DATA

PREPARATION FOR AUTOMATED MACHINE LEARNING

WHITE PAPER

DATA PREPARATION FOR

AUTOMATED MACHINE LEARNING
BY JEN UNDERWOOD

© 2017 Impact Analytix, LLC - All rights reserved. 1


 DATA PREPARATION FOR AUTOMATED MACHINE LEARNING

TABLE OF CONTENTS
Introduction ........................................................................................................................ 3
Art and Science of Data Preparation ...................................................................... 3
Automation for Faster Data Preparation ................................................................ 3
DataRobot for Machine Learning ........................................................................... 4

Where to Start ..................................................................................................................... 5
Machine Learning Lifecycle .................................................................................... 5
Plan for Data Collection .......................................................................................... 5
Avoid Overfitting and Underfitting ......................................................................... 7
Collect and Structure Data .................................................................................................. 8
Gather Data ............................................................................................................ 8
Beware of Bias ...................................................................................................... 11
Explore and Profile ............................................................................................................ 11
Understand Your Data .......................................................................................... 12
Detect Leakage ..................................................................................................... 12
Find and Reduce Errors ........................................................................................ 13
Improve Data Quality ........................................................................................................ 15
Engineer Features ............................................................................................................. 18
Conclusion ......................................................................................................................... 20
Recommended Next Steps ................................................................................... 20

JANUARY 2018 – WHITE PAPER COMMISSIONED BY DATAROBOT

© 2017 Impact Analytix, LLC - All rights reserved. 2


 DATA PREPARATION FOR AUTOMATED MACHINE LEARNING

INTRODUCTION
The beauty of the human mind in combination with automated machine learning
empowers amazing predictive insights that might never be found using manual
techniques. Since the quality of predictive output relies on the quality of input, proper
data preparation is a critical success factor for achieving optimal machine learning results.

THE ART AND SCIENCE OF DATA PREPARATION

The iterative process of preparing data for automated machine learning is both an art and
a science. The art of data preparation requires knowledge of the business – or “domain
expertise” – to select the right problems to solve, identify crucial input data, and carefully
transform and engineer informative features to maximize predictive model accuracy. The
science of data preparation involves cleansing and normalizing collected data, selecting
influencer features, and generating training, testing, and validation data sets for
automated machine learning.
Although automated machine learning solutions may provide safeguards to prevent
common mistakes, you’ll still want to learn how to correctly prepare, shape, and format
your data to create great models. Providing the wrong data, irrelevant data, or
improperly prepared data undermines model performance for generalization.
Essentially you will want to design an input dataset with feature independence,
“You can expedite
explainable variance, and maximum information gain to find signals in the
the highly iterative
noise.
predictive data
preparation process with AUTOMATION FOR FASTER DATA PREPARATION
automated machine
To identify relationships in data — “the signals”— and isolate distracting,
learning.”
irrelevant data — “the noise” — you can expedite the highly iterative predictive
data preparation process with automated machine learning. In minutes, you can
find the most relevant features and pinpoint specific areas in your dataset where
prediction errors occur to help you focus efforts on the right data and reduce
experimentation time.
After running basic input data and evaluating the results, you will enhance input data, add
features, build another model, and review performance once again. You’ll continue this
process until your model meets performance objectives.
Ideally subject matter experts that understand the business process and data source
nuances will assist in the data preparation process. Depending on your project, data
preparation might be a one-time activity or a periodic one. As new insights are revealed,
it is common to experiment further.

© 2017 Impact Analytix, LLC - All rights reserved. 3


 DATA PREPARATION FOR AUTOMATED MACHINE LEARNING

DATAROBOT FOR MACHINE LEARNING

DataRobot is the world’s most advanced automated machine learning
platform. DataRobot automates the machine learning process from
data ingestion to deployment. It delivers immediate value and
unmatched ease-of-use, and no complicated math or scripting is
“Domain knowledge and
required.
best practices
DataRobot includes an array of data preparation features, used by the world’s
automating feature engineering to find key insights and hidden leading data scientists
patterns. This invaluable technology expedites analytical have been uniquely
investigation across millions of variable combinations that baked into DataRobot
would be far too time-consuming for manual human blueprints.”
exploration.
For optimal machine learning model performance, domain
knowledge and best practices used by the world’s leading data
scientists have been uniquely baked into DataRobot blueprints. Users of all
skill levels can safely apply machine learning with its built-in optimizations and
safeguards.
DataRobot supports popular advanced machine learning techniques and open source
tools such as Apache Spark, H2O, Scala, Python, R, and TensorFlow. Using drag-and-drop,
point-and-click guided menu options, DataRobot users can simply and quickly create
predictive models with automated machine learning. The process is simple:

