As the Big Data market has evolved, the focus has shifted from data operations (storage, access and processing of data) to data science (understanding, analyzing and forecasting from data). And as new models are developed, organizations need a process for deploying analytics from research into the production environment. In this talk, we'll describe the five stages of real-time analytics deployment: Data distillation Model development Model validation and deployment Model refresh Real-time model scoring We'll review the technologies supporting each stage, and how Revolution Analytics software works with the entire analytics stack to bring Big Data analytics to real-time production environments.

1. Revolution Confidential
R eal-Time P redic tive
A nalytic s with B ig Data
from deployment to produc tion
David M S mith @ revodavid
V P Marketing and C ommunity
R evolution Analytic s

2. Today we’ll dis c us s : Revolution Confidential
 What is “real-time predictive analytics with
big data?”
 Case study: UpStream Software
 Real-time big-data stack
 Five phases of deployment to production
 Just what do we mean by “Big Data” and
“Real-Time”, anyway?
 Recommendations
 Q&A
2

3. R eal-Time P redic tive A nalytic s with B ig Data
Revolution Confidential
 Real Time
 Milliseconds? Seconds? Hours? Days?
 In production, continuously updated
 Big Data
 Size, flow rate, diversity
 Conflict with ‘real time’
 Predictive Analytics
 Rear view: description, aggregation, tabulation
 Prediction, inference, statistical modeling
3

4. Revolution Confidential
C as e S tudy
www.upstreamsoftware.com
“Given that our data sets are already
in the terabytes and are growing
rapidly, we depend on Revolution R
Enterprise’s scalability and fast
performance — we saw about a 4x From: “How Big Data is Changing Retail
performance improvement on 50 Marketing Analytics”
million records. It works brilliantly.” http://bit.ly/upstream-webinar
John Wallace, CEO Upstream Software

5. UpS tream S oftware Revolution Confidential
UpStream’s Big Data Analytics engine
analyzes and optimizes marketing mix for
UpStream’s retailer clients
 Customer analytics and segmentation
 Revenue/ action attribution among channels and
marketing programs
 Customized Next Best Action per individual
prospect or customer
 Event-Triggered Marketing: responses to
consumer actions
5

6. B ig Data Revolution Confidential
 Demographics: consumer, product, market
 Actions: web clicks, email clicks, mobile app usage,
call center logs, social, search …
 Outcomes: impressions, touches, orders (retail, online,
mobile)
6

7. P redic tive A nalytic s Revolution Confidential
• Time-to-event models
UPSTREAM DATA • Exploratory data analysis CUSTOM VARIABLES
FORMAT (PMML)
• GAM survival models
• ETL • Scoring for inference
• N marketing channels • Scoring for prediction
• Behavioral variables
• 5 billion scores per day
• Promotional data per customer
• Overlay data

8. R eal Time Revolution Confidential
http://www.infoworld.com/d/big-data/williams-sonoma-uses-big-data-zero-in-customers-206641
November 8, 2012 8

9. P redic tive vs Des c riptive A nalytic s Revolution Confidential
 Retail sales are influenced by stores near the
home
 Newer customers are more sensitive to marketing
 Google keywords perform worse than you think
 Display advertising performs better than you think
 Online sales benefit more from seasonal events
than retail stores
9

10. R eal-time Revolution Confidential
B ig Data
P redic tive
A nalytic s
S tac k
10

11. Data L ayer Revolution Confidential
 Structured data
 RDBMS, NoSQL, Hbase, Impala
 Unstructured data
 Hadoop MapReduce
 Streaming data
 Web, social, sensors, operational systems
 Some descriptive analytics done here
11

12. A nalytic s layer Revolution Confidential
 Predictive analytics technology
 Development environment: build models
 Production environment:
 Deploy real-time scoring
 Engine for dynamic analytics
 Local data mart
 Static, periodically updated from data layer
 Improves performance
12

13. Integration L ayer Revolution Confidential
 Connective tissue between end-user
applications and analytics engine
 Engines for flow control, real-time scoring
 Rules engine / CEP engine
 API for Dynamic Analytics
 Brokers communication between app developers
and data scientists
13

14. Dec is ion L ayer Revolution Confidential
 End-user interface to analytics system
 Customers
 Operations
 Business analysts
 C-suite
 Familiar & simple interfaces
 Supports a range of end-user technologies
 Level of UI complexity dictated by need
14

