Quantitative Research and Analytics, Proprietary and Confidential1 Ryan Michaluk [email protected]

Quantitative Research and Analytics, Proprietary and Confidential

1

Driving Better Business Results with Machine Learning on Hadoop

Ryan [email protected]

Introduction to Skytree

Quantitative Research and Analytics

Since 2012

Leading Platform for ML on Big Data

Data

Size

Model Complexit

y

Some laws of data science


PredictiveAccuracy

# ofIterations

Data

Size

Model Complexit

y



PredictiveAccuracy

# ofIterations But:

Computation Time

Data

Size

Model Complexit

y

Computation Time



PredictiveAccuracy

# ofIterations

Allstate - We are …


Largest public personal lines insurer in US

16 million households

40,000 employees

11,000 agencies

4 brands

Auto, home, life, retirement

Quotes – personal, item, offer

Policies – personal, item, history

Claims – loss, participant, repair, adjuster notes

Telematics – events, mileage

Agencies – policy, sales, region, internal

Data is the foundation of our business


Pricing

Fraud Prevention

Underwriting

Marketing

Customer Experience

Data drives our decisions


Pricing

Fraud Prevention

Underwriting

Marketing

Customer Experience

Make the best

possible decisions

Data drives our decisions


Statistical models that find patterns

Descriptive (good)What happened

Predictive (better)What will happen

Prescriptive (best)What action will produce the best outcome

What is machine learning


Action

Action

Action

Predictive Model

Result

Result

Result

Predictive model


Best ResultDo This

Prescriptive model


Action

Action

Action

Predictive Model

Result

Result

Result

Make data driven decisions

Adapt to change

Power scales with amount of data

Machine learning has many benefits


Data Challenges - accessible, usable

Resource Challenges - people, tools, time

Cultural Challenges - information is valuable,algorithms are useful

… but can be difficult to implement


Machine learning creates a big impact


Model Metric

New Algorithm .367

Old Algorithm .350

Old without most important variable

.339

Machine learning creates a big impact


Model Metric

New Algorithm .367

Old Algorithm .350

Old without most important variable

.339

Improvement

comparable to including

the most important variable

Better algorithms take more time

GLM Random Forest GBM

K-Means Topological Methods

More data takes more time

Machine learning is a process

Machine learning is hard


Change Paramet

ers

Build Model

Validate Model

ModelDevelopment

Cycle

Machine learning is iterative


Data Size

Model Complexity

Iteration TimeComputati

on Speed

Iteration Time

# of Iterations

Data Scientist Time

Data scientist time


Avoid hard / large problems

Reduce data size

Reduce model complexity

Reduce # of iterations

Increase computation speed

Reducing time requires tradeoffs


Avoid hard / large problems

Reduce data size

Reduce model complexity

Reduce # of iterations

Increase computation speed

Reducing time requires tradeoffs


Machine

Learning on

Hadoop

Challenges existAlgorithms don’t parallelize easilyMore than just model training

OptionsBuild your own

Exactly what you want, maybeReally hardLarge opportunity cost

Vended solution

Is ML on Hadoop right for you


Bring ML to data

Scalable ML environment

Improve existing solutions

Tackle new projects

Data scientists have more time to solve problems

ML on Hadoop is right for Allstate


ML on Hadoop is right for Allstate


Good for the Business

Good for Data Scientists

Questions


Documents

Quantitative Research and Analytics, Proprietary and Confidential1 Ryan Michaluk [email protected]