Enviar búsqueda
Cargar
1440 track2 roberts
•
4 recomendaciones
•
1,483 vistas
Rising Media, Inc.
Seguir
Pasha Roberts
Leer menos
Leer más
Datos y análisis
Denunciar
Compartir
Denunciar
Compartir
1 de 43
Descargar ahora
Descargar para leer sin conexión
Recomendados
1030 track1 heiler
1030 track1 heiler
Rising Media, Inc.
1120 track 3 prendki_using our laptop
1120 track 3 prendki_using our laptop
Rising Media, Inc.
1345 keynote roberts
1345 keynote roberts
Rising Media, Inc.
1000 track1 gland_sims
1000 track1 gland_sims
Rising Media, Inc.
1555 track1 alam
1555 track1 alam
Rising Media, Inc.
1000 track2 boire
1000 track2 boire
Rising Media, Inc.
1030 track 3 rolleston_using our laptop
1030 track 3 rolleston_using our laptop
Rising Media, Inc.
1645 track 1 bress_using his laptop
1645 track 1 bress_using his laptop
Rising Media, Inc.
Recomendados
1030 track1 heiler
1030 track1 heiler
Rising Media, Inc.
1120 track 3 prendki_using our laptop
1120 track 3 prendki_using our laptop
Rising Media, Inc.
1345 keynote roberts
1345 keynote roberts
Rising Media, Inc.
1000 track1 gland_sims
1000 track1 gland_sims
Rising Media, Inc.
1555 track1 alam
1555 track1 alam
Rising Media, Inc.
1000 track2 boire
1000 track2 boire
Rising Media, Inc.
1030 track 3 rolleston_using our laptop
1030 track 3 rolleston_using our laptop
Rising Media, Inc.
1645 track 1 bress_using his laptop
1645 track 1 bress_using his laptop
Rising Media, Inc.
1120 track1 taylor
1120 track1 taylor
Rising Media, Inc.
1555 track 1 huang_using his mac
1555 track 1 huang_using his mac
Rising Media, Inc.
1530 track2 reid
1530 track2 reid
Rising Media, Inc.
1615 track1 schleicher
1615 track1 schleicher
Rising Media, Inc.
925 plenary rexer_using our laptop
925 plenary rexer_using our laptop
Rising Media, Inc.
1055 track3 soules
1055 track3 soules
Rising Media, Inc.
940 diamond sponsor sengupta
940 diamond sponsor sengupta
Rising Media, Inc.
1120 track1 grossman
1120 track1 grossman
Rising Media, Inc.
1530 track 3 gunther_using our laptop
1530 track 3 gunther_using our laptop
Rising Media, Inc.
1315 keynote jopia_shareable
1315 keynote jopia_shareable
Rising Media, Inc.
1000 track 1 groves_using our laptop
1000 track 1 groves_using our laptop
Rising Media, Inc.
1645 track 3 porter
1645 track 3 porter
Rising Media, Inc.
1620 track1 lee
1620 track1 lee
Rising Media, Inc.
1530 track1 rosenbaum
1530 track1 rosenbaum
Rising Media, Inc.
840 plenary elder_using his laptop
840 plenary elder_using his laptop
Rising Media, Inc.
Data Quality Analytics: Understanding what is in your data, before using it
Data Quality Analytics: Understanding what is in your data, before using it
Domino Data Lab
20131108 - Big Data presentation at the 2nd GDG Brunei Darussalam DevFest
20131108 - Big Data presentation at the 2nd GDG Brunei Darussalam DevFest
Izam Ryan
920 plenary elder
920 plenary elder
Rising Media, Inc.
From Insight to Impact (Chicago Summit - Keynote)
From Insight to Impact (Chicago Summit - Keynote)
Open Analytics
The Value of Predictive Analytics and Decision Modeling
The Value of Predictive Analytics and Decision Modeling
Decision Management Solutions
1505 Statistical Thinking course extract
1505 Statistical Thinking course extract
Jefferson Lynch
Rethinking Risk-Based Project Management in the Emerging IT initiatives.pptx
Rethinking Risk-Based Project Management in the Emerging IT initiatives.pptx
Inflectra
Más contenido relacionado
La actualidad más candente
1120 track1 taylor
1120 track1 taylor
Rising Media, Inc.
1555 track 1 huang_using his mac
1555 track 1 huang_using his mac
Rising Media, Inc.