15. R eal-Time Deployment Revolution Confidential
Five phases of deploying real-time predictive
analytics with big data to production:
1. Data Distillation
2. Model Development
3. Validation and Deployment
4. Real-time Scoring
5. Model refresh
16

16. P has e 1: Data Dis tillation Revolution Confidential
The data in the Data Layer isn’t yet ready for
predictive modeling. We need to:
 Extract features from unstructured text
 Topic modeling, sentiment analysis
 Combine disparate data sources
 Filter for populations of interest
 Select relevant features and outcomes for
modeling
 Export to the data mart for predictive modeling
17

17. E xample: Data Dis tillation in Hadoop Revolution Confidential
Log Files
Event Streams HDFS Load Map-Reduce Structured
Data
rmr
Scraped Text
Unstructured Analytics
Data Data Mart
Revolution R Enterprise for Hadoop: bit.ly/r-hadoop
18

18. A utomating the E xtrac tion P roc es s Revolution Confidential
 This process needs to be repeatable and
maintainable
 Create a re-usable R script:
 ASCII: rxDataStep
 SQL: rxImport, ROracle, RMySQL
 NoSQL: Rcassandra, rmongodb
 Hadoop: rmr, rhbase, Rhive
 Appliances: nza, teradataR, PL/R
 Feeds: rjson, XML, RCurl, twitteR, python
 R script runs from Analytics Layer
 Processing: in Data Layer
 Output: structured file for Data Mart
19

19. Data Revolution Confidential
Dis tillation
P roc es s
XD
F
20

20. P has e 2: Model development Revolution Confidential
 Goal: create a predictive model that is
 Powerful
 Robust
 Comprehensible
 Implementable
 Key requirements for Data Scientists:
 Flexibility
 Productivity
 Speed
 Reproducibility
21

21. T he Model Development C yc le Revolution Confidential
Download the White Paper
R is Hot
Variable bit.ly/r-is-hot
Transformation
Feature
Selection Model Predictive
Data Mart Sampling Estimation
Aggregation Model
Model
Model
Comparison /
Refinement
Benchmarking
22

22. Tec hnology c ons iderations Revolution Confidential
 Moving Big Data is slow
 So just move it once, to data mart near the development
and production environments
 Static data needed for modeling cycle
 But have a plan to refresh and update
 Data layer not optimized for predictive analytics
 But good for descriptive (data distillation)
 Revolution R Enterprise optimizations for the
analytics layer
 R; XDF file format; Parallel External Memory Algorithms
23

23. P redic tive Modeling A lgorithms Revolution Confidential
Algorithm Example Applications Big Data
ETL, data distillation, record/variable selection,
Data Step
variable transformation
✓✓
Descriptive Statistics Exploratory Data Analysis, Data Validation ✓✓
Tables & Cubes Reporting, contingency analysis ✓✓
Correlation / Covariance Factor Analysis, Value at Risk ✓✓
Linear regression Forecasting, Net present value estimation ✓✓
Logistic Regression Response modeling, offer selection ✓✓
Generalized Linear
Models
Capital reserve estimation, climate modeling ✓✓
K-means clustering Customer Segmentation ✓✓
Decision Trees Dynamic pricing, classification, variable importance ✓✓
Model Prediction Real-time Scoring (decisions, offers, actions) ✓✓
R CRAN packages Everything else
Parallel & distributed Simulations, By-Group analysis, ensemble models,
computing with R custom applications
✓
24

24. High performanc e with dis tributed c lus ters
Revolution Confidential
Data Scientists / Modelers Grid computing cluster
Revolution R Enterprise
Platform LSF
Microsoft HPC Server
Microsoft Azure
Ad-hoc grids (SNOW)
Shared SMP server
Hadoop / HDFS
25

25. P has e 3: Model Validation and Deployment Revolution Confidential
 Scoring rules map parameters to outputs
 Parameters: information known at real time
 prospect ID, product, webpage, …
 Scores: values inferred in real time
 prices, recommendations, actions, …
 Validation: backtest with historical data
 Deployment: code running in real time
26

