Machine learning is hot right now and for good reason. We're going to break down what you need to know about what goes into a model and give you four machine learning models your business should have in production right now.
The 4 Machine Learning Models Imperative for Business Transformation
1. The Four Machine Learning Models Imperative
for Business Transformation
2. There isn’t a one-size-fits-all approach to
building a machine learning model. The
way you code, deploy and assimilate the
model into your organization is complex.
Putting the proper frameworks in place
will help ensure you’re able to organically
uncover insights that will move the needle
in your business. That starts with laying
the groundwork for the production of your
models.
Building Effective Machine Learning Models.
EXPERIMENT
Clear Need
Hypothesis
EXPLORE
Research
Observe
Prioritize
Insight
Revise
Hypothesis
Brainstorm
Concepts
DIRECTIONCONCEPT
PAIN POINT RESOLVEPAIN POINT
3. When it comes to leveraging the assets at
your disposal and formulating a strategy
centered around advanced analytics,
where do you start? With the 3 Ps —
people, processes and platforms.
Training your team to put new platforms
to use and adopt new processes is critical.
As you make your people more aware of
the possibilities from predictive machine
learning, you’re in a better position to
deliver a beautiful end-to-end experience.
Bringing the 3 Ps Into Machine Learning.
Learn & Predict
People
Explore & Labels
Process
Collect & Store
Platforms
MACHINE LEARNING HIERARCHY OF NEEDS
4. Rules Based (SQL) vs Machine Learning.
Standard Query Language (SQL) is a basic, rules-based language used to
communicate with databases.
Benefits of SQL
● Results are easily explainable and
decipherable.
● Disseminating the knowledge from
SQL is straightforward.
● Gaining adoption is easier and
faster because key players already
possess a level of context around
the findings.
Limitations of SQL
● Dirty data could dramatically skew
results and rob the model of accuracy.
● The models do not evolve with time
because they are typically manual and
fixed.
● Reporting is historical, making it
harder to predict future outcomes.
5. In their simplest form, both statistics and
machine learning offer deeper insight
around a given data set using pattern
recognition, outlier identification,
exception-based modeling and more.
These approaches differ in that one
emphasizes discipline (statistics) and the
other emphasizes momentum and
dynamic perpetual improvement.
Statistics vs. Machine Learning.
Databases
Statistics
Neurocomputing
Machine
Learning
STATISTICAL MODELING VS. MACHINE LEARNING
Pattern
Recognition
AI
6. DEEP LEARNING
Deep learning is required to leverage non-linear tasks. This learning style offers
more potential to deepen a code base by allowing the model to separate a data
set into various features rather than relying on human input.
Deep Learning vs. Machine Learning.
MACHINE LEARNING
Input Feature Extraction + ClassificationInput OutputFeature Extraction Classification Output
Cat
Not Cat Not Cat
Cat
VS.
Input
7. Deep Learning vs. Machine Learning.
Accuracy
Involvement of
Domain Expert
Draw Set That
Can Be Analyzed
Correlations
Moderate with a false positive
rate at up to 5%
Extremely high with false
positive rate at nearly zero
Not requiredRequired for feature engineering &
extraction
Only 2.5-5% of available data
Only simple linear correlations
Non-linear correlations i.e.
correlations that exist in complex
patterns, rather than simple 1-1
correlations
Process 100% of available raw
data
8. Manual workflows using algorithms are no longer efficient. Modern machine
learning allows for minimal latency and impressive accuracy on models due to
the automated aspects of many of the platforms available.
The Vast AI and Machine Learning Landscape.
9. Machine learning platforms empower a
larger user base to configure, deploy,
adopt and maintain a given model or set
of models.
There are many platforms out there.
Choosing the best one for your
organization starts with running a cost-
benefit analysis using a critical set of
parameters and requirements as it
relates to your specific organization.
Choosing the Best Platform for You.
SELECTING THE PLATFORM FOR YOUR
ORGANIZATION’S NEEDS
MACHINE LEARNING PROVIDERS
SCORE
CAPABILITIES
Open Source
Available?
Usability (non
technical user)
Integrations
Scalability
Data Exploration
and Prep
Cost
10. Finding the Right Learning Structure.
● Supervised: Maps an input variable
to an output variable to make
predictions based on a given data set.