1530 track2 reid
1530 track2 reid
Rising Media, Inc.
1615 track1 schleicher
1615 track1 schleicher
Rising Media, Inc.
925 plenary rexer_using our laptop
925 plenary rexer_using our laptop
Rising Media, Inc.
1055 track3 soules
1055 track3 soules
Rising Media, Inc.
940 diamond sponsor sengupta
940 diamond sponsor sengupta
Rising Media, Inc.
1120 track1 grossman
1120 track1 grossman
Rising Media, Inc.
1530 track 3 gunther_using our laptop
1530 track 3 gunther_using our laptop
Rising Media, Inc.
1315 keynote jopia_shareable
1315 keynote jopia_shareable
Rising Media, Inc.
1000 track 1 groves_using our laptop
1000 track 1 groves_using our laptop
Rising Media, Inc.
1645 track 3 porter
1645 track 3 porter
Rising Media, Inc.
1620 track1 lee
1620 track1 lee
Rising Media, Inc.
1530 track1 rosenbaum
1530 track1 rosenbaum
Rising Media, Inc.
840 plenary elder_using his laptop
840 plenary elder_using his laptop
Rising Media, Inc.
Data Quality Analytics: Understanding what is in your data, before using it
Data Quality Analytics: Understanding what is in your data, before using it
Domino Data Lab
20131108 - Big Data presentation at the 2nd GDG Brunei Darussalam DevFest
20131108 - Big Data presentation at the 2nd GDG Brunei Darussalam DevFest
Izam Ryan
920 plenary elder
920 plenary elder
Rising Media, Inc.
From Insight to Impact (Chicago Summit - Keynote)
From Insight to Impact (Chicago Summit - Keynote)
Open Analytics
The Value of Predictive Analytics and Decision Modeling
The Value of Predictive Analytics and Decision Modeling
Decision Management Solutions
La actualidad más candente
(20)
1120 track1 taylor
1120 track1 taylor
1555 track 1 huang_using his mac
1555 track 1 huang_using his mac
1530 track2 reid
1530 track2 reid
1615 track1 schleicher
1615 track1 schleicher
925 plenary rexer_using our laptop
925 plenary rexer_using our laptop
1055 track3 soules
1055 track3 soules
940 diamond sponsor sengupta
940 diamond sponsor sengupta
1120 track1 grossman
1120 track1 grossman
1530 track 3 gunther_using our laptop
1530 track 3 gunther_using our laptop
1315 keynote jopia_shareable
1315 keynote jopia_shareable
1000 track 1 groves_using our laptop
1000 track 1 groves_using our laptop
1645 track 3 porter
1645 track 3 porter
1620 track1 lee
1620 track1 lee
1530 track1 rosenbaum
1530 track1 rosenbaum
840 plenary elder_using his laptop
840 plenary elder_using his laptop
Data Quality Analytics: Understanding what is in your data, before using it
Data Quality Analytics: Understanding what is in your data, before using it
20131108 - Big Data presentation at the 2nd GDG Brunei Darussalam DevFest
20131108 - Big Data presentation at the 2nd GDG Brunei Darussalam DevFest
920 plenary elder
920 plenary elder
From Insight to Impact (Chicago Summit - Keynote)
From Insight to Impact (Chicago Summit - Keynote)
The Value of Predictive Analytics and Decision Modeling
The Value of Predictive Analytics and Decision Modeling
Similar a 1440 track2 roberts
1505 Statistical Thinking course extract
1505 Statistical Thinking course extract
Jefferson Lynch
Rethinking Risk-Based Project Management in the Emerging IT initiatives.pptx
Rethinking Risk-Based Project Management in the Emerging IT initiatives.pptx
Inflectra
The Datafication of HR: Graduating from Metrics to Analytics
The Datafication of HR: Graduating from Metrics to Analytics
Visier
Cloud: Risk profiles, Preference and Decisions in Cloud
Cloud: Risk profiles, Preference and Decisions in Cloud
Amazon Web Services
Scalable agile metrics to optimize flow and reduce delivery risks
Scalable agile metrics to optimize flow and reduce delivery risks
Adonis El Fakih
Datascience101presentation4
Datascience101presentation4
Salford Systems
DFSS short
DFSS short
Chaitanya Chenna
Siemens Resource Planning Case Study
Siemens Resource Planning Case Study
Greg Bailey
9. Risk.ppt
9. Risk.ppt
Kameswara Rao Poranki
9. Risk.ppt
9. Risk.ppt
RajeevRanjan338430
project managment.ppt
project managment.ppt
HinaAsghar16
9. Risk.ppt
9. Risk.ppt
HemaLatha208756
9. Risk.ppt
9. Risk.ppt
HinaAsghar16
9. Risk.ppt
9. Risk.ppt
AyidAlmgati
9. Risk.ppt
9. Risk.ppt
HinaAsghar16
9. Risk.ppt
9. Risk.ppt
AayushSharma261
Dear Dad
Dear Dad
Nina Vazquez
Project Management 08
Project Management 08
Tathagat Varma
Stop Talking About What You Do and Start Talking About What You Deliver - Pro...