o Ingest data sources

o Select a target variable to predict
o Automatically generate features, extract balanced samples, build and iterate
through 100s of machine learning models
o Visually explore top performing models and key findings
o Easily deploy and operationalize models
Machine learning development steps that used to take weeks or months of effort can now
be completed in hours. By embedding DataRobot automated machine learning model
intelligence into your reporting or business processes, you can quickly close the loop
between insight and action.
Since each data set and business objective can be unique with varied challenges, we have
provided the following guidelines to help get you started. We also share essential tips and
additional resources for further study.

© 2017 Impact Analytix, LLC - All rights reserved. 4


 DATA PREPARATION FOR AUTOMATED MACHINE LEARNING

WHERE TO START
The machine learning process begins with Business Understanding. This initial step
focuses on defining the right problem to solve and recognizing the business objectives
and requirements. After selecting a problem, you will collect and assess data. During the
Data Understanding step, you will get familiar with available data sources, identify data
quality problems, and perform exploratory analysis. Then, in the Data Preparation step,
you will cleanse the data, shaping and transforming it into a flattened format for loading
into the automated machine learning platform.

MACHINE LEARNING LIFECYCLE

Figure 1: Overview of the Machine Learning Process

For the purposes of this white paper, we will concentrate on collecting data and preparing
it properly. We will not cover the entire machine learning lifecycle.
Before you begin the data collection and data preparation process, it is assumed that you
already have selected, defined, and isolated a business problem to solve that is a viable
candidate for machine learning. You should also have chosen at least one metric that you
want to better understand. If you need more information on those steps in the machine
learning process, please refer to our previous white paper, Moving from BI to Machine
Learning.

PLAN FOR DATA COLLECTION

As you develop requirements for machine learning model data collection, contemplate
the business process and review it from different perspectives. Consider what happens at
each step, what data is captured, where it is stored, if data history or changes are
retained, and if that data truly reflects the real world for resulting predictions. Often data
is collected in line-of-business applications or data warehouses for other reasons. Existing

© 2017 Impact Analytix, LLC - All rights reserved. 5


 DATA PREPARATION FOR AUTOMATED MACHINE LEARNING

data may be missing situational context such as location, environmental conditions, and
other relevant variables for predicting an outcome. Document known issues and
preferred data that could be added in the future.

TO IMPROVE INPUT DATA COLLECTION, DIAGRAM THE BUSINESS PROCESS

FLOW. IDENTIFY THE STEPS, ENVIRONMENTAL CONDITIONS, SCENARIOS,
PEOPLE, SYSTEMS, AVAILABLE DATA, AND MISSING DATA.

As you continue planning to gather data for your machine learning modeling project,
you’ll need to confirm decision-level metric granularity. Granularity refers to a unit of
analysis. A unit might be an opportunity, customer, or transaction. Granularity is
determined by the business objectives and how your model will be used
operationally. Ask stakeholders how decisions will be made from the predictive
“To illustrate these models. Are they based on a single customer, transaction, or event, or are they
concepts, we will be based on aggregate data over time?
referring to a publicly
available Bank Marketing To illustrate these concepts, we will be referring to a publicly available Bank
Data Set1 from UCI’s Marketing Data Set1 from UCI’s machine learning repository. The sample data
machine learning set contains partially prepared data to predict client term deposits collected
repository.” during the bank’s telemarketing campaigns.
In the Bank Marketing Data Set, the desired outcome to predict is client term
deposits. This is a binary yes or no outcome in the sample, but it could have
alternatively been a total amount figure to maximize deposits. Don’t always limit yourself
to collecting one outcome variable while assembling data. Think about other questions
that might be asked and data that would make sense to include.
Potential influencer features for the example client term deposit outcome include client
demographics such as age, job, marital status, and education. Past credit and loan
repayment information is also important to know. Other features chosen included
campaign contacts, previous marketing campaign outcomes, and several external social
and economic environmental attributes such as employment rate.

NOTE: WHEN SELECTING FEATURES, YOU WILL WANT TO EXTRACT THE

MAXIMUM INFORMATION FROM THE MINIMUM NUMBER OF INPUT VARIABLES.