● Unsupervised: Does not contain an
output variable. Learns about the
data set via pattern recognition.
● Reinforced (or Semi-Supervised): A
combination of unsupervised and
supervised learning where an
algorithm is used to formulate
conclusions.
DIFFERENT TYPES OF MACHINE LEARNING
11. Framing and Problem Definition.
Knowing where to start with machine learning can be hard. Asking
questions up front can help steer you in the right direction.
● How do you plan to identify the
target metric for which you want to
predict?
● How do you define the limitations of
the target metric?
● What level of context and data
breadth are you able to include?
● Is there any third-party data you can
leverage to add incremental value to
the model?
● How can we assimilate insights (both
from the development and
deployment) of our model into the
appropriate communication lanes,
processes and organizational
materials/literature so we’re allowing
for digital economies of scale?
12. Every organization has a level of
analytical maturity. Knowing where
your organization falls can help you
start to understand what’s
available to you and identify any
areas where you can improve in
order to drive more business value
from your data.
Having an understanding of where
you are currently is vital for laying
out short- and long-term strategies.
Analytics Maturity and Self-Awareness.
DATA SCIENCE’S ROLE IN BUSINESS TRANSFORMATION
REPORTING
ANALYSIS
MONITORING
FORECASTING
PREDICTIVE
PRESCRIPTIVE
PREDICTIVE STATISTICS
MACHINE LEARNING
DESCRIPTIVE STATISTICS Data
Insights
Advanced
Analytics
BI&ANALYTICSCOMPLEXITY
BUSINESS VALUE
13. A framework turns a fuzzy concept into a concrete set of rules that will help your
team take the predictions of the machine learning model and turn it into action
steps which will drive the business forward.
A Framework for Deployment.
Resilient Trusted Prevalent Measurable AdvancementAdvancement
PeoplePlatforms Process
14. ● Validate Use Case
● Data Finalization
● Explore and Diagnose
● Cleanse
● Develop
● Features
● Build
● Infer
● Publish
● Deploy
● Consume
Successfully Deploying Machine Learning Models.
15. To know if the model is giving accurate
results, you must analyze your predictions
using one or more of the following metrics:
● Confusion or Error Matrix
● Accuracy
● Recall or Sensitivity to TPR (True
Positive Rate)
● Precision
● Specificity or TNR (True Negative
Rate)
● F1 Score
Using Fit Statistics to Validate Machine Learning.
16. A regression model estimated a function
(f) from input variables (x) into continuous
output variables (y), as a range of values.
Analyze whether regression models are
accurate by looking at the baseline
output of the model; using a variety of
statistics such as:
● Mean Squared Error (MSE)
● Root Mean Squared Error (RMSE)
● Mean Absolute Error (MAE)
● R Squared (R2)
● Adjusted R Squared (R2)
Fit Statistics for Regression Models.
17. Local Interpretable Model-
Agnostic Explanations (LIME) is a
record explainer mechanism — an
important technique to leverage
when filtering through the
predicted outcomes from any
machine learning model. This
technique is powerful and fair
because it focuses more on the
inputs and outputs from the
model, rather than the model itself.
Using LIME to Understand a Model’s Predictions.
Models
Model finds a
customer has an
89% propensity
to churn
LIME pulls out the
various datasets
And predictions...
...and makes small
tweaks to the inputs
Dataset &
Predictions
Pick Step Explanations Human Makes
a Decision
LIME then generates
explanations about why a
prediction was made and a
variable’s impaction the
outcome.
A human can then
make an informed
decision about what
to do with the model’s
findings
18. Digital economies of scale means the more you leverage data and analytics
across an organization, the more complete, robust, accurate, usable and
valuable they become. As your informational ecosystem becomes more valuable,
the cost of enacting a given (subsequent) digital initiative decreases as previous
work is repurposed.
Digital Economies of Scale via Model Chaining.
SUCCESSIVE MODELING
19. Requisites for Operationalizing Machine Learning.
There’s a lot that goes in
the backend of creating a
machine learning
predictive model, but all
of those efforts are for
naught if you don’t
operationalize your model
effectively with proper
amount of forethought,
scoping, preparation,
building, and inferring.