Stop Talking About What You Do and Start Talking About What You Deliver - Pro...
Tim Creasey
Asking When, Not If in Predictive Modeling
Asking When, Not If in Predictive Modeling
Andrea Kropp
Similar a 1440 track2 roberts
(20)
1505 Statistical Thinking course extract
1505 Statistical Thinking course extract
Rethinking Risk-Based Project Management in the Emerging IT initiatives.pptx
Rethinking Risk-Based Project Management in the Emerging IT initiatives.pptx
The Datafication of HR: Graduating from Metrics to Analytics
The Datafication of HR: Graduating from Metrics to Analytics
Cloud: Risk profiles, Preference and Decisions in Cloud
Cloud: Risk profiles, Preference and Decisions in Cloud
Scalable agile metrics to optimize flow and reduce delivery risks
Scalable agile metrics to optimize flow and reduce delivery risks
Datascience101presentation4
Datascience101presentation4
DFSS short
DFSS short
Siemens Resource Planning Case Study
Siemens Resource Planning Case Study
9. Risk.ppt
9. Risk.ppt
9. Risk.ppt
9. Risk.ppt
project managment.ppt
project managment.ppt
9. Risk.ppt
9. Risk.ppt
9. Risk.ppt
9. Risk.ppt
9. Risk.ppt
9. Risk.ppt
9. Risk.ppt
9. Risk.ppt
9. Risk.ppt
9. Risk.ppt
Dear Dad
Dear Dad
Project Management 08
Project Management 08
Stop Talking About What You Do and Start Talking About What You Deliver - Pro...
Stop Talking About What You Do and Start Talking About What You Deliver - Pro...
Asking When, Not If in Predictive Modeling
Asking When, Not If in Predictive Modeling
Más de Rising Media, Inc.
1415 track 1 wu_using his laptop
1415 track 1 wu_using his laptop
Rising Media, Inc.
Matt gershoff
Matt gershoff
Rising Media, Inc.
Keynote adam greco
Keynote adam greco
Rising Media, Inc.
1620 keynote olson_using our laptop
1620 keynote olson_using our laptop
Rising Media, Inc.
1530 track 2 stuart_using our laptop
1530 track 2 stuart_using our laptop
Rising Media, Inc.
1530 track 1 fader_using our laptop
1530 track 1 fader_using our laptop
Rising Media, Inc.
1415 track 2 richardson
1415 track 2 richardson
Rising Media, Inc.
1215 daa lunch owusu_using our laptop
1215 daa lunch owusu_using our laptop
Rising Media, Inc.
1215 daa lunch a bos intro slides_using our laptop
1215 daa lunch a bos intro slides_using our laptop
Rising Media, Inc.
915 e metrics_claudia perlich
915 e metrics_claudia perlich
Rising Media, Inc.
855 sponsor movassate_using our laptop
855 sponsor movassate_using our laptop
Rising Media, Inc.
1615 plack using our laptop
1615 plack using our laptop
Rising Media, Inc.
1530 rimmele do not share
1530 rimmele do not share
Rising Media, Inc.
1325 keynote yale_pdf shareable
1325 keynote yale_pdf shareable
Rising Media, Inc.
1115 fiztgerald schuchardt
1115 fiztgerald schuchardt
Rising Media, Inc.
1000 kondic do not share
1000 kondic do not share
Rising Media, Inc.
905 keynote peele_using our laptop
905 keynote peele_using our laptop
Rising Media, Inc.
Stephen morse sharable
Stephen morse sharable
Rising Media, Inc.
Elder shareable
Elder shareable
Rising Media, Inc.
1115 ramirez using our laptop
1115 ramirez using our laptop
Rising Media, Inc.