1
UCI machine learning repository data set https://archive.ics.uci.edu/ml/datasets/bank+marketing

© 2017 Impact Analytix, LLC - All rights reserved. 6


 DATA PREPARATION FOR AUTOMATED MACHINE LEARNING

AVOID OVERFITTING AND UNDERFITTING

Overfitting and underfitting are common mistakes for beginners who are preparing
machine learning modeling data. Overfitting captures the noise in your data with an
overly complex, unreliable predictive model. Essentially what happens is the model
memorizes unnecessary details. When new data comes in, the model fails.
If you don’t have enough features, your model might be oversimplified and suffer from
underfitting issues. Underfitting occurs when a statistical algorithm cannot capture the
underlying patterns in the data.
Why are overfitting and underfitting problematic? The machine learning model will have
too many prediction errors to be useful for decision-making.

Figure 2: Common Data Preparation Issues

For categorical data, overfitting can occur if a high number of categories are observed
with a small number of observations per category. These types of variables hold less
information for predictive value. For time series data, overly complex mathematical
functions that describe the relationship between the input variable and the target
variable can also lead to overfitting. In the most extreme form of overfitting, individual
identifiers are inadvertently used as machine learning inputs. Individual identifiers can
perfectly model existing data, but would only by chance reliably model and predict
outcomes for other data.
Thus, there is a delicate balance between being too specific with too many features and
too vague with not enough features. Designing machine learning model features with just
the right amount of predictive information gain and precision is a key skill in the art of
data preparation.

© 2017 Impact Analytix, LLC - All rights reserved. 7


 DATA PREPARATION FOR AUTOMATED MACHINE LEARNING

COLLECT AND STRUCTURE DATA

After reviewing the business process and planning the machine learning input data
requirements, you’ll delve into the data collection and shaping process.

GATHER DATA
Machine learning algorithms assume that each record is independent and are not related
to other records. If relationships exist between records, you will want to create a new
variable called a feature in a column within the row of data to capture that behavior.
Unlike third-normal form transactional or dimensional patterns used in business
intelligence, machine learning requires data to be input as a “flattened” table, view, or
comma separated (.csv) flat file of rows and columns. Your view will need to contain an
outcome metric and target variable, along with input predictor variables. This data
representation for machine learning is called the feature matrix.

Figure 3: Example Prepared Data Set

If you have data stored in several tables in a data warehouse or relational database
format, you will need to use record identifiers to join fields from multiple tables to create
a single unified, flattened “view.” For many target variables, input data is captured at
various business process steps in multiple data sources. A sales process might have data
in a CRM, email marketing program, Excel spreadsheet, and/or accounting system. If that
is the case, you will want to identify the fields in those systems that can relate, join, or
blend the different data sources together.
Prepared data should be collected at a level of analytical granularity upon which you can
make decisions. Choose a granularity that is actionable, understandable, and useful in the

© 2017 Impact Analytix, LLC - All rights reserved. 8


 DATA PREPARATION FOR AUTOMATED MACHINE LEARNING

event you incorporate the results into your existing business process or application. For
example, if you want to make daily sales forecasts, you need to input data at a day level
rather than week, month, or year.
If you are trying to capture changes in data over a certain time period, check if
“Shaping data involves your data source is only keeping the current state values of a record. Most data
subject matter expert warehouses are designed to save different values of a record over time and do
thought to creatively not overwrite historical data values with current data values. Transactional
select, create, and application data sources such as Salesforce only contain the current state value
transform variables for for a record. If you want to get a prior value, you need to have a snapshot of
maximum influence.” the historical data stored or keep the prior value data in custom fields on the
current record.
While structuring input data, ensure that it is clean and consistent. The order and
meaning of input variables should remain the same from record to record. Inconsistent
data formats, “dirty data,” and outliers can undermine the quality of analytical findings.

HOW MUCH DATA TO COLLECT

The actual number of records is not always easy to determine and depends on patterns
in your data. If you have more noise in your data, you will need more data to overcome
it. Noise in this context means unobserved relationships in the data that are not captured
by the input predictor variables.

Figure 4: Sample Size Estimation

To determine minimum data set sizes, consider the dimensionality of your data and
pattern complexity.2
o For small models with a few variables, 10 to 20 records per variable value may be
sufficient.
o For more complex models, ~100 records per variable value may be needed to
capture patterns.
o For complex models with ~100 input variables, you will need a minimum of
10,000 records in the data for each subset (training, testing, and validation).

