KDD 2017

Transcript

1. KDD 2017
   Oey, Michael Julio
   Gregorius Edwadr
   

   View Slide

2. KDD
   ● Held by Special Interest Group on Knowledge Discovery and Data
   Mining ( SIGKDD )
   ● Part of Association for Computing Machinery ( ACM )
   ● SIGKDD has become an official ACM SIG since 1998
   ● ACM SIGKDD has hosted an annual conference since 1995
   ● Considered as the most influential forum for knowledge discovery and
   data mining research
   GDP Labs Confidential
   

   View Slide

3. KDD
   2012
   Beijing
   2013
   Chicago
   2014
   New York
   2015
   Sydney
   2016
   San Fransisco
   2017
   Halifax
   GDP Labs Confidential
   

   View Slide

4. KDD 2017
   ● Held at Halifax, Canada
   ● August 13th - 17th 2017
   ● Consists of :
   ○ Tutorial
   ○ Workshop
   ○ Main Conference
   ■ Keynote
   ■ Paper presentations
   ● Sponsored by a lot of named tech companies
   from around the world
   GDP Labs Confidential
   

   View Slide

5. KDD 2017
   ● From Theory to Data Product - Applying Data Science Methods to Effect
   Business Change
   ● A/B Testing at Scale : Accelerating Software Innovation
   ● Workshop on Causal Discovery
   ● Anomaly Detection in Finance
   ● Platforms and Infrastructure
   ● Applied Machine Learning
   ● Deep Learning
   ● Intelligent Systems and Data Science
   ● KDD Panel : The Future of Artificially Intelligent Assistants
   ● Clustering
   ● Web Applications
   GDP Labs Confidential
   

   View Slide

6. From Theory to
   Data Product :
   Applying Data
   Science Methods
   to Effect Business
   Change
   ● Advanced analytic entry points
   ○ The Technology Directive
   ○ The Field of Dreams
   ○ The Ambitious Executive
   ○ The Smart Competitor
   ● Are you asking the right questions?
   ○ Business valuable questions
   ● Agile Approach to Data Driven
   Decision Making
   ○ What is agile?
   ○ Managing uncertainty
   ● Data science is expensive, but if you
   don’t do it, you will be left behind
   Danielle Leighton
   Lindsay Brin
   Janet Forbes
   T4G Limited
   GDP Labs Confidential
   

   View Slide

7. Google Vizier - A
   Service for Black
   Box Optimization
   Daniel Golovin
   Google Research
   ● Used on many applications
   ○ A/B Testing
   ○ Machine Learning
   ○ Physical Design
   ○ Robotics
   ● Vizier
   ○ Easy to use
   ○ Reliable
   ○ Scalable
   ○ Flexible
   ○ State of the art
   ● They tried Vizier to bake chocolate
   chip cookies
   GDP Labs Confidential
   

   View Slide

8. TFX : TensorFlow
   Extended - A
   Production Scale
   ML Platform
   Heng Tze Cheng
   Google Research
   ● Productionizing ML pipeline is hard
   ● End to end ( from data to serving )
   ● Design principles :
   ○ One ML platform for many
   products
   ○ Continuous training and serving
   ○ Human in the loop
   ○ Reliable and scalable
   ● Steps :
   ○ Data analysis
   ○ Data validation
   ○ Data transformation
   ○ Trainer
   ○ Model validation & evaluation
   ○ Serving
   ● Documentation -> passive way
   ● Education -> active way ( teach, talk )
   ● Automation -> key principles
   GDP Labs Confidential
   

   View Slide

9. Designing AI at
   Scale to Power
   Everyday Life
   Rajesh Parekh
   Facebook
   ● Data drives all product at Facebook
   ● FB Learner Flow
   ○ Help non ML people to use ML
   ○ 70% users are non AI-experts
   ○ 25% engineers are active users
   ● Applied Machine Learning at Facebook
   ○ Computer Vision
   ○ Deep Text
   ○ Speech and Video
   ○ AI Powered Camera
   ● What’s next
   ○ Multi modal learning
   ○ Transfer learning
   ○ Multi lingual modeling
   ○ Weakly supervised /
   unsupervised learning
   GDP Labs Confidential
   

   View Slide

10. Industrial Machine
    Learning
    Josh Bloom
    General Electric
    ● Data produced by machines will
    overcome data produced by people
    ● Industrial Machine Learning
    ○ Preventive Maintenance
    ○ Anomaly / Failure detection
    ○ Etc
    ● Industrial level is more dangerous than
    social level
    ● Optimization metric
    ○ Higher accuracy != higher value
    ○ Drive towards higher
    precision/low FPR
    ● Models
    ○ Physical model
    ○ Data driven model
    ● Interpretability and accuracy trade off
    ● Small improvement helps a lot
    ● Hard to convince everyone to
    transition
    GDP Labs Confidential
    

    View Slide

11. KDD 2017
    ● Deep Learning for Personalized Search and Recommender Systems
    ● Recent Advances in Feature Selection: A Data Perspective
    ● 2017 Edition of AdKDD and TargetAd
    ● Three Principles of Data Science: Predictability, Stability, and
    Computability
    ● Supervised Learning
    ● Deep Learning
    ● Intelligent Systems and Data Science
    ● Hands-On Tutorial Declarative, Large-Scale Machine Learning with
    Apache SystemML
    ● Hands-On Tutorial Tensor Flow
    GDP Labs Confidential
    

    View Slide

12. Online Advertising
    GDP Labs Confidential
    

    View Slide

13. How ad works
    GDP Labs Confidential
    

    View Slide

14. How ad works
    GDP Labs Confidential
    

    View Slide

15. How ad works
    GDP Labs Confidential
    

    View Slide

16. How ad works
    Advertiser
    Publisher
    GDP Labs Confidential
    

    View Slide

17. Ad Pricing
    ● Popular Model
    ○ CPM -> cost per mille impression
    ○ CPC -> cost per click
    ○ CPA -> cost per acquisition
    ● Others
    ○ CPF -> cost per follower
    ○ CPV -> cost per view
    ○ CPI -> cost per install
    ○ CPD -> cost per download
    GDP Labs Confidential
    

    View Slide

18. Criteo
    GDP Labs Confidential
    

    View Slide

19. Criteo
    GDP Labs Confidential
    

    View Slide

20. Attribution Model
    GDP Labs Confidential
    

    View Slide

21. Attribution Probability
    GDP Labs Confidential
    

    View Slide

22. Attribution Aware Bidder
    GDP Labs Confidential
    

    View Slide

23. Bidding
    GDP Labs Confidential
    

    View Slide

24. Result
    GDP Labs Confidential
    

    View Slide