Applied_Data_Science_Presented_by_Yhat

1ŷhat | Applied Data Science
Applied Data Science
Practical guide to building data-driven products beyond
analysts’ laptops
ŷ
Presented by
Yhat
http://yhathq.com/
March 2014

What are the byproducts of data science?
What are predictive models?
How can data science improve products?
How do you go from insight to prototype to production application?
This is a white paper about data science teams and how companies apply
their insights to the real world. You’ll learn how successful data science
teams are composed and operate and which tools and technologies they
are using.
We discuss the byproducts of data science and their implications beyond
analysts’ laptops and answer the question of what to do with predictive
models once they’re built. Lastly, we inspect the post-model-building
process to highlight the most common pitfalls we see companies make
when applying data science work to live data problems in day-to-day
business functions and applications.

Describing data science
In the wake of an increasingly digital
economy, businesses are racing to build
operational knowledge around the vast
sums of data they produce each day. And
with data now at the center of almost every
business function, developing practices for
working with data is critical regardless of your
company’s size or industry.
“Data science,” one of many recently
popularized terms floating amidst the myriad
of buzzwords and big data hoopla, is a field
concerned with the extraction of knowledge
from data. Practitioners—aptly named “data
scientists”—are those charged with solving
complex and sophisticated problems related
to data usually employing a highly diversified
blend of scientific and technical tools as well
as deep business and domain expertise.
The central goal of data
science
As is the case with any analytical project, the
central goal in data science is to produce
practical and actionable insights to improve
the business. That is to say, data scientists
overcome complexities involved in data
to empower businesses to make better
operational decisions, optimize processes,
and improve products and services used by
customers and non-technical employees.
“What distinguishes data science itself from the tools and techniques
is the central goal of deploying effective decision-making models to a
production environment. “
– John Mount & Nina Zumel, Practical Data Science with R

Profile of a typical data
science project
With broad strokes, a data science project
begins with some question, need, or goal
in mind and with varying degrees of focus.
Accordingly, a data scientist’s primary task at
the start of a new project is to refine the goal
and develop concrete project objectives.
Analysts will first conduct a preliminary
survey of the data, applying domain
knowledge to develop a clear and succinct
problem definition to serve as the principal
object of study.
Identify relevant data sets
With a narrow and expressive definition of
the problem, data scientists can begin to
evaluate different data sets to identify which
variables are likely to be relevant to the
problem they are trying to solve. Evaluating
which data sets should be used for the
project, however, is not an activity performed
in isolation. Most companies have numerous
data sets, each highly diverse in shape,
composition and size. Analysts may or may
not be familiar with a particular data source,
how to query it, where it comes from, what it
describes or even that it exists.
For these reasons, quantitative analysts are
usually working in proximity to or in direct
collaboration with engineers, marketers,
operations teams, product managers, and
other stakeholders to gain a robust and
intimate understanding of the data sources
at their disposal.
Cross-functional collaboration
Collaboration at this stage is not only
valuable for identifying which data are
relevant to a problem but also for ensuring
the ultimate viability of any resulting solution.
Hybrid teams composed of stakeholders in
EXAMPLE:
Detecting erroneous car listings at Vast.com, the
premiere marketplace for data powering vertical
search in automotive, travel and real estate.
BACKGROUND:
Vast ingests car-listing data from thousands of
suppliers and publishes listings to thousands of
marketplaces that trust the data are accurate.
The listing data itself is initially created manually
by users and is therefore vulnerable to human
error.
INITIAL BUSINESS NEED:
Identify inaccurate car listings on the fly and fix
them before they reach users.
FOCUSED PROBLEM DEFINITION:
We treat this as a binary classification problem
where, given a car listing created by a user, we
predict whether the asking price is sensible
given the vehicle’s characteristics (e.g. make,
model, year, odometer reading, etc.).

separate functions produce deeper collective
understanding of both the problem and the
data at the center of any project. Knowing
how a data set is created and stored, how
often it changes, and its reliability are critical
details that can make or break the feasibility
of a data product.
For example, consider a new credit-
scoring algorithm more accurate than
previous methods but that relies on data
no longer sold by the credit bureau. Such
circumstances are common today given
that data sets are so diverse and subject
to frequent change. By incorporating
interdepartmental expertise in the early
stages of model development, companies
dramatically reduce the risk of pursuing
unanswerable questions and ensure data
scientists are focusing attention on the most
suitable data sets.
Model-building
After firming up the project’s definition and
completing a preliminary survey of the data,
analysts enter the model-building phase of
analytics lifecycle. The notion of “model” is
often obscure and can be difficult to define,
even for those well versed in data science
vocabulary.
A statistical model, in short, is an abstract
representation of some relationship between
variables in data. In other words, a model
describes how one or more random, or
independent, variables relate to one or
more other dependent variables. A simple
linear regression model might, for example,
describe the relationship between years of
education (X) and personal income (y).
But linear regression is far from the only
way to represent the relationships in data,
and identifying the right algorithms and
machine learning methods for your problem
is largely an exploratory exercise. Data
scientists apply knowledge of the business
and advanced research skills to identify
those algorithms and methods most likely
to be effective for solving a problem. Many
and perhaps most data science studies are
bound up with solving some combination of
clustering, regression, classification, and/or
ranking problems. And within each of these
categories are numerous algorithms that may
or may not be suitable for tackling a given
We need to
look into it

