hat

Data Science in Practice

Five common applications of data science with concrete,
real-life use cases

Presented by
Yhat
http://yhat.com/
June 2016

hat

| Data Science in Practice

What is data science...actually?

How do real companies use data science to make products
and operations better?
What does the data science lifecycle look like?
In our first whitepaper, Applied Data Science, we translated the hype-y
term of data science into the plain english definition of using data to
make better decisions, optimize processes and improve products and
services. We also described the central goal of data science: getting
statistical models into production.

In this whitepaper we introduce five common applications of data

science that build upon those concepts. Our goal is to debunk the
impression that data science is some type of obscure black magic and
give you concrete examples of how it is applied in reality. Youll learn how
real companies are using data science to make their products and dayto-day operations better. Last but not least, we describe the data science
life cycle and explain Yhats role in getting models into production.

hat

| Data Science in Practice

APPLICATION 1:
RECOMMENDER SYSTEMS
Recommender systems, also known as
recommender engines, are one of the most
well known applications of data science.
Recommender systems are a subclass of
information filtering systems, systems that
cut through the noise of all options and
present users with just the subset of options

WHAT IS A RECOMMENDER SYSTEM?

A model that filters information to present users
with a curated subset of options theyre likely to find
appealing
HOW DOES IT WORK?
Generally via a collaborative approach
(considering users previous behavior) or contentbased approach (based on discrete assigned
characteristics)
WHAT IS A REAL USE CASE?
Tendril uses recommendation models to match
eligible customers with new or existing energy
products

theyll find appealing. The data being filtered

can range from products on an e-commerce

approach to filtering. Collaborative filtering

site to dating matches that appear as you

considers a users previous behavior, as well

search for the one.

as the behavior of similar users. Contentbased filtering provides recommendations

Recommender Systems

based on discrete attributes or assigned

characteristics.
Data scientists at energy software company

Content Filtering
(location, age, gender)

Collaborative Filtering
(previous behavior, similar users)

Tendril opted for a hybrid approach that

combines both collaborative and contentbased filtering. Tendril provides analytics
and consumer solutions to energy suppliers,

Recommender
System

including which energy products consumers

would most likely consider. We use Support
Vector Regression models to predict
household energy consumption to provide
our clients with in-depth, personalized
information about their customers, explains

Recommender systems offer a more

Mark Gately, Data Analytics Manager at

intelligent approach to information filtering

Tendril. This detailed information is also

than a simple search algorithm by introducing

used in recommendation models, which help

users to items they might not have otherwise

match eligible customers with new or existing

discovered. Recommender systems generally

energy products.

take either a collaborative or content-based

hat

| Data Science in Practice

APPLICATION 2:
CREDIT SCORING

One such company is Ferratum Bank,

a pioneer in financial technology and
mobile consumer lending since 2005. We
developed complex statistical and machine

If you have ever applied for a credit card

learning models to enable smarter lending

or a loan, youre likely already familiar with

decisions, explains Scott Donnelly, Director

the concept of credit scoring. What you

of Business Lending at Ferratum Bank. By

may be less aware of is the set of decision

getting creative with our approach and

management rules evaluating how likely

adopting innovative technologies, weve been

an applicant is to repay debts behind the

able to reinvent how both consumers and

scenes.

businesses obtain loans. This has allowed us

to reach prospective customers that in the

WHAT IS CREDIT SCORING?

A model that determines an applicants
creditworthiness for a mortgage, loan or credit card

past may have been overlooked by traditional

HOW DOES IT WORK?

A set of decision management rules evaluates how
likely an applicant is to repay debts

APPLICATION 3:
DYNAMIC PRICING

WHAT IS A REAL USE CASE?

Ferratum Bank uses machine learning models to
reach prospective customers that may have been
overlooked by traditional banking institutions

banking institutions.

You walk out of the store, arms full of

groceries, only to realize that a torrential
downpour began as you perused the

The first general purpose credit scoring

produce inside. You struggle to retrieve your

algorithm, now known as the FICO score, was

phone, check your favorite ride app and are

introduced in 1989. The FICO score is still

dismayed to find...a 2.1x surge!? Welcome to

one of the most widely used models in the

your first lesson on dynamic pricing.

United States today, though peer-to-peer and

direct lending organizations have focused
on developing new techniques over the
past few years. These new machine learning
models and algorithms capture innovative
factors and relationships that traditional
loan scorecards couldnt, like how applicants
manage monthly cash flow or whether
friends or community members would
endorse the applicant.

hat

| Data Science in Practice

WHAT IS DYNAMIC PRICING?