Más de Rising Media, Inc.
(20)
1415 track 1 wu_using his laptop
1415 track 1 wu_using his laptop
Matt gershoff
Matt gershoff
Keynote adam greco
Keynote adam greco
1620 keynote olson_using our laptop
1620 keynote olson_using our laptop
1530 track 2 stuart_using our laptop
1530 track 2 stuart_using our laptop
1530 track 1 fader_using our laptop
1530 track 1 fader_using our laptop
1415 track 2 richardson
1415 track 2 richardson
1215 daa lunch owusu_using our laptop
1215 daa lunch owusu_using our laptop
1215 daa lunch a bos intro slides_using our laptop
1215 daa lunch a bos intro slides_using our laptop
915 e metrics_claudia perlich
915 e metrics_claudia perlich
855 sponsor movassate_using our laptop
855 sponsor movassate_using our laptop
1615 plack using our laptop
1615 plack using our laptop
1530 rimmele do not share
1530 rimmele do not share
1325 keynote yale_pdf shareable
1325 keynote yale_pdf shareable
1115 fiztgerald schuchardt
1115 fiztgerald schuchardt
1000 kondic do not share
1000 kondic do not share
905 keynote peele_using our laptop
905 keynote peele_using our laptop
Stephen morse sharable
Stephen morse sharable
Elder shareable
Elder shareable
1115 ramirez using our laptop
1115 ramirez using our laptop
Último
办理学位证纽约大学毕业证(NYU毕业证书)原版一比一
办理学位证纽约大学毕业证(NYU毕业证书)原版一比一
fhwihughh
20240419 - Measurecamp Amsterdam - SAM.pdf
20240419 - Measurecamp Amsterdam - SAM.pdf
Human37
Generative AI for Social Good at Open Data Science East 2024
Generative AI for Social Good at Open Data Science East 2024
Colleen Farrelly
Indian Call Girls in Abu Dhabi O5286O24O8 Call Girls in Abu Dhabi By Independ...
Indian Call Girls in Abu Dhabi O5286O24O8 Call Girls in Abu Dhabi By Independ...
dajasot375
DBA Basics: Getting Started with Performance Tuning.pdf
DBA Basics: Getting Started with Performance Tuning.pdf
John Sterrett
专业一比一美国俄亥俄大学毕业证成绩单pdf电子版制作修改
专业一比一美国俄亥俄大学毕业证成绩单pdf电子版制作修改
yuu sss
RS 9000 Call In girls Dwarka Mor (DELHI)⇛9711147426🔝Delhi
RS 9000 Call In girls Dwarka Mor (DELHI)⇛9711147426🔝Delhi
jennyeacort
NLP Project PPT: Flipkart Product Reviews through NLP Data Science.pptx
NLP Project PPT: Flipkart Product Reviews through NLP Data Science.pptx
Boston Institute of Analytics
From idea to production in a day – Leveraging Azure ML and Streamlit to build...
From idea to production in a day – Leveraging Azure ML and Streamlit to build...
Florian Roscheck
Top 5 Best Data Analytics Courses In Queens
Top 5 Best Data Analytics Courses In Queens
dataanalyticsqueen03
RadioAdProWritingCinderellabyButleri.pdf
RadioAdProWritingCinderellabyButleri.pdf
gstagge
Machine learning classification ppt.ppt
Machine learning classification ppt.ppt
amreenkhanum0307
Predicting Salary Using Data Science: A Comprehensive Analysis.pdf
Predicting Salary Using Data Science: A Comprehensive Analysis.pdf
Boston Institute of Analytics
NLP Data Science Project Presentation:Predicting Heart Disease with NLP Data ...
NLP Data Science Project Presentation:Predicting Heart Disease with NLP Data ...
Boston Institute of Analytics
RABBIT: A CLI tool for identifying bots based on their GitHub events.
