© 2013 DataSong, all rights reserved / 234 Front Street, 3rd Floor, San Francisco CA 94111
Time-to-Event Models Consumer Interaction Insight
October 2013
Today’s Presenters
Tess A Nesbitt, PhD
Senior Data Scientist
John Wallace
Founder & CEO
Agenda
About us
Problem statement
High level modeling approach.
Use cases
Scoring systems
Q&A
DataSong at a Glance
Approaching $1 trillion in revenue analyzed. $2 billion in marketing spend under our lens.
Experienced 60 person team based out of San Francisco, with offices in Seattle, LA and India.
Founded in 2003 with a proven history of solving difficult analytics problems. Evolved from consulting
through close partnerships with clients.
Customer interaction insight that powers applications for customer level revenue attribution,
targeting, media optimization
Actionable and accurate information that drives customer acquisition and revenue growth for
modern direct marketers.
Patented big data approach models behavior at the individual consumer level.
DataSong Offerings
1. A regression modeling framework for prediction and inference
2. Automation of modelsets in Hadoop
3. Enterprise grade scoring in Hadoop
Modelset Creation: Current State
Flatten out the data
• 1. Aggregate a fact table (sum, count)
• 2. Join a dimension to a fact table and aggregate it (sum, count)
• 3. Superpose time
• If we have a dimension with a cardinality of 25 and 6 time periods of interest, that’s 150 variables for 1 dimension
AccountNo #SiteVisits
123456 5
AccountNo #Visits_SEO #Visits_EmailClick #Visits_SEM #Visits_...
123456 3 1 1 …
AccountNo #Visits_SEO_1Mo #Visits_SEO_3Mo #Visits_SEO_6Mo #Visits_SEO_...
123456 1 2 3 …
In Our Opinion
“Feature Engineering”
• Creating good variables is many times more important than choice of algorithm
Don’t lose track of time
• Age old practice of flattening data into 1 row per customer with 1000s of variables is limiting
Aggregations can obfuscate
Time series without customer- level data overlook important causal relationships
New Challenges for Predictive Modeling
More and more of our input data is generated from log files
• Large observational data (or if you want to call it Big Data, you can)
• We are approaching an infinite number of variables to test
Increasing # of use cases for real time scoring
Increasing # of opportunities to use models for inference
Understanding the Baseline Hazard
What Are We Doing About it?
Survival Response Model
• Explains differences in response rate as we change exposure to marketing
• Know what was significant and what wasn’t
Account ID-level analysis follows customers and cookies over time
Time-dependent Outcome had an event or was censored
Time-dependent Covariates the effect of an event is not constant
Time-varying Covariates time may modify an event effect
Controls for non-marketing effects:
Baseline Hazard Rate
Customer-driven activity many customers are driven by loyalty vs. marketing
Anniversary Effects many sales driven by season demand vs. marketing
CUSTOMER INTERACTION
BEHAVIORAL
LOYALTY
ID L
EV
EL
TIM
ES
TA
MP
DA
TA
MARKETING
SERVICE
LOYALTY
TELEMATIC
Prior Transactions
Impressions
DM
Referring clicks
In-store service
Call center
Inbound email/forms
Redemptions
Point balances
GPS data
Smart devices
EXAMPLE
CUSTOMER
SERVICE
CUSTOMER INTERACTION OBJECTIVE TIME APPROACH OUTCOME
BEHAVIORAL
LOYALTY
SITE VISIT
SU
BS
CR
IPT
ION
-CE
NT
RIC
ID
LE
VE
L T
IME
ST
AM
P D
AT
A
INFERENCE
PREDICTION
TIME-TO-EVENT
POINT-IN-TIME
Response
Model MARKETING
SERVICE
LOYALTY
TELEMATIC
PRICE/
PROMOTION
COMPETITION
SEASONALITY
UPGRADE
LEAVE
DEFAULT MA
CR
O D
AT
A
SITE VISIT
PURCHASE
CUSTOMER
SERVICE
CUSTOMER INTERACTION OBJECTIVE TIME APPROACH OUTCOME
BEHAVIORAL
LOYALTY
SITE VISIT
SU
BS
CR
IPT
ION
-CE
NT
RIC
ID
LE
VE
L T
IME
ST
AM
P D
AT
A
INFERENCE
PREDICTION
TIME-TO-EVENT
POINT-IN-TIME
Voluntary
Churn
Model
MARKETING
SERVICE
LOYALTY
TELEMATIC
PRICE/
PROMOTION
COMPETITION
SEASONALITY
UPGRADE
LEAVE
DEFAULT MA
CR
O D
AT
A
SITE VISIT
PURCHASE
CUSTOMER
SERVICE