TRAINING, TESTING, AND VALIDATION DATA SETS
The most common strategy is to split data into training, testing, and validation data sets.

Applied Predictive Analytics: Principles and Techniques for the Professional Data Analyst by Dean Abbott
2

© 2017 Impact Analytix, LLC - All rights reserved. 9


 DATA PREPARATION FOR AUTOMATED MACHINE LEARNING

These data sets are usually created through a random selection of records without
replacement, meaning each record belongs to one and only one subset. All three data
sets should reflect your real-world scenario.
Cross Validation
Keep in mind that DataRobot includes industry standard k-fold cross validation features
that divide the data into k subsets with the holdout repeated k times. Each time, one of
the k subsets is used as the test set and the other k-1 subsets are combined to form a
training set. DataRobot’s k-fold feature enables you to independently choose how much
data you want to use in testing.

Time Series Considerations

Data that changes over time should be reflected in your input data set. When time
sequences (Contact Made > Quote Provided > Deal Closed) are important in predictions,
proportionally collecting data from those different time periods is important as well. The
key principle is to provide data that reflects what actually happens at the right level of
outcome metric granularity.
When collecting data, think about
the balance of your values in your
raw data. In our example, how
many prospects do you have at
different ages and stages of life?
Do they have loans? What are the
balances on outstanding loans?
How many defaulted on a loan?
How recent was the default,? What
is the prospect’s income history,
credit ratings, debt-to-income
ratio, and so on?
Figure 5: Sampling Data
Think Proportionally
When extracting a subset of data, be sure to include approximately the same proportion
of variables in your DataRobot input data set that you see in the real-world data. If you
provide more records of one variable value, you can accidentally bias the machine
learning model’s predictions, thereby diminishing performance. If you have data sets with
millions or billions of rows, it is much less likely that you will encounter accidental bias in
data preparation.

© 2017 Impact Analytix, LLC - All rights reserved. 10


 DATA PREPARATION FOR AUTOMATED MACHINE LEARNING

SAMPLING
The choice of the optimal sampling method3 for a given problem especially depends on
the character of the dataset and the desired proportion of the subsets. Each method has
its advantages but also its limitations.
For simple, nearly uniformly distributed datasets, the method of simple random sampling
may be sufficient. For naturally well-ordered time series data, highly efficient
deterministic approaches such as convenience and systematic sampling can achieve
reliable results. When dealing with complex, high-dimensional data, more sophisticated
and stratified sampling techniques can reduce the bias and variance of model error.
Unbalanced Two-Class Problems
Data sets seldom come with evenly distributed samples. Unbalanced data is a common
issue to remediate. Fraud or failure rate data are examples of unbalanced two-class
problems. Analyzing unbalanced data creates useless results with exceptionally high error
rates.
To build predictive models on unbalanced data, you need to apply sampling techniques
that increase the minority class proportion with downsampling or upsampling to create a
balanced data set. After training a machine learning model with a balanced training set,
you will validate performance of the model with the real-world, unbalanced, unseen test
set.

BEWARE OF BIAS
While you accumulate data, consider potential biases.4 Human nature is consciously and
unconsciously biased. Cognitive biases are tendencies to think in certain ways that can
lead to irrational judgment. Outcome, omission, and many other bias types can easily
creep into the data collection process.
If unknown bias exists, it is basically an unjustified assumption that your input data
reflects reality. Any model built on such assumptions reflects only the distorted reality
and will perform poorly. To reduce potential bias, test hypotheses, poke holes in your
own ideas, welcome challenges, and conduct peer reviews of your data collection and
sampling thought processes. Machine learning projects should be group projects and not
done in isolation.

EXPLORE AND PROFILE

Now you will assess the condition of your source data. DataRobot automates several
aspects of initial data examination. DataRobot also automates sampling to avoid
conventional sampling and overfitting issues. During this step, you’ll visually look for

3
Common Types of Data Sampling Methods https://en.wikipedia.org/wiki/Sampling
4
The Cognitive Bias Codex https://upload.wikimedia.org/wikipedia/commons/a/a4/The_Cognitive_Bias_Codex_-
_180%2B_biases%2C_designed_by_John_Manoogian_III_%28jm3%29.png

© 2017 Impact Analytix, LLC - All rights reserved. 11


 DATA PREPARATION FOR AUTOMATED MACHINE LEARNING

trends, extreme values, outliers, exceptions, skewed data, incorrect values, and
inconsistent and missing data.