problem.
To that end, the model-building phase is
characterized by rigorous testing of different
algorithms and methods drawing from one or
more of these problem classes (i.e. clustering,
regression, classification, and ranking) with
the ultimate goal being to identify the “best”
way to model some underlying business
phenomenon. “Best,” importantly, will take
on a different meaning depending on the
problem, the data, and the situational
nuances tied to the project. For example, the
“best” way to model the quality of the Netflix
recommendation system is very different
from the “best” way to model the quality of a
credit-scoring algorithm.

Actionable data science &
applications in operations
When a data science project progresses
beyond the model-building phase, the core
question is how best to take advantage of the
insights produced. This is a critical junction
and one ultimately determines the practical
ROI your data science investment.
Extracting value from data is like any other
value chain. Companies expend resources to
convert raw material—in this case data—into
valuable products and services suitable for
the market.
As is the case with any value chain, a product
gains value as it progresses from one lifecycle
stage to the next. Therefore, the manner in
which activities in the chain are carried out
is important as it often impacts the system’s
value.
Consider the product recommendations
example again—our goal is to increase
average order size for shoppers on our
website by recommending other products
users will find relevant.
1. Refine the problem definition
2. Survey the raw material and evaluate
which data to include in the model
3. Rigorously test modeling techniques
4. Identify a winning modeling strategy for
implementation
5. Integrate recommendations into the
website to influence customers
Common sense indicates that progressing
through step four without achieving step five
falls short of the objective. But, sadly, this
is a common scenario among companies
developing data science capabilities. Similarly,
it is often the case that hypotheses are
disproved only after companies have invested
A data product provides actionable information without exposing
decision makers to the underlying data or analytics.
Examples include: Movie Recommendations, Weather Forecasts,
Stock Market Predictions, Production Process Improvements, Health
Diagnosis, Flu Trend Predictions, Targeted Advertising.
– Mark Herman, et al., Field Guide to Data Science

substantial time and effort engineering
large-scale analytics implementations for
models which later prove to be suboptimal
or entirely invalid.
Data Driven Products
Value Produced
REAL
WORLD
SCORING
REAL
WORLD
SCORING
STRATEGIC
INSIGHTSSTRATEGIC
INSIGHTS
CLEAN
DATA
CLEAN
DATA
GET RAW
DATA
GET RAW
DATA
Stages of the Analytics Project Life CycleStages of the Analytics Project Life Cycle
Expert data teams
Management
Customers & Front
Line Employees
Amount of Work
Why building data driven
products is hard
One issue is that the effort expended during
each lifecycle stage is not necessarily equal
to the value created. Cleaning data requires
considerable effort from data engineers,
data warehouse experts, and data scientists,
but produces no significant utility for the
customer directly.
To be clear, cleaning data is valuable and
often necessary. It is an activity that makes
modeling and application building easier and
in some cases possible. For the most part,
however, the early activities of a data science
project rarely produce noticeable differences
in the quality of your product at least as far
as the customer is concerned.
For this reason, successful data science
teams are often intensely focused on
concrete, non-technical business objectives.
By keeping an eye on success criteria
expressly dealing with an end-user or front-
line employee, teams are less likely to declare
a project completed prematurely.
Another obstacle in data-driven product
building is rooted in tools. Data scientists
use incredibly different tools than those
used by application developers responsible

for building customer-facing apps. Although
both groups use code to solve problems,
the byproducts of data science are not
necessarily compatible with other systems
or vice versa. This is an important distinction
and one that is often underestimated or
overlooked by managers albeit inadvertently.
Scientific programming languages like R
and Python are ideal for cleaning, exploring,
and modeling data which explains the
overwhelming popularity of these tools
among data scientists. Likewise, application
developers gravitate towards tools designed
specifically for building scalable web and
mobile apps and creating compelling user
experiences—think .NET, Ruby on Rails,
Node.js, or JVM based web and mobile
frameworks.
Analysts use tools like R and Python to
perform advanced statistical procedures that
may take many tens or hundreds of lines of
code in another language. For example, R’s glm
function, a command that fits a generalized
linear model capable of predicting a True/
False outcome, has no native equivalent in
Java or Ruby. As a consequence, data science
prototypes built using this type of built-in
scientific functionality must be ported to
another environment via a notoriously difficult,
time consuming, and error-prone process in
order to be useful in production.
This operating paradigm is especially alarming
considering the inherent exploratory nature
of data science in the first place. Conceiving
hypotheses, designing tests and measuring
results are core to the data science job
function and primary reasons these
professionals are in such demand. Operational
practices that inhibit data scientists from
deploying and running experiments are, by
logical extension, counterproductive and
should be avoided.