Modeling price as a function of supply, demand,
competitor pricing and exogenous factors
HOW DOES IT WORK?
Generalized linear models and classification trees
are popular techniques for estimating the right
price to maximize expected revenue
WHAT IS A REAL USE CASE?
Turo uses dynamic pricing models to suggest prices
to the people who list and rent out cars

Businesses use dynamic pricing algorithms

ago we started to model supply and demand

to model rates as a function of supply,

dynamics, so working on dynamic pricing was

demand, competitor pricing, and exogenous

an intuitive next step.

factors (e.g. weather or time). Many fields,

from airline travel to athletics admission

We quickly realized that the gap between

ticketing, employ dynamic pricing to

model development and model deployment,

maximize expected revenue. The nuts and

in production, was much bigger than

bolts of dynamic pricing strategies vary

expected. It requires a very wide spectrum of

widely, though generalized linear models and

skills: from knowledge of statistical modeling

classification trees are popular techniques for

to software architecture best practices. We

estimating the right (lowest/highest) price

use Yhats platform, ScienceOps, to transform

that consumers are willing to pay for a book,

our dynamic pricing prototype into a

a flight, or a cab.

production-ready algorithm in the languages

our Data Science team prefers to work in, R

Dynamic Pricing

Price (e.g. fare of ride)

D2

and Python.

D1
P2
P1

APPLICATION 4:
CUSTOMER CHURN
Churn rate describes the rate at which
customers abandon a product or service.
Understanding customers likelihood to churn
is particularly important for subscription-

Q1

Q2

Quantity (e.g. # of rides)

based models, everything ranging from

traditional cable or gym memberships to
recently popularized monthly subscription

Turo, a peer-to-peer car rental service

operating in over 2,500 cities, uses dynamic
pricing to suggest prices to the people who
list and rent out their cars on the platform.
Dynamic pricing helps us to balance supply
and demand and ensure that both our
travelers and our hosts are getting a fair
market deal, explains Jrme Selles, Director
of Data Science and Analytics. Three years

hat

| Data Science in Practice

boxes.
Data scientists looking to predict customer
churn may consider a variety of algorithms
for the job, such as support vector machines,
random forest, or k-nearest-neighbors.
Beyond the accuracy of a given model, data
scientists must also balance the tradeoff
between precision (correctly predicting a

churning customer) and recall (how many

not graduating more efficiently, so they can

predictions were actually successful). So

intervene and help those students graduate.

whats better? Classifying every churning

customer but occasionally mislabeling a
non-churning customer? Or identifying fewer
churning customers, but not mislabeling nonchurners? Its a difficult decision that requires
in-depth knowledge of the business case and
years of experience.

APPLICATION 5:
FRAUD DETECTION
Financial technology, or FinTech, companies
offer financial services like banking, investing,
and payment processing via software, rather
than through traditional banking institutions.

WHAT IS CUSTOMER CHURN?

Predicting which customers are going to abandon
a product or service
HOW DOES IT WORK?
Data scientists may consider using support vector
machines, random forest or k-nearest-neighbors
algorithms
WHAT IS A REAL USE CASE?
EAB combines data from transcripts, standardized
test scores, demographics and more to identify
students at risk of not graduating

These are familiar questions for the data

scientists at EAB, the education division of
The Advisory Board Company. EAB provides
data driven applications and insights to
hundreds of institutions of higher education.
A key component of our Student Success
Collaborative product, used by academic
advisors and other administrators, is a
predictive model of student graduation, says
Harlan Harris, Director of Data Science. We
combine data from transcripts, standardized
test scores, demographics, and other facts
about students to provide a graduation
risk score. Colleges and universities use
these scores to identify students at risk of

hat

| Data Science in Practice

Companies processing massive volumes of

financial transactions also need a quantifiable
way to detect and prevent fraudulent
transactions from being processed.
WHAT IS FRAUD DETECTION?
Detecting and preventing fraudulent financial
transactions from being processed
HOW DOES IT WORK?
Fraud detection is a binary classification problem:
is this transaction legitimate or not?
WHAT IS A REAL USE CASE?
Via SMS Group uses a combination of complex
data lookups and decision algorithms written in
R and implemented in PHP to assess whether a
loan applicant is fraudulent

Traditional fraud detection presents a fairly

straightforward problem: Is a transaction
legitimate or not? Otherwise called a binary
classification problem. This can be trickier
than it seems, especially when you have
thousands (or even millions) of legitimate
transactions occurring for every instance
of fraud. To add insult to injury, a single
occurrence of fraud can cost a company an
exorbitant amount of money. To combat this,

some data science teams pair supervised

data can be used to inform decisions,

classification techniques with anomaly

optimize processes, and improve products

detection algorithms to identify outliers and

and services across a very wide range of

pick out suspicious behavior.

business problems. Regardless of the specific

question at hand, the data science lifecycle

Dmitrijs vovs is responsible for managing

always culminates with selecting the winning

risk at VIA SMS Group, where over 60MM

model strategy and implementing it into an

is loaned to consumers across 6 countries

application where real business value can be

every year. The risk analytics team at VIA

realized.