UNDERSTAND YOUR DATA

As you begin to explore and understand your data, DataRobot provides data profiles for
every feature that include how many values are unique or missing, as well as the
statistical mean, standard deviation, median, minimum, and maximum value. You
can also review the distributions of each feature using a histogram with optional
“By looking at these
bin settings and apply transformations to normalize your data.
findings, the business can
immediately appreciate Informative Features immediately ranks variables that provide the most
the variables that truly information gain for building optimal machine learning models. Knowing which
influence outcomes to areas of your data most influence the outcome is invaluable on its own. This
get immediate value.” information can guide the business to focus limited time and resources on the
activities that matter most.
Another one of DataRobot’s data preparation strengths is the array of different
visualization tools that identify influential features, rank them, and uncover errors. The
Feature Ranking report measures how much each feature by itself contributes to the
accuracy of a machine learning model.

Figure 6: Feature Impact

DETECT LEAKAGE
Leakage is the accidental inclusion of outcome information that would not be legitimately
available for predictions. If you have inadvertent leakage, you’ll likely notice it in the
feature ranking report when a feature has an exceptionally high impact.

© 2017 Impact Analytix, LLC - All rights reserved. 12


 DATA PREPARATION FOR AUTOMATED MACHINE LEARNING

Figure 7: Leakage Example

In our Bank Marketing Data Set example, duration was identified in the DataRobot
Feature Impact report as a leaked feature with 100% impact. The next best performing
feature carries a little more than 10% impact. If you see an exceptionally high impact,
verify process flow timing for that feature.

TO AVOID LEAKAGE, YOU NEED TO CONSIDER THE TIMING AND ORDER OF

EVENTS TO ENSURE THAT OUTCOME DATA IS NOT USED AS INPUT DATA.

FIND AND REDUCE ERRORS

DataRobot Model X-Ray and Reason Codes capabilities can also provide deep insights for
enhancing data preparation. Model X-Ray enables interactive, visual exploration of
machine learning model performance. You can easily see where a model makes mistakes
by selecting input features. It shines a light on issues that might not get detected using
other tools. Model X-Ray allows machine learning model designers to concentrate on
where the most model performance improvements can be made in the data preparation
process.

© 2017 Impact Analytix, LLC - All rights reserved. 13


 DATA PREPARATION FOR AUTOMATED MACHINE LEARNING

Figure 8: Model X-Ray

Reason Codes unveil what values within a feature drive the model’s results. This tool is
crucial for machine learning model development processes that are subject to regulatory
compliance or legal scrutiny. With Reason Codes, you can discover which combinations
of feature values trigger a specific machine learning outcome. This information can also
be useful throughout the iterative data preparation process to incrementally improve
results.

Figure 9 Reason Codes

© 2017 Impact Analytix, LLC - All rights reserved. 14


 DATA PREPARATION FOR AUTOMATED MACHINE LEARNING

IMPROVE DATA QUALITY

We recommend that you address data quality issues as early as possible. Here are a few
tips for handling common data issues in the data preparation process.
Correcting Incorrect Values

Machine learning models assume the input data is correct. If you are seeing errors from
source applications that should get fixed, a best practice is to try and resolve the issue at
the source system versus in a data preparation process.
Treat incorrect values as missing if there is a minimal amount and you can’t easily
determine correct values. If there are a lot of inaccurate values, try to determine what
happened to repair them. If you do make changes to data, document your reasoning. Also
capture initial context and changed values with a flag to identify changes. The pattern in
your data might be hidden in those incorrect values.
Skewed Variables
For continuous variables, review the distributions, central tendency, and variable spread
in DataRobot. These are measured using various statistical metrics and visualization
methods such as histograms. Continuous variables should be normally distributed. If not,
reduce skewness with transformations or by experimenting with bin sizes for optimal
prediction.
When a skewed distribution needs to be corrected, the variable is transformed by a
function that has a disproportionate effect on the tails of the distribution. Log transforms
like log(x), logn(x), log10(x); the multiplicative inverse (1/x); square root transform sqrt(x);
or power (xn) are the most frequently used corrective functions.
In the table to the left,
several issues and formulas
to minimize skew are shown.
The before and after charts
illustrate how different
skewed variable
transformations can be used
to normalize feature
distributions.