A more efficient means of
shipping data products
The value companies get from data science is
governed largely by the pace at which teams
can conceive and test new ideas and bring
data-driven products to market. At Yhat, we
are re-envisioning how data scientists apply
their work to real world problems to help
companies move from proof of concept
to data-driven product faster and more
efficiently than ever before.
Through our cloud-based and on-premise
applications, analysts deploy and update
predictive routines using tools they know and
love without impacting application developers
or production systems. The platform is
designed not only to play-nice with the
leading tools and workflows but to naturally
extend them to provide a highly streamlined
process for moving from insight to prototype
to full-scale production application efficiently.
The platform includes native packages for R
and Python, a simple and elegant command
line interface for deploying, testing, and
managing models in production, and an
administration and reporting web app for
monitoring and tracking model performance
and usage. Models that would otherwise
need to be ported manually from analysts’
laptops into production are instead made
instantly accessible via several universally
understood and standards-compliant
protocols. Invoke models in real-time via
REST and low latency streaming APIs or score
data offline via on-demand or scheduled
batch jobs.
By reducing or, in some cases eliminating,
the handoff from data scientist to engineer,
we help analysts package and ship their
insights as ready-to-use products and avoid
long implementation periods. Through Yhat,
you can quickly deploy analytical prototypes,
update and improve your recommendation
systems, make credit policies changes, try
different learning methods, or incorporate
new data sets into workflows at an unheard
of pace and without impacting your
development roadmap.
Kludge Valley
Timespenttodeployandtestmodels
Number of models to deploy and test
Typical Data Science Team ŷhat
ComplexityExplosion

Works Cited
Chandrasekaran, Swami. “Curriculum via Metro Map.” Pragmatic Perspectives. Swami
Chandrasekaran, 8 July 2013. Web. 12 Mar. 2014. <http://nirvacana.com/thoughts/wp-content/uploads/2013/07/
RoadToDataScientist1.png>. A pragmatic visual inspired by transportation metro maps to depict the learning path to
becoming a data scientist. Each area / domain is represented as a “metro line”, with the stations depicting the topics you
must learn / master / understand in a progressive fashion.
Harris, Jeanne G., et al., The Team Solution to the Data Scientist Shortage. Chicago: Accenture
Institute for High Performance, n.d. Web. 12 Mar. 2014. <http://www.accenture.com/SiteCollectionDocuments/PDF/
Accenture-Team-Solution-Data-Scientist-Shortage.pdf> Today’s data scientist shortage won’t go away soon, yet companies
need them more than ever to leverage the value of big data. The solution is to build teams of data scientists instead of
seeking soloists.
Herman, Mark, et al., comps. The Field Guide to Data Science. Strategy and Technology Consulting
Firm | Booz Allen Hamilton. Booz Allen Hamilton, n.d. Web. 12 Mar. 2014. <http://www.boozallen.com/media/file/The-Field-
Guide-to-Data-Science.pdf>. Our goal in creating the Field Guide to Data Science is to capture what we have learned and to
share it broadly. We want this effort to help drive forward the science and art of Data Science.
Mount, John, and Nina Zumel. Practical Data Science With R. 7th ed. Greenwich:
Manning Pubns, n.d. Web. 12 Mar. 2014. <http://www.manning.com/zumel/PracticalDataScienceR_MEAP_ch01.pdf>. This book
combines technical content with practical, down to earth, advice on how to practice the craft of data science.
Porter, Michael E. Competitive Strategy. New York: Free Press, 1998. Print.
Witten, Ian H., and Eibe Frank. Data Mining: Practical Machine Learning Tools and Techniques. N.p.:
n.p., 2005. The Morgan Kaufmann Series in Data Management Systems 2. Manning Publications Company. Web. 12 Mar.
2014. <http://home.ustc.edu.cn/~hshl05/addition/Practical Machine Learning Tools and Techniques.pdf>.

About
Yhat, Inc. provides the leading data science operations platform for turning analytics projects
into products. We improve data science operations by connecting people, processes, and
systems into easily managed documentation, applications, and reports.
Through Yhat, data scientists train, deploy and update models; change features and algorithms
in production without downtime; configure recurring jobs to score data and more. The
platform tracks how your predictive models perform over time to provide non-technical
stakeholders new levels of transparency into data science efficacy and ROI.
Contact Us
http://yhathq.com
info@yhathq.com
(917) 719-5959

Applied_Data_Science_Presented_by_Yhat

Recommended

More Related Content

What's hot (20)

Viewers also liked (15)

Similar to Applied_Data_Science_Presented_by_Yhat (20)

Applied_Data_Science_Presented_by_Yhat