Big Data Day LA 2016/ Use Case Driven track - Shaping the Role of Data Science: An Evolution towards Prescriptive Analytics as Key Driver in Revenue Acceleration, Thomas Sullivan, Chief Data Scientist, IRIS.TV
At IRIS.TV, our business builds algorithmic solutions for video recommendation with the end goal to deliver a great user experience as evidenced by users viewing more video content. This talk outlines our reasons for expanding from a descriptive/predictive approach to data analytics toward a philosophy that features more prescriptive analytics, driven by our data science team.
Big Data Day LA 2015 - The Big Data Journey: How Big Data Practices Evolve at...
Semelhante a Big Data Day LA 2016/ Use Case Driven track - Shaping the Role of Data Science: An Evolution towards Prescriptive Analytics as Key Driver in Revenue Acceleration, Thomas Sullivan, Chief Data Scientist, IRIS.TV
Semelhante a Big Data Day LA 2016/ Use Case Driven track - Shaping the Role of Data Science: An Evolution towards Prescriptive Analytics as Key Driver in Revenue Acceleration, Thomas Sullivan, Chief Data Scientist, IRIS.TV (20)
Developer Data Modeling Mistakes: From Postgres to NoSQL
Big Data Day LA 2016/ Use Case Driven track - Shaping the Role of Data Science: An Evolution towards Prescriptive Analytics as Key Driver in Revenue Acceleration, Thomas Sullivan, Chief Data Scientist, IRIS.TV
2. Unleashing the Power of Prescriptive Analytics as a
Revenue Driver
Dr. Thomas J. Sullivan
Chief Data Scientist
Nicholas Oswald
Data Scientist
3. Outline
• About IRIS.TV
• Path Toward Prescriptive Analytics
• Mathematical Formulation
• Conditions for Success
• Human vs. Machine Tradeoff
• The Feedback Loop
• Use Case: Online Video
4. Personalized Video Programming
What is IRIS.TV?
The IRIS.TV Video Programming Platform is a
lightweight API made up of three components
easily integrated into your existing video
environment
• Video personalization across all devices
• Automated data structuring
• Business Intelligence & Programming
Management
6. Path Toward Prescriptive Analytics
“LEVER”
An available resource
that may be used to
generate an expected
response
7. A Simple Mathematical Explanation
With an understanding of:
• The expected outcome that changing lever(s) will have on a goal (e.g. increase
monthly revenue)
• A desired goal (optimize use of current resources and/or achieve a different level of
outcome)…
…prescriptive analytics can be used as a decision aid when identifying how levers
can be adjusted (“courses of action”)
EXAMPLE:
𝑦 = 𝛼 + 𝛽𝑥
𝑥 =
𝑦∗
− 𝛼
𝛽
Predict y:
Prescribe an estimated value of x that will
generate a desired outcome, y*:
8. Predictive Model Variable Type
1. Exogenous model variables: those input variables that can not be
modified by anyone (e.g. time, weather)
2. Immovable Levers: Resources over which the consumer of prescriptive
analytics has no control (e.g. a different department)
3. Movable Levers: Resources that may be re-allocated, as prescribed,
toward achieving a desired outcome (e.g. money, people, computing cycles)
Prescriptions, though considerate of all types of input variables, focus on
the movable levers, particularly those that result in greatest impact
9. Higher Dimensional Solution Frontier
Feasible Solution Area
Constraints may
make some
outcomes infeasible
𝑦 = 𝛼 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + …
10. In What Type of Environment Can Prescriptive Analytics be Useful?
• Levers and constraints are known
• The relationship between the outcome and the input levers are known or
estimated (with cause-effect established to an acceptable level)
• Consumer(s) of prescriptions has the ability and the willingness to change
levers
• Feedback loop exists:
– Historical Data Exists (Descriptive)
– Data are turned into predictive (via modeling, simulation, NN, etc) analytics
– Data-driven prescriptions are generated and implemented
– Observed effects - along with subject matter expertise - are used to validate and
refine prescriptive analysis
• Trust in prescriptions can evolve over time
12. CASE STUDY: NBA Finals
Background
• Week before NBA Finals
• Major sports client serving
videos
• Seeking prescriptions to
improve Video Lift
13. Prescriptive Illustrations
Levers that can be used to influence Video Lift
Video Lift = Recommended Views/User Experiences
* Video Video VideoVideo
* * *
14. Available Levers
Note: Levers ordered by expected effect on
Video Lift
Correct Supply / Demand Imbalance
Add Video Category Metadata
Reduce Average Length of Video
Add Valid Video Source ID
Add Video
Adjustable Levers for Increasing Video Lift
Potential Increase in Video Lift
15. Relevant Movable Levers
• Supply / Demand Analysis of portfolio assets (to meet audience
needs and retain them longer)
• Pilot for decomposing asset length and completeness of categories
(increase video lift)
• Heat maps for illustrating when viewers are on the site (staff
interaction and ad placement)
By Moving these 3 levers users will watch 20% more videos
20. Final Prescriptions
• Supply / Demand imbalance
• Metadata - Context
• Video Length
– Prescriptions based on decomposition of the three main levers in the previous
slides to isolate expected value of moving levers individually or jointly
– Better outcome expected if done before Monday noon when audience
engagement peaks
21. What is the Appropriate Balance Between Human and Machine?
Iterative interaction – as far as
possible, human informs machine,
machine infers rules
Lots of subject matter
expertise/little access to
suite of analytical tools
Little subject matter
expertise / insufficient
access to suite of
analytical tools
22. Closing Thoughts
“This is not a race against the machines.
If we race against them, we lose.
This is a race with the machines.”
-The Inevitable: Understanding the 12 Technological Forces That Will Shape Our Future