SMS Group use advanced algorithms to

assess whether an applicant is fraudulent

For data scientists who build their models

prior to considering whether or not to

in open source programming languages

underwrite the requested loan. We write our

like R and Python, the path to production

decision algorithms in the R programming

application is not always clear. Mobile and

language and implement them into our

web applications are built using platforms

web and mobile apps in the server-side

and frameworks like .NET, Ruby on Rails,

language of PHP. By using R, we can leverage

Java, PHP or Node.js, which cannot consume

a combination of complex data lookups

models written in R or Python. As a result,

and state of the art algorithms to identify

many models are abandoned after months

fraudulent transactions, explains Dmitrijs.

of work before they ever see the light of

day. Alternatively, data scientists advanced

THE DATA SCIENCE LIFECYCLE

As the five applications above demonstrate,

statistical procedures may be tossed

over the fence to engineers and manually
recoded into another language, a notably

Now what?

So when can we
go live with the
new model?

hat

| Data Science in Practice

Any of you know

what Gradient
Boosting is?

difficult, time consuming and error-prone

rely on ScienceOps to take data science

process.

models from prototype to production.

YHATS ROLE

To find out more about how your business

Yhats data science operations

system, ScienceOps, eliminates the

can deploy models rapidly, frequently and

reliably with ScienceOps, get in touch with
the Yhat team or schedule a demo today.

counterproductive barrier between data

scientists and engineers by making R and
Python models accessible via REST API.
Instead of translating models, data scientists
can deploy models to ScienceOps, where
automatically generated API endpoints make
production integration quick & easy.

WHAT IS SCIENCEOPS?
Yhats data science operations system that
eliminates the barrier between data scientists and
engineers
HOW DOES IT WORK?
ScienceOps makes R and Python models accessible
via REST API and provides a platform to monitor,
manage and scale data science models
WHAT IS A REAL USE CASE?
ScienceOps is used by companies around the
globe, including each of those highlighed in the five
applications above

Our core mission at Yhat is to allow data

scientists to deploy predictive models rapidly,
frequently and reliably, but we recognize
that a data scientists job does not end there.
Beyond the initial step of implementing
models, ScienceOps also provides the ability
to monitor, manage and scale models.
Companies around the globe, including each
of those highlighted in the use cases above,

hat

| Data Science in Practice

Works Cited
Chiang, Eric. Predicting Customer Churn with Scikit-learn. The Yhat Blog. Yhat, 20 Mar. 2014.
Web.
Huang, Cheng-Lung, Mu-Chen Chen, and Chieh-Jen Wang. Credit Scoring with a Data Mining

Approach Based on Support Vector Machines. Expert Systems with Application 33 (2007):

847-56. Web.

Leskovec, Jure, and Jeffrey Ullman. Recommendation Systems. Mining of Massive Data Sets.

Ed. Anand Rajaraman. 2.1 ed. Cambridge: Cambridge UP, 2014. 307-41. Print.

Phua, Clifton, Vincent Lee, Kate Smith, and Ross Gayler. A Comprehensive Survey of Data

Mining-based Fraud Detection Research. Web. <https://arxiv.org/pdf/1009.6119.pdf>.

Yhat. Applied Data Science: Practical Guide to Building Data-driven Products beyond Analysts

Laptops. New York: Yhat, 2014. Print.

hat

| Data Science in Practice

About
Yhat (pronounced Y-hat) provides an end-to-end data science platform for developing,
deploying, and managing real-time decision APIs.
Yhats flagship product, ScienceOps, enables data scientists to transform static insights into
production-ready decision making APIs that integrate seamlessly with any customer- or
employee-facing app. Yhat also created Rodeo, an open source integrated development
environment (IDE) for Python.

Contact Us
http://yhat.com
[email protected]
(718) 855-2107
45 Main Street
Suite #707
Brooklyn, New York 11201

hat

| Data Science in Practice

10