WELCOME
WHO ARE WE? David Brader Noah Erickson Data Scientist Director, Process Improvement Titan America LLC Titan America LLC 3
WHAT WE LEARNED… Highway To The Danger Zone -the key to staying out of it… Mediators, Moderators and M2’s Oh My -when variables model differently than in the process world Making Data Science Meaningful In Real-Time 4
EVER NOTICE EVERYONE IS A DATA SCIENTIST? And they almost invariably have the wrong perspective? - Like this one, the typical Senior Manager And Above: 5
CEO & E-SUITE PERSPECTIVE Data Science 6
SIMPLE RIGHT? …JUST GET THE DATA… 7
STILL GETTING THE DATA… 8
NO SUCH THING AS BAD DATA, JUST OPPORTUNITIES FOR A LOT OF DATA CLEANING 9
DATA SCIENCE – THE REAL PYRAMID 10
WHY DOMO AND DOMO DATA SCIENCE? Standard Playbook Each Step Is Somewhat Automated Within It’s Function But Often Manually Transported Through To The Next Stage Analytics Data Visualize Preparation (ML) But Being Forward Looking, and In Order To Ensure Proper Human Capital Consumption (i.e. lazy) Our Vision Was…. 11
DATA SCIENCE CONSUMPTION AS A CONTINUOUS OPERATING PROCESS Near Real Data Time Process Acquisition At Present Within DOMO: Parameter and Preparation Manipulation >100 Data Sources Ø Every Step Happens 20,300 Unique Columns Ø VISUALIZATION Automatically >381 M Rows Ø As New Data Is High Level Low Level Uploaded In Alert and Alert and Process Data Being Logged Every Ø Notification Notification Near Real Time 10 Minutes at 5 s Frequency Data Analysis (ML) 12
WE WANTED TO BE ABLE TO RAPIDLY DEPLOY THAT MODEL So We Teamed Up With DOMO Data Science – We just needed them to show us how to use the R and Python Tiles in DOMO… But They Made Us Do All This Other Work Data Glossary and Definition Sheets – Give Them Process Flow Charts (Actual – Process Not Data Flow) Have Meetings….. – Why? We Were Already In The Data Science Zone?!? 13
…THEN, A WEEKEND GETS RUINED… 14
SHOW ME…. Turns Out They Really Read That 300+ Page Dissertation I Sent Them! – We Were Here THE FUNDAMENTAL THEORETICAL RELATIONSHIP BETWEEN THE PRIMARY PREDICTIVE FEATURE AND THE VALUE WE WERE TRYING TO PREDICT WAS NOT PRESENT IN OUR DATA!?! 15
NOAH (& TONY) GOT BACK TO WORK…. Expanded Dataset To Encompass More Variation And Proper Relationships And…. Last Version Of Modeling Gave Us A 33% Improvement In The Baseline R 2 From Historical Modeling In The Literature! 16
SUCCESS!....? But, we had an inline model that outperforms those in the literature, why maybe? Remember we wanted something that would be inline and near real time, turns out the DOMO Datascience Process showed us something very different… 17
WE WEREN’T READY FOR THE REAL WORLD… The bulk of the features that were significant in predicting the target value, are ones that we currently ONLY GET AT THE SAME TIME as the value we wanted to predict, which is 2 days after the actual processing WE NEED NEW MEANS BY WHICH TO GET THE NECESSARY DATA ANALYZED AND CAPTURED REAL TIME! THIS MEANS CAPEX! NEW ANALYTICAL EQUIPMENT ON ORDER 18
NOW FOR A LOOK UNDER THE HOOD…. 19
LESSONS LEARNED The Process Will Set You Free – Or At Least Keep You Out Of Trouble Be Prepared For Features & Variables That Will Behave Far Differently In A Real World Operating Process Than They Do In Controlled Situations Institutionalizing Your Solutions Requires Them To Operate At The Same Frequency As Your Decision Making 20
THANK YOU
Recommend
More recommend
