h2o world - translating advanced analytics for business users - conor jensen
TRANSCRIPT
Translating advanced analytics for business users11 Nov 20152015 H2O WorldConor JensenZurich North America
© 2015 Zurich American Insurance Company. All rights reserved.
2
1. Identify the analytics consumer
2. Hire the right talent
3. Develop the analytics collaboratively
4. Tell a story with your data
Developing analytics for business users
© 2014 Zurich American Insurance Company. All rights reserved.
Who will be the end user?
3
Analytics for machines: computers execute on the outputs of the analytics
Analytics for humans: humans – often non-data scientists – execute on the outputs
Source: HBR, The Question to Ask Before Hiring a Data Scientist, Michael Li, August 6, 2014
© 2015 Zurich American Insurance Company. All rights reserved.
4
Finding the right analytics talent is key
“Traditional” view of the Data Scientist continuum:
Researcher/Mathematician
Applied scientist/Statistician
Data Engineer/Data Analyst
Computer Scientist
Data Scienti
st
This is an extremely important and helpful frame of reference…
© 2015 Zurich American Insurance Company. All rights reserved.
5
Talent must be tailored to your organization
• Understand what kind of talent your organization needs and where your gaps are
• Don’t focus on hiring the unicorn, hire people with experience developing models for humans
• The most important language any analyst should know is that of your business
However this omits the most important dimension
Computer Science
Stat
isti
cs
Busin
ess
© 2015 Zurich American Insurance Company. All rights reserved.
6
An amazing analytics team isn’t enough
Which of these sound like your organization?
“In God we trust, all others bring data”-The Elements of Statistical Learning
“There are lies, damn lies, and statistics”-The Autobiography of Mark Twain
Or
A good analytics team is a necessary, but not sufficient condition for creating a data driven organization
© 2015 Zurich American Insurance Company. All rights reserved.
7
Collaborate throughout the process
• Determine what problems need to be solved
• Define the problem you’re trying to solve
• Create multiple solutions and discuss results with end users
• Iterate frequently with the end users to make sure the solution satisfies the problem
• Create a robust feedback loop for models in production
Data Preparation
Insight Generation
Solution Design
Prioritization
Develop Analytics
Plan
Solution Design
Implemen-tation
Lifecycle Managemen
t
Prioritization
Illustrative analytics process
Examples where you should be collaborating:
© 2015 Zurich American Insurance Company. All rights reserved.
8
Be mindful of trade offs during development
Accuracy
Stability (fitting) Speed
Clarity
While these aren’t necessarily mutually exclusive goals, you need to prioritize which are most important for any given project
There are many potentially competing priorities
© 2015 Zurich American Insurance Company. All rights reserved.
9
Data science isn't about cool graphics and visualizations; it's about telling �� ��a story
There’s more than one way to tell a story
The story should be tailored for the audience
© 2015 Zurich American Insurance Company. All rights reserved.
10
1. You don’t have to explain to the user how the model works… …however you must explain how users can best use the model
2. Iterate frequently with selected end users during the model development, and actively solicit feedback after models are deployed
3. Tailor the message and the visualization to the audience
4. Explicitly acknowledge outliers and explain how users should evaluate them so they don’t erode confidence in the model
5. Be explicit about where the data behind the model is thin and the users need to apply more judgment
6. Though very important, use visualizations only when they support the story effectively
Tips for effective storytelling
© 2015 Zurich American Insurance Company. All rights reserved.
11
Visualizations should be intuitive
“Graphical elegance is often found in simplicity of design and complexity of data”
The Visual Display of Quantitative Information, Edward Tufte
$2-5B
$750Mto 1B
$1-2B
240
49
81
55
55
W&RT
261
39
75
74
73
FI
322
73
104
71
74
T&A
148
26
38
34
50
T&C
205
36
44
68
13
39
33
4057
MFG
$5B+
Other
510
123
167
139
81
F&A
125
Cust
omer
rev
enue
ba
nd
Industry group
Bad examples
© 2015 Zurich American Insurance Company. All rights reserved.
12
Three steps to effectively develop advanced analytics for business users once you’ve identified the right audience:1. Hire and develop analytics professionals with business knowledge2. Build your analytics in collaboration with the end users3. Present the data in a way that users can consume
The role of the business user is to interpret and apply the model results; so they must know where to trust the model, and where to rely on their own judgment
Always remember, the models are just models and it is our responsibility to educate the users
Finally, predictive analytics that are inputs into a decision making process must be developed for, with, and ultimately understood by the people making the decision
Parting thoughts
13© 2014 Zurich American Insurance Company. All rights reserved.
Discussion
Q & A