Figure 10: Transformations for Skewness

© 2017 Impact Analytix, LLC - All rights reserved. 15


 DATA PREPARATION FOR AUTOMATED MACHINE LEARNING

High-Cardinality

High-cardinality fields are categorical attributes that contain a very large number of
distinct values. Examples include names, ZIP codes, and account numbers. Although these
variables can be highly informative, high-cardinality attributes are rarely used in
predictive modeling. The main reason is that including them will vastly increase the
dimensionality of the data set, making it difficult or even impossible for most algorithms
to build accurate predictive models.

Redundant, Highly Correlated Variables

Duplicate, redundant, or other highly correlated variables that carry the same information
should be minimized. DataRobot algorithms will perform better without collinear
variables. Collinearity occurs when two or more predictor variables are highly correlated,
meaning that one can be linearly predicted from the others with a substantial degree of
accuracy.
To identify high correlation between two continuous variables, review scatter plots. The
pattern of a scatter plot indicates the relationship between variables. The relationship
can be linear or non-linear. To find the strength of the relationship, compute correlation.
Correlation varies between -1 and +1.
If you have two variables that
are almost identical and you
do want to retain the
difference between them,
consider creating a ratio
variable as a feature. Another
approach is to use Principal
Component Analysis (PCA)
output as input variables.
To avoid the collinearity issue,
do not include multiple
variables that are highly
correlated or data that is from
the same reporting hierarchy.
Often those fields provide Figure 11: Scatter plots for correlation detection
obvious insights. For example,
customers who live in the city
of Tampa also happen to live in the state of Florida.

© 2017 Impact Analytix, LLC - All rights reserved. 16


 DATA PREPARATION FOR AUTOMATED MACHINE LEARNING

Missing values

The most common repair for missing values is imputing a likely or expected value using a
mean or computed value from a distribution. If you use the mean, you may be reducing
your standard deviation thus the distribution imputation approach is more reliable.
Another approach is to remove any record with missing values. Don’t get too ambitious
with filtering out missing values. If you delete too many records, you will undermine the
real-world aspects in your analysis. As you address missing values, do not lose the initial
missing value context. A common data preparation approach is to add a column to the
row to flag data was missing coded with a 1 or 0.
Extreme Values and Outliers
Outliers are values that exceed three standard deviations from the mean. Many machine
learning algorithms are sensitive to outliers since those values affect averages (means)
and standard deviations in statistical significance calculations. If you come across unusual
values or outliers, confirm that these data points are relevant and real. Often, odd values
are errors.
If the extreme data points are accurate, predictable, and something you can count on
happening again, do not remove them unless those points are unimportant. You can
reduce outlier influence by using log transformations or converting the numeric variable
to a categorical value with binning.

© 2017 Impact Analytix, LLC - All rights reserved. 17


 DATA PREPARATION FOR AUTOMATED MACHINE LEARNING

ENGINEER FEATURES
Feature creation is the art of extracting more information from existing data to improve
the predictive power of machine learning algorithms. You are making the data you already
have more useful. Strong features that precisely describe the process being predicted can
improve pattern detection and enable more actionable insights to be found.
Creating features from several combined variables and ratios usually provides
higher model accuracy than any single-variable transformation because of the
“Feature engineering is information gain associated with data interactions. If you ever saw the movie or
often the determining read the book “Moneyball: The Art of Winning an Unfair Game” by Michael
factor in whether a Lewis, you’ll know how baseball analysis was revolutionized with new
machine learning performance metrics like On-Base Percentage (OBP) and Slugging Percentage
modeling project is (SLG). With feature engineering, you will be using a fundamentally similar
successful or not.” approach.

HUMAN INGENUITY AND CREATIVITY REQUIRED

Feature engineering is challenging because it depends on leveraging human intuition
to interpret implicit signals in data sets that machine learning algorithms use.
Consequently, feature engineering is often the determining factor in whether a machine
learning modeling project is successful or not. This step in the process is experimental and
usually the bottleneck in automated machine learning processes.
Although collected raw data fields can be used as-is in DataRobot without
transformations, supplemental data, or calculations to train machine learning models,
you’ll almost always want to add more perspective to your data set by designing features.
Engineered features provide better context to differentiate patterns in the data.