CUSTOMER INTERACTION OBJECTIVE TIME APPROACH OUTCOME
BEHAVIORAL
LOYALTY
SITE VISIT
SU
BS
CR
IPT
ION
-CE
NT
RIC
ID
LE
VE
L T
IME
ST
AM
P D
AT
A
INFERENCE
PREDICTION
TIME-TO-EVENT
POINT-IN-TIME
Involuntary
Churn
Model
MARKETING
SERVICE
LOYALTY
TELEMATIC
PRICE/
PROMOTION
COMPETITION
SEASONALITY
UPGRADE
LEAVE
DEFAULT MA
CR
O D
AT
A
SITE VISIT
PURCHASE
SITE VISIT
CUSTOMER
SERVICE
PURCHASE
CUSTOMER INTERACTION OBJECTIVE TIME APPROACH OUTCOME
BEHAVIORAL
LOYALTY
SU
BS
CR
IPT
ION
-CE
NT
RIC
ID
LE
VE
L T
IME
ST
AM
P D
AT
A
INFERENCE
PREDICTION
TIME-TO-EVENT
POINT-IN-TIME
Simple
Attribution
Model
MARKETING
SERVICE
LOYALTY
TELEMATIC
PRICE/
PROMOTION
COMPETITION
SEASONALITY
UPGRADE
LEAVE
DEFAULT MA
CR
O D
AT
A
SITE VISIT
CUSTOMER
SERVICE
PURCHASE
CUSTOMER INTERACTION OBJECTIVE TIME APPROACH OUTCOME
BEHAVIORAL
LOYALTY
SU
BS
CR
IPT
ION
-CE
NT
RIC
ID
LE
VE
L T
IME
ST
AM
P D
AT
A
INFERENCE
PREDICTION
TIME-TO-EVENT
POINT-IN-TIME
Incremental
Attribution
Model
MARKETING
SERVICE
LOYALTY
TELEMATIC
PRICE/
PROMOTION
COMPETITION
SEASONALITY
UPGRADE
LEAVE
DEFAULT MA
CR
O D
AT
A
Customer
3
Customer
2
Customer
1
What Would the Model Say?
JANUARY FEBRUARY MARCH APRIL MAY JUNE
PURCHASE
CA
TA
LO
G
EM
AIL
CA
TA
LO
G
EM
AIL
EM
AIL
EM
AIL
CA
TA
LO
G
EM
AIL
$100 PURCHASE
PURCHASE $100 PURCHASE
PURCHASE $100 PURCHASE PURCHASE
DAYS SINCE TREATMENT SALES ALLOCATION
customer sales Catalog Email Retarget Cumulative
Orders Catalog Email Retarget Brand Loyalty
#1 $ 100 20 40 0 1 $ 95.66 $ 0.02 $ - $ 4.32
#2 $ 100 20 10 0 1 $ 77.52 $ 18.16 $ - $ 4.32
#3 $ 100 20 10 0 2 $ 69.94 $ 17.74 $ - $ 12.32
Functions Used Purpose
rxImport read in data from flat files
READ/WRITE rxDataStep read from XDF file, output to xdf file
rxReadXdf read from XDF file, can output to dataframe
rxSummary calculate summary stats on XDF file
rxCrossTabs build contingency tables of factors
EDA rxCube build contingency tables of factors
rxHistogram create histograms of numeric vars
rxQuantile calculate quantiles of numeric vars
rxLogit build logistic regression models
MODELING rxPredict score data from xdf with specifed model
rxRocCurve evaluate false and true positives of models
rxDTree* build classification and regression trees
Revolution R Enterprise ScaleR Functions Used
Run time for 30MM rows
and 30 variables is
approx 5 min
Prediction: Current State
How did we deliver?
Propensity Score (LOW HIGH)
Other models only use one dimension to predict likelihood to purchase: PROPENSITY
Prediction: DataSong Approach
Incrementality Metric
Sensitivity
Score
● Breakthrough results from adding customer sensitivity score: 14% increase in response rate
● Reallocated marketing circulation: Identified best prospects to not mail that were likely to purchase without receiving catalog
Propensity Score (LOW HIGH)
(LO
W
HIG
H)
Response modeling single channel: swap set usage
INCREMENTALITY metric predicts sensitivity of the next marketing treatment
Scoring Discussion
Scoring systems are like picture frames: good art is never without one
Your best model may never see the light of day
• Sharing your parameter estimates isn’t enough
Who should own scoring ?
• IT: Production support, high uptime mentality
• Analytics: often missing the software engineering discipline
Scale
Analytics teams should be able to manage dozens of models and score billions of records everyday
DataSong Architecture
• ETL
• N marketing channels
• Behavioral variables
• Promotional data
• Overlay data
• Functions to read Hadoop output; xdf creation
• Exploratory data analysis
• GAM survival models
• Scoring for inference
• Scoring for prediction
• 5 billion scores per day
per customer
DATASONG DATA
FORMAT (DDF)
CUSTOM VARIABLES
(PMML)
DataSong Contact
1. A regression modeling framework for prediction and inference
2. Automation of modelsets in Hadoop
3. Enterprise grade scoring in Hadoop
Linked In: www.linkedin.com/company/datasong
Facebook: www.facebook.com/datasong
Twitter: www.twitter.com/datasong
Phone: 877.540.5910
Email: [email protected]
Top Related