MATURITY
SCALE
OPERATIONALIZING MACHINE LEARNING PREDICTIVE MODELS
Intervention
Plan
Collaborate
& Infer
Adopt
& Apply
Pervasiveness
& Expansion
Attributable
Value
Change
Agent
Business
Application
ML
Positioning
Concept Drift
Controls
20. Although there’s a lot to the setup for any model, the key to building one
that works has less to do with the technicalities, and everything to do with
knowing what you want the model to solve.
The Four Data Models Businesses Need.
● Lead/Opportunity Conversions
● Attrition and Customer Retention
● Lifetime Value Model
● Employee Retention Model
21. What Do You Want Out of Machine Learning?
Organizations can benefit from machine
learning models across all stages of the
bow tie funnel.
● At the top, you’re looking at how to
convert your customers.
● In the middle, you’re aiming for
retention.
● Across the board, you’re hoping to
improve the lifetime value (LTV) of
your customer, and the internal
employee experience to keeps your
team on board and engaged.
22. The lead/opportunity conversions model
identifies why a consumer in the
engagement stage buys, and which
products or services they have a higher
propensity to buy.
This allows for a lower marketing spend
while simultaneously generating higher
return on marketing investment.
Lead/Opportunity Conversions Model.
LEAD/OPPORTUNITY CONVERSIONS MODEL
23. This model helps you identify the
emotional and logical triggers of
customers with a higher propensity to buy
to fuel a strategy of more personalized
marketing through predicting what the
buyer needs to hear, and when to
accomplish precision messaging.
50% of customers want personalized
offers and recommendations specific to
their needs.
Lead/Opportunity Conversions Model.
24. The attrition/customer retention model
can help you understand which
customers are most likely to churn out
and when.
By understanding the propensity to churn,
you’re better able to gauge
product/market fit and the health of your
overall organization against incoming
competitors.
Attrition/Customer Retention Model.
ATTRITION/CUSTOMER RETENTION MODEL
25. Old Logo:
Attrition/Customer Retention Model.
TRIANGLE PEST CONTROL CASE STUDY
New Logo:
Triangle Pest Control analyzed their
customer’s propensity to churn and
determined that their buyers had no
reason to remain brand loyal to their
organization. They took an empathetic
stance to improve the customer’s
experience and lower attrition rates by:
● Redesigning their logo
● Retraining their employees
26. As a result of their redesigned strategy, Triangle Pest Control started seeing the
benefits in the first year. As a result of their empathetic efforts and fast reaction
to their deeper customer understanding, they achieved:
Attrition/Customer Retention Model.
27. Lifetime Value Machine Learning Model.
LIFETIME VALUE MACHINE LEARNING MODEL
The Lifetime Value model looks across the
entire bow tie funnel to identify which
customers are likely to have a higher LTV,
and should be invested in early on.
Pattern recognition can be used to
leverage data and analytics to predict
where to make strategic changes. You
must train your machine learning models
to spot patterns against the backdrop of
the most important, empathetic set of
questions — the ones which stem around
your buyer’s WHY.
28. Lifetime Value Machine Learning Model.
When you ask why a customer does
something, not what they want, you’re
letting your machine learning model
focus on insights-centric questions
without faults in data or human bias
skewing the results. These questions
include:
● Why are customers buying from you?
● Why are they leaving?
● Why are they talking about you?
● Why is the market shifting?
29. Employee Retention Model.
The employee retention model predicts
which employees have a high propensity
to churn, potentially avoiding the typically
high costs associated with turnover.
Using annual reviews and exit interviews
to identify problem areas with
engagement is too late. Using predictive
machine learning gets you the data you
needed to take action before it’s too late.
Employee
30. Employee Retention Model.
It’s better to use pattern recognition to
find the employees that are at risk of
leaving, and target those employees with
incentives. Monitor critical factors like:
● Monthly income
● Overtime
● Age
● Distance from home
● Total working years
● Years at the company
● Years with the current manager
31. To see the biggest results from the
insights pulled from the data,
organizations need to get the buy-in from
teams and executives company wide.
Data visualization takes data and uses it
to tell the story that will gain that buy-in
and get teams to understand how
initiatives are impacting their role,
department and the company as a whole.
Bringing a Modern Approach to Your Organization