RABBIT: A CLI tool for identifying bots based on their GitHub events.
natarajan8993
Call Us ➥97111√47426🤳Call Girls in Aerocity (Delhi NCR)
Call Us ➥97111√47426🤳Call Girls in Aerocity (Delhi NCR)
jennyeacort
毕业文凭制作#回国入职#diploma#degree澳洲中央昆士兰大学毕业证成绩单pdf电子版制作修改#毕业文凭制作#回国入职#diploma#degree
毕业文凭制作#回国入职#diploma#degree澳洲中央昆士兰大学毕业证成绩单pdf电子版制作修改#毕业文凭制作#回国入职#diploma#degree
yuu sss
1:1定制(UQ毕业证)昆士兰大学毕业证成绩单修改留信学历认证原版一模一样
1:1定制(UQ毕业证)昆士兰大学毕业证成绩单修改留信学历认证原版一模一样
vhwb25kk
Call Girls In Dwarka 9654467111 Escorts Service
Call Girls In Dwarka 9654467111 Escorts Service
Sapana Sha
GA4 Without Cookies [Measure Camp AMS]
GA4 Without Cookies [Measure Camp AMS]
📊 Markus Baersch
Último
(20)
办理学位证纽约大学毕业证(NYU毕业证书)原版一比一
办理学位证纽约大学毕业证(NYU毕业证书)原版一比一
20240419 - Measurecamp Amsterdam - SAM.pdf
20240419 - Measurecamp Amsterdam - SAM.pdf
Generative AI for Social Good at Open Data Science East 2024
Generative AI for Social Good at Open Data Science East 2024
Indian Call Girls in Abu Dhabi O5286O24O8 Call Girls in Abu Dhabi By Independ...
Indian Call Girls in Abu Dhabi O5286O24O8 Call Girls in Abu Dhabi By Independ...
DBA Basics: Getting Started with Performance Tuning.pdf
DBA Basics: Getting Started with Performance Tuning.pdf
专业一比一美国俄亥俄大学毕业证成绩单pdf电子版制作修改
专业一比一美国俄亥俄大学毕业证成绩单pdf电子版制作修改
RS 9000 Call In girls Dwarka Mor (DELHI)⇛9711147426🔝Delhi
RS 9000 Call In girls Dwarka Mor (DELHI)⇛9711147426🔝Delhi
NLP Project PPT: Flipkart Product Reviews through NLP Data Science.pptx
NLP Project PPT: Flipkart Product Reviews through NLP Data Science.pptx
From idea to production in a day – Leveraging Azure ML and Streamlit to build...
From idea to production in a day – Leveraging Azure ML and Streamlit to build...
Top 5 Best Data Analytics Courses In Queens
Top 5 Best Data Analytics Courses In Queens
RadioAdProWritingCinderellabyButleri.pdf
RadioAdProWritingCinderellabyButleri.pdf
Machine learning classification ppt.ppt
Machine learning classification ppt.ppt
Predicting Salary Using Data Science: A Comprehensive Analysis.pdf
Predicting Salary Using Data Science: A Comprehensive Analysis.pdf
NLP Data Science Project Presentation:Predicting Heart Disease with NLP Data ...
NLP Data Science Project Presentation:Predicting Heart Disease with NLP Data ...
RABBIT: A CLI tool for identifying bots based on their GitHub events.
RABBIT: A CLI tool for identifying bots based on their GitHub events.
Call Us ➥97111√47426🤳Call Girls in Aerocity (Delhi NCR)
Call Us ➥97111√47426🤳Call Girls in Aerocity (Delhi NCR)
毕业文凭制作#回国入职#diploma#degree澳洲中央昆士兰大学毕业证成绩单pdf电子版制作修改#毕业文凭制作#回国入职#diploma#degree
毕业文凭制作#回国入职#diploma#degree澳洲中央昆士兰大学毕业证成绩单pdf电子版制作修改#毕业文凭制作#回国入职#diploma#degree
1:1定制(UQ毕业证)昆士兰大学毕业证成绩单修改留信学历认证原版一模一样
1:1定制(UQ毕业证)昆士兰大学毕业证成绩单修改留信学历认证原版一模一样
Call Girls In Dwarka 9654467111 Escorts Service
Call Girls In Dwarka 9654467111 Escorts Service
GA4 Without Cookies [Measure Camp AMS]
GA4 Without Cookies [Measure Camp AMS]
1440 track2 roberts
1.
How to Use
Survival Analytics to Predict Churn (Or, Why You Shouldn't Use Logistic Regression to Predict Attrition) Talent Analytics, Corp. Pasha Roberts, Chief Scientist June 20, 2017 Chicago USA Demo Code: https://github.com/talentanalytics/class_survival_101/ www.talentanalytics.com © 2017 Talent Analytics, Corp 1 / 41
2.
