Slide 1
Slide 1 text
Water they doing with those meters?
Amelia Taylor
A consulting project for Valor Water Analytics:
Insight Fellow
Slide 2
Slide 2 text
Utility companies have leaks.
Slide 3
Slide 3 text
Analytics Companies Have a Problem
Slide 4
Slide 4 text
Analytics Companies Have a Problem
Prototype
Model
Slide 5
Slide 5 text
Analytics Companies Have a Problem
Prototype
Model
Meter Tampering Scores
Slide 6
Slide 6 text
Analytics Companies Have a Problem
Prototype
Model
Meter Tampering Scores
New Scores
Using new
supervised data
Slide 7
Slide 7 text
Data Existing
Model Scores
&
Labels
Algorithms
Slide 8
Slide 8 text
Data Existing
Model Scores
&
Labels
SMOTE
Resample
Algorithms
Slide 9
Slide 9 text
Data Existing
Model Scores
&
Labels
SMOTE
Resample
Random
Forest
Classification
New Scores
Algorithms
Slide 10
Slide 10 text
Data Existing
Model Scores
&
Labels
SMOTE
Resample
Random
Forest
Classification
Gradient
Descent New Scores
Algorithms
Slide 11
Slide 11 text
Metrics
My algorithm: Random
Forest with SMOTE
Company Prototype Model
ROC Curves
Increased AUC by 70%
Slide 12
Slide 12 text
Metrics
My algorithm: Random
Forest with SMOTE
Company Prototype Model
ROC Curves
Increased AUC by 70%
Slide 13
Slide 13 text
Metrics
My algorithm: Random
Forest with SMOTE
Company Prototype Model
ROC Curves
Increased AUC by 70%
Slide 14
Slide 14 text
Actionable Insights
Quickly identified features not performing as expected.
Tampered Untampered
Slide 15
Slide 15 text
Actionable Insights
Quickly identified features not performing as expected.
Tampered Untampered
Slide 16
Slide 16 text
Actionable Insights
Untampered Tampered
99.924% 0.076%
Untampered Tampered
50% 50%
Quickly identified features not performing as expected.
Before SMOTE After SMOTE
Tampered Untampered
Slide 17
Slide 17 text
Deliverables
1. Python Modules & API
Slide 18
Slide 18 text
Deliverables
1. Python Modules & API
2. Scalable platform for improving scores from
multiple prototype models.
Slide 19
Slide 19 text
Deliverables
1. Python Modules & API
2. Scalable platform for improving scores from
multiple prototype models.
3. Improved the true positive rate (recall) balancing
that with the false positive rate.
AUC Increased 70%
Slide 20
Slide 20 text
About Me Amelia Taylor
Green: Unbiased Squangle
PhD Mathematics
University of Kansas