interpretation
● Chat through LINE translation bot
● Frequent business trips— important to talk face to face
Team Communication