Analytics Design Patterns

A first attempt at formalizing a set of Analytics Design Patterns created for the data2day conference.

Transcript

1. analytics design
   patterns
   Valentin Zacharias, 27.11.2014
   

   View Slide

2. A design pattern is a general reusable solution
   to a commonly occurring problem in the
   design of analytics solutions.
   It is a best practice template for how to
   solve a problem that can be used in many
   different situations.
   Patterns establish a language to organize
   best practices.
   Patterns establish a knowledge baseline.
   (partly adapted from Wikipedia’s definition of software design patterns)
   

   View Slide

3. motivation
   how to scale delivery of analytics services?
   (beyond algorithms and tools)
   how to organize analytics know how and best
   practices?
   how to systematically bridge the gap between
   business need and tools / algorithms?
   

   View Slide

4. goal (of this presentation)
   spread and discuss ideas on how to structure
   know how and best practices in analytics
   present some interesting analytics design
   patterns that might be of use
   

   View Slide

5. me
   ● 13 year experience in software driven
   innovation
   o AI researcher at FZI / PhD Student at KIT
   o Manager of a research division at FZI concerned
   with all aspects of information driven decisions
   o Big Data consultant with codecentric
   o Analytics Consultant / Data Scientist with Daimler
   TSS
   

   View Slide

6. structure
   Generic CRISP-DM
   as organizational
   structure (for this
   talk)
   

   View Slide

7. business understanding
   Source: Crisp-DM Process
   

   View Slide

8. project patterns
   ● Data First: Find, understand and utilize
   patterns in a given data set
   ● Decision Driven: Identify and realize
   concrete analytics use case
   ● Metrics Model: Create systems to
   understand and optimize a business's core
   driver’s of growth
   ● Integration First: Break open data silos and
   harness integration
   

   View Slide

9. value patterns
   ● Data Driven Business
   o Operational Excellence
   o Customer Intimacy
   o Product Leadership
   ● Analytics as added Value
   o For Capitel Goods
   No unplanned Downtime
   Asset Optimization
   Enterprise Optimization
   ● Data & Analytics as Business
   

   View Slide

10. data
    understanding
    Source: Crisp-DM Process
    

    View Slide

11. data understanding
    patterns
    Data Source Visualization
    Event Log Exploration
    Query Based Data Extracts (w/wo sampling)
    Analytical Record
    Reverse Pivot
    Data Audit
    

    View Slide

12. data source visualization
    ● Database structure
    visualization
    ● Colors:
    o Blue=numeric
    o Green=categoric
    o Yellow=data/time
    o Red = other
    o White = missing
    

    View Slide

13. Analytical Record (1/2)
    Most Data Mining tools and methods demand
    their input structured around one kind of
    observation (rows) with a number of values
    (columns).
    The extract that represent the elements of
    interest (customer/product/process instance..)
    in this way is called the Analytical Record
    

    View Slide

14. Analytical Record (2/2)
    ● e.g. a feature vector representing a hosts
    behaviour at a point of time based on log
    data from IDS/Firewall (for intrusion
    detection)
    ● e.g. a feature vector representing a
    customer’s recent interactions with the
    company (calls to hotline, use of services)
    from different transactional / log systems
    

    View Slide

15. data preparation
    Source: Crisp-DM Process
    

    View Slide

16. data preparation patterns
    ● handling outliers
    o Winsorization
    o Trimming
    o Spatial Sign
    ● data weighting and balancing
    o Case Weighting
    o Oversampling / Undersampling
    o Partitioning
    

    View Slide

17. data preparation patterns
    ● handling missing values
    o Case Deletion
    o Available Case Analysis
    o Single Imputation
    Mean / Reasonable Value
    Regression
    o Multiple Imputation
    Regression
    o Maximum Likelihood
    

    View Slide

18. data preparation patterns
    ● data transformation
    o Generalization Abstraction
    o Temporal Abstraction
    o Rates
    o Dummy Variables for Categories
    o Centering and Scaling
    o Polynomials
    o Interaction Terms
    o Box Cox transformation
    

    View Slide

19. data preparation patterns
    ● Feature Selection
    ○ Filter / Proxy Based
    ○ Subset Selection
    ○ Embedded
    ○ Dimensionality Reduction
    

    View Slide

20. modeling
    Source: Crisp-DM Process
    

    View Slide

21. modeling
    ● Descriptive Modeling
    o density estimation
    o dependency modeling
    o cluster analysis / segmentation
    ● Predictive Modeling
    o Classification
    o Regression
    ● Pattern / Rule Discovery
    ● Content Based Retrieval
    ● Recommendation
    

    View Slide

22. Evaluation
    ● Training/Test/Validation
    ● LOOCV / K-Fold CV
    ● Bootstrap
    ● Expected Value and Cost Benefit Matrix
    

    View Slide

23. Deployment
    Source: Crisp-DM Process
    

    View Slide

24. Deployment Patterns
    ● PMML
    ● A/B Testing
    ● Canary Pattern
    ● (Visualizations)
    

    View Slide

25. Credits
    

    View Slide

26. end.
    valentin zacharias
    www.vzach.de
    me@vzach.de
    

    View Slide