Ben Zamanzadeh - Data Science @ Datapop - LA Data Science meetup - January 2015

Transcript

1. Marketing Knowledge Graph
   Data Science & Machine Learning
   Dr. Ben Zamanzadeh
   CTO DataPop
   

   View Slide

2. 2
   • Distributed Store Front (Omni-Channel Retail)
   – Official Site, Search & PLA, Mobile, Social, Affiliates, Ad Networks,
   Shopping Sites, Digital Catalogs, Streaming Networks, Gaming & TV
   Consoles, Brick2Click, etc.
   • Digital Merchandising
   – Manage online Perception of Products & Brands
   • Smart Catalog
   – Integrate Knowledge of Consumer with Catalog
   • Rapidly Changing Landscape
   – Dynamic Market Conditions,
   – New Advertising methods, New Channels,
   – More Sophisticated competition
   • Large & Complex Operational & Data Scale
   Evolution of eCommerce
   

   View Slide

3. 3
   Dark Data
   Lack of visibility caused by:
   •  Big Dark Data
   •  Massive amount of noisy & complex data
   •  Waves in the Ocean Dilemma
   •  Disjointed operations
   •  Machine Learning Models lack of context
   •  Need Human to generate Actionable Insight
   

   View Slide

4. 4
   Improvement Cycle
   Convert Big Data into Knowledge Graph
   Generate Actionable Insights Semantic Advertising
   Beats Studio
   Over-Ear
   Headphones –
   Red –
   Beats by Dre
   

   View Slide

5. 5
   • Marketing Knowledge Graph
   – Marketing Knowledge Graph (MKG) consists of a semantic network
   of brands, products, retailers, consumer intent & sentiments in
   addition to the advertising performance statistics and history for all
   publisher’s and channels.
   – Marketing Knowledge Graph is a knowledge base used by DataPop’s
   Semantic Advertising to generate and enhance Advertising
   Campaigns.
   • SEMANTIC ADVERTISING
   – Semantic Advertising achieves meaningful and optimized
   advertising through the use of semantic networks linking
   brands, products, retailers, to the consumer intent and
   sentiments using performance metrics (AKA Marketing
   Knowledge Graph).
   Marketing Knowledge Graph & Semantic Advertising
   

   View Slide

6. 6
   • Search Knowledge Graph
   – Used for Semantic Search: It is used by search ranking algorithms
   to return most relevant results for search queries. Also it is
   intended to engage Searcher to explore Knowledge Graph for better
   answers. KG is a loosely connected network of vast number of
   entities spread around a very wide range of topics.
   • Marketing Knowledge Graph
   – Used for online Marketing & Advertising of Products and Services. It
   is used by data mining and analysis systems as well as advertising
   campaign management systems.
   – MKG is a tightly coupled semantic network, which is limited to
   products, brands, consumer intent, retailers and various advertising
   channels & publishers. Advertising performance is a major
   component of MKG, which changes rapidly due to changes in
   consumer behavior and eCommerce competitive landscape.
   • 
   Difference Between MKG and Search Knowledge Graph
   

   View Slide

7. 7
   Marketing Knowledge Graph
   MOST PINNED
   PRODUCT
   HALLOWEEN
   MINDSET
   PRICE MINDSET
   $
   

   View Slide

8. 8
   Going Strong with Marketing Knowledge Graph
   

   View Slide

9. 9
   • Named Entity Recognition
   – Conditional Random Fields
   – Support Vector Machine
   • Supervised Category Recognition
   –  Support Vector Machine
   –  Max Entropy
   –  Naïve Bayes
   –  Ensemble Methods
   • Unsupervised Linguistic Category Derivation
   • Topic Modeling, Explicit Semantic Analysis
   – Semantic Relatedness
   –  Semantic “Area Density”, Semantic Distance
   Our R&D Focus : Natural Language Processing & Machine Learning
   

   View Slide

10. 10
    • Max Entropy
    – Bag of words and features, Minimize un-observed
    assumptions
    – Maximize Information Entropy : Closest to uniformity
    • Advantages
    – Good for large set of categories
    – Fairly Fast predictor
    – Works better with mildly noisy data
    – Does not assume independence of words in the bag of words
    – Worked well with 3 M training set
    – Confidence score has almost linear relationship with F1
    • Disadvantages
    – Slower training
    Comparison of Statistical Methods
    

    View Slide

11. 11
    Max Entropy Confidence Score
    

    View Slide

12. 12
    • Support Vector Machines
    – Vector based solutions, using Cosine Similarity
    – Establish support vectors that isolate each category hyper-plane
    • Advantages
    – Strong learner (usually best in Vector based)
    – Fast training and prediction
    – F1 similar to Max Ent (slightly below)
    – Good for text, image, bioinformatics
    • Disadvantages
    – Complex training, Optimization
    – Scores not as useful
    – Sensitive to noise
    – Better with longer documents, more signal
    – Better with smaller number of classes
    Comparison of Statistical Methods
    

    View Slide

13. 13
    • Naïve Bayes
    – Bag of words, Statistical counting
    • Disadvantages:
    – Weaker learner at 3 M samples
    – Assumed words are independent in bag of words
    (not true)
    • Advantages
    – Adaptive
    • Recent Research on Training Data Size
    Comparison of Statistical Methods
    

    View Slide

14. 14
    ü  “Good Data” is the King
    ü  Unsupervised
    ü  Non-Human Supervised
    ü  Max Entropy
    ü  SVM
    ü  Ensemble
    ü  CRF
    ü  Mix of Rules & ML
    What Doesn’t Work
    Ⓧ Naïve Bayes
    Ⓧ Human Gen Supervised
    Ⓧ Standalone methods
    Ⓧ Just use ML methods
    Ⓧ Biased Data
    Ⓧ Noisy Data
    Named Entity Recognition & Categorization
    What Works
    

    View Slide

15. 15
    Sample Problem for Topic Modelling
    

    View Slide

16. 16
    • Lobbyist Objective
    –  Lobbyist gets five minutes on the senate floor to address senators
    –  Lobbyist is given 5 different subjects that he can pitch to the senators
    –  He can only pitch one of the subjects
    –  Does not know who is in room ahead of time, senators are scattered in the
    room and are busy talking to each other in groups.
    • He is given few seconds and a fast laptop
    –  Either select a group of senators and join them to present his case (any of
    the 5). He has to choose the most influential group that also is most
    interested in one of the subjects.
    –  Or go on the podium and address all senators and take the risk that he may
    loose all of the audience quickly if majority of room is not interested.
    • What he knows about each senator
    –  Level of influence, subcommittee membership, Past voting history
    –  What subjects senator is interested (from 1000’s of possible subjects)
    –  What each group is talking about right now (conversation subject)
    –  He can also see who is in each group (Floor map)
    Statement of problem: Lobbyist on the Senate Floor
    

    View Slide

17. 17
    Visualization of Semantic “Area Density”
    

    View Slide

18. 18
    Technology Stacks & Languages
    

    View Slide

19. We’re Looking
    Research Engineer Position
    Must Have Experience:
    Java, UIMA, Mallet, ClearTK
    NER, Ensemble Methods
    

    View Slide

20. Thank You
    Dr. Ben Zamanzadeh
    

    View Slide