Aggregations
Some commonly computed aggregate features including the mean (average), most
recent, minimum, maximum, sum, multiplying two variables together, and ratios made
by dividing one variable by another. Note DataRobot automatically generates date and
time aggregation features.
Ratios
Ratios can be excellent feature variables. Ratios can communicate more complex
concepts such as price-to-earnings ratio, where neither price nor earnings alone can
deliver this insight.
Transformations
Transformation refers to the replacement of a variable by a function. For instance,
replacing a variable x by the square or cube root or logarithm x is a transformation. You
transform variables when you want to change the scale of a variable or standardize the
values of a variable for better understanding. Variable transformation can also be done

© 2017 Impact Analytix, LLC - All rights reserved. 18


 DATA PREPARATION FOR AUTOMATED MACHINE LEARNING

using categories or bins to create new variables. An example transformation might be

binning continuous Lead Age into Lead Age Groups or Loan Amount into Loan Amount
Categories.

FEATURE ENGINEERING TECHNIQUES

Here are a few popular feature engineering ideas shared on Data Science Central5 that
can be used to extract more information from your input data:

1. Single variable transformations
2. Ratio or frequency of categorical variables
3. Combine important variables
4. Compute variable interactions
5. Change data types
6. Compute relative differences
7. Cartesian transformations
8. Bin transformations
9. Window time series data
10. Reframe continuous variables
11. One hot encoding for sequence problems
12. Sparse value coding

The possibilities for feature engineering are limited only by your own human ingenuity
and creativity. Feature engineering truly is the human art of data preparation for
automated machine learning.

5
Data Science Central https://www.datasciencecentral.com/profiles/blogs/feature-engineering-data-scientist-s-secret-
sauce-1

© 2017 Impact Analytix, LLC - All rights reserved. 19


 DATA PREPARATION FOR AUTOMATED MACHINE LEARNING

CONCLUSION
In this white paper, we briefly introduced basic data preparation for machine learning
concepts. We discussed how to plan your project and collect, organize, structure, and
shape data in a machine learning-friendly format. We also bestowed vital tips for feature
engineering to help you master the art of data preparation for automated machine
learning models.
As you progress in your automated machine learning model journey, each area of the
data preparation process should be further researched. For additional reading, the
following books are highly recommended:

o Data Preparation for Data Mining by Dorian Pyle

o Data Preprocessing in Data Mining by Salvador García and Julián Luengo
o Applied Predictive Analytics: Principles and Techniques for the Professional Data
Analyst by Dean Abbott
o Feature Engineering for Machine Learning Models: Principles and Techniques for
Data Scientists by Alice Zheng and Amanda Casari
Note that DataRobot also provides classes that cover these and other related topics.

RECOMMENDED NEXT STEPS

For additional information on automated machine learning, please contact an expert at
DataRobot. It’s easy to get started.
o DataRobot
www.datarobot.com
o Data Preparation Essentials for Automated Machine Learning webinar recording
www.datarobot.com/webinar/data-prep/
o DataRobot AI Acceleration Packages
www.datarobot.com/product/enterprise-ai/
o DataRobot Courses
www.datarobot.com/education/all-courses/

© 2017 Impact Analytix, LLC - All rights reserved. 20


 DATA PREPARATION FOR AUTOMATED MACHINE LEARNING

About DataRobot
DataRobot offers an enterprise machine learning platform that empowers users of all skill levels to make better
predictions faster. Incorporating a library of hundreds of the most powerful open source machine learning algorithms,
the DataRobot platform automates, trains and evaluates predictive models in parallel, delivering more accurate
predictions at scale. DataRobot provides the fastest path to data science success for organizations of all sizes. For
more information, visit www.datarobot.com.

About the Author
Jen Underwood, founder of Impact Analytix, LLC, is an analytics industry expert with a unique blend of product
management, design and over 20 years of “hands-on” advanced analytics development. In addition to keeping a
pulse on industry trends, she enjoys digging into oceans of data. Jen is honored to be an IBM Analytics Insider, SAS
contributor, and former Tableau Zen Master. She also writes for InformationWeek, O’Reilly Media, and other tech
industry publications.

Jen has a Bachelor of Business Administration – Marketing, Cum Laude from the University of Wisconsin, Milwaukee
and a post-graduate certificate in Computer Science – Data Mining from the University of California, San Diego. For
more information, visit www.jenunderwood.com.

© 2017 Impact Analytix, LLC - All rights reserved. 21