▸ ▸ ▹ ▹ ▸ ▸ Who is Talent
Analytics? Predictive Modeling Platform – Advisor™ We predict employee attrition and performance pre-hire. Much like credit risk modeling: Predict likelihood to pay / default on mortgage, before extending credit Predict likelihood to perform / leave role or company early, before extending job offer PAAS (Prediction As A Service) Seamless deployment of predictions into talent acquisition process www.talentanalytics.com © 2017 Talent Analytics, Corp 2 / 41
3.
Turnover and Tenure: A
Timey-Wimey Relationship www.talentanalytics.com © 2017 Talent Analytics, Corp 3 / 41
4.
▸ ▸ Two Different (But
Related) Things Turnover: Percent of employees (customers) that terminate within a period of time Tenure: Employees’ (customers’) length of time with company Is it a Wave or a Particle? www.talentanalytics.com © 2017 Talent Analytics, Corp 4 / 41
5.
Low Tenure, High
Turnover www.talentanalytics.com © 2017 Talent Analytics, Corp 5 / 41
6.
High Tenure, Low
Turnover www.talentanalytics.com © 2017 Talent Analytics, Corp 6 / 41
7.
High Tenure, High
Turnover (in 2015) www.talentanalytics.com © 2017 Talent Analytics, Corp 7 / 41
8.
Low Tenure, Low
Turnover (in Nov, Dec) www.talentanalytics.com © 2017 Talent Analytics, Corp 8 / 41
9.
▸ ▸ ▸ Main Problem: We
Can't See the Future Everyone will terminate. Someday, somehow... But when? Technical Lingo: Right Censoring www.talentanalytics.com © 2017 Talent Analytics, Corp 9 / 41
10.
▸ ▸ ▸ Each Side Has
Missing Information Turnover Has No Information About Tenure Treats a temp the same as a seasoned veteran Ignores patterns - such as people quitting right after a bonus Tenure Has No Information About Turnover High Tenure = staying power or retirement risk? Each Side Varies Depending on When You Look www.talentanalytics.com © 2017 Talent Analytics, Corp 10 / 41
11.
▸ ▸ Annual Attrition Timing
is Arbitrary What if you could see attrition at every point in time? “One Year” only matters to accountants and astronomers Cumulative Breakeven matters more to the business www.talentanalytics.com © 2017 Talent Analytics, Corp 11 / 41
12.
Survival 101 www.talentanalytics.com ©
2017 Talent Analytics, Corp 12 / 41
13.
▸ ▸ ▸ ▸ ▸ ▸ Survival Analytics Medical Roots
- How long will a patient survive a disease? Engineering Roots - How long until a machine fails? Social Science - How long do people live? Employment - How long until an employee terminates? or promotes? In General - “Expected duration of time until one or more events happen” Artfully combines turnover and tenure Lots of competing jargon and syntax from many application domains: Failure Rate, Hazard Rate, Force of Mortality, ... www.talentanalytics.com © 2017 Talent Analytics, Corp 13 / 41
14.
▸ ▹ ▹ ▹ ▸ ▸ ▸ Hazard Rate: Conditional Daily
Risk of Termination Every day in tenure there is a (small) probability that the employee will terminate or the customer will drop... or the patient will die... or the event will occur Conditional on survival to the prior day Like an actuarial life table Can rise at different times: e.g. training, review or bonus time www.talentanalytics.com © 2017 Talent Analytics, Corp 14 / 41
15.
▸ ▸ ▸ ▸ Survival Rate: Daily
Probability of Being Employed at Time t How likely are you to be working here, t days after being hired? Near 100% on Day 1 (some people actually never show up) Downhill from there - usually not linear One role at a time www.talentanalytics.com © 2017 Talent Analytics, Corp 15 / 41
16.
▸ ▸ ▸ ▸ ▸ ▸ ▹ ▹ Survival Rate: Daily
Probability of Being Employed at Time t How likely are you to be working here, t days after being hired? Near 100% on Day 1 (some people actually never show up) Downhill from there - usually not linear One role at a time This is the full picture Simple transformation of the Hazard Curve Best to use statistical tool (R, SAS) Apparently possible in Excel www.talentanalytics.com © 2017 Talent Analytics, Corp 15 / 41
17.
Attrition is Built
Into Survival Curves www.talentanalytics.com © 2017 Talent Analytics, Corp 16 / 41
18.
Shelf Pattern Sparse
Pattern Many Shapes www.talentanalytics.com © 2017 Talent Analytics, Corp 17 / 41
19.