26. Validation for produc tion Revolution Confidential
Real Outcomes
Historical
Data
Scores
Scoring
Rules
 Refresh data mart
 Rebuild model, withholding a validation set
 Random sampling
 Create accuracy measure
 ROC, positive rate, false positive rate
 Measure and monitor
27

27. Deployment with R c ode Revolution Confidential
Outcomes
Historical
Scores
R Model R Model
Data
Data
New
Object Object
 Scoring rules captured as “R model objects”
 Move R code directly from development
environment to production server
 No recoding: Lowest cost, greatest speed &
reliability
 Most flexibility (not limited in model choice)
 Easy to validate with custom accuracy measures
28

28. Managing and S c aling Deployed R C ode Revolution Confidential
 Revolution R Enterprise includes RevoDeployR
 Code management & tracking
 Security & resource allocation
 Scalability for real-time demands
 Web Services API
 Scoring and dynamic analytics
 Data visualization
 custom reporting
 Ad-hoc analytics
 Interactive data apps
29

29. C ode C onvers ion for S c oring Revolution Confidential
 Business rules (e.g. IBM ILOG)
 Create directly with R code from model object
 Recode in other languages
 SQL
 C++, etc.
 Low latency
 Slow and costly
deployment
 Potential for errors
30

30. S c oring via P MML Revolution Confidential
 Some predictive models can be expressed in
the PMML standard (www.dmg.org)
 Not all, but growing all the time
Scores
R Model PMML
Data
PMML
Object pmml
 Easily generate PMML with R “pmml” package
 PMML Scoring Engine: Zementis Adapa
 Databases and appliances may support PMML
 But supported standards vary
31

31. P has e 4: R eal-Time S c oring Revolution Confidential
 Scoring triggered by Decision Layer, brokered by
Integration Layer
 R code
 Revolution R Enterprise Server
 In-appliance (e.g IBM PureData System for Analytics)
 SQL
 In-database
 Rules
 Compiled code
 Bespoke engines
 May be using hardware from data layer
 But not the actual data
32

32. R eal-Time Revolution Confidential
S c oring:
review
iLog
SQL
Near Real
Time 33

33. P has e 5: Model R efres h Revolution Confidential
 Deployed scoring rules no longer connected
to Data Layer or Data Mart
 Enables real-time performance
 Need to refresh using recent data
 Model refresh process
 Use data extract script
 Re-run model script
 Statistical review OR automated validation
34

34. S c heduled B atc h Updates Revolution Confidential
 Periodic model refresh
 Weekly
 Daily
Revolution R Enterprise  Hourly
Production Server Cluster
 Automated
Scheduler validate/deploy process
Excel
RevoDeployR Server
Web Services API
35

35. R e-developing the model Revolution Confidential
 Even then, the underlying structure of the
data may change:
 Important variables become non-significant
 Non-significant variables become important!
 New data sources become available
 Model accuracy drift is a warning sign
 Return to Phase 2 or Phase 1
36

36. Revolution Confidential
Kilobytes / second
Megabytes / second
How big is
‘B ig data’?
Megabytes
Gigabytes  Terabytes
Petabytes  Exabytes
38

37. Revolution Confidential
Seconds or less
Milliseconds
How fas t is
‘R eal time’?
Daily  Hourly
Hours  Minutes
Daily  Hourly
Weeks  Days
39

38. R ec ommendations Revolution Confidential
 Architect your Predictive Analytics Stack
 Best of breed vs single-vendor stack
 Diversify data sources and decision apps
 R = Predictive Analytics
Revolution R Enterprise provides:
 Analytics Layer: big data, performance
 Integration Layer: scalability, reliability
 Get Help to Get Started
 Consider an SI partner for architecture, best
practices and implementation advice
40

39. R es ourc es Revolution Confidential
 Revolution R Enterprise : R for Big Data
 www.revolutionanalytics.com/products
 Revolution Analytics Consulting Services
 www.revolutionanalytics.com/services
 Contact us: bit.ly/hey-revo
 Rhadoop : Connecting R and Hadoop
 bit.ly/r-hadoop
 Contact David Smith
 david@revolutionanalytics.com
 @revodavid
 blog.revolutionanalytics.com
41

40. T hank you. Revolution Confidential
The leading commercial provider of software and support for the popular
open source R statistics language.
www.revolutionanalytics.com 650.646.9545 Twitter: @RevolutionR
42