Heavy Terms During
Training Heavy Terms After 1-Year Bonus Cliffs or Kinks in the Survival Curve www.talentanalytics.com © 2017 Talent Analytics, Corp 18 / 41
20.
Everyone Has a
Survival Curve www.talentanalytics.com © 2017 Talent Analytics, Corp 19 / 41
21.
Predicting Survival To the
Individual www.talentanalytics.com © 2017 Talent Analytics, Corp 20 / 41
22.
▸ ▸ ▸ ▹ ▸ ▸ ▸ Don't Use Logistic
Methods to Predict Attrition! Commonly Done: Predict one-year attrition with logistic Use tenure as a variable along with others Common Result: “The biggest cause of termination is tenure” Not an actionable coef cient Ignores nuances available to survival methods Mishandles current employee tenure Less accurate www.talentanalytics.com © 2017 Talent Analytics, Corp 21 / 41
23.
▸ ▸ ▸ ▸ Proportional Hazard Most Survival
models are “Proportional Hazard” Start with Base Survival rate Convert to Base Cumulative Hazard rate Multiply Cumulative Hazard by a single factor (hence “Proportional Hazard") Linear changes to Hazard lead to a rotation of Survival Curve Predictive Goal: Predict the amount of Survival Curve rotation for each candidate www.talentanalytics.com © 2017 Talent Analytics, Corp 22 / 41
24.
▸ ▹ ▹ ▹ ▹ ▸ ▸ ▸ ▸ Proportional Hazard Prediction
Methods Predicting a Single Number Cox regression Vanilla Cox Stepwise ElasticNet or LASSO Variable selection with Random Forest Decision Trees Random Forests Neural Networks SVM, etc There are also other survival algorithms, including parametric methods www.talentanalytics.com © 2017 Talent Analytics, Corp 23 / 41
25.
▸ ▸ ▸ ▸ Build/Validate a Survival
Model With Thresholds The Model Predicts: “Dark Blue” hires will survive the longest “Light Blue” hires will survive above-median “Grey” survival curve is current median survival. “Light Red” and “Dark Red” candidates are not likely to survive. www.talentanalytics.com © 2017 Talent Analytics, Corp 24 / 41
26.
Deeper Dive: Predicting Survival with
Proportional Hazard Follow Demo Code on GitHub: https://github.com/talentanalytics/class_survival_101/ www.talentanalytics.com © 2017 Talent Analytics, Corp 25 / 41
27.
▸ ▸ ▸ ▸ ▸ ▸ Formulas and Jargon Hazard
Function: Probability of termination at time t, conditional on survival up to time t Ranges from 0 to 1 (typically very small) Cumulative Hazard Function: Cumulative conditional probability of termination up to time t Ranges from 0 to 4ish (theoretically to In nity) Survival Function: Probability that termination will be later than time t Ranges from 1 to 0 Theoretically the formulas above are continuous measures, but in practice are discrete daily. www.talentanalytics.com © 2017 Talent Analytics, Corp h(t) H(t) = h(x) dx = −log(S(t))∫ t 0 S(t) = e −H(t) 26 / 41
28.
Calculating Survival in
R Four Simple Steps with R survival package: 1. Gather and Prepare the Data 2. Calculate Baseline Survival Function 3. Model Proportional Hazard with Cox Regression 4. Validate Model www.talentanalytics.com © 2017 Talent Analytics, Corp (t)S0 27 / 41
29.
▸ ▸ ▸ Calculating Survival in
R Four Simple Steps with R survival package: 1. Gather and Prepare the Data 2. Calculate Baseline Survival Function 3. Model Proportional Hazard with Cox Regression 4. Validate Model 5. Deploy Model into hiring 6. Hire people likely to stay on the job longer 7. Save and earn money Avoid attributing savings to soft factors like ripple effect, morale, engagement Hard savings from reduced Replacement Cost Hard earnings from increased Employee Lifetime Value www.talentanalytics.com © 2017 Talent Analytics, Corp (t)S0 27 / 41
30.
▸ ▹ ▸ ▹ ▹ ▹ 1) Gather Employee
Data > attr.data emp.id hire.date term.date factor.x factor.y factor.z <int> <dttm> <dttm> <dbl> <dbl> <dbl> 1 52 2016-02-23 <NA> 0.09555369 1.515028 102.3811 2 261 2014-10-30 <NA> 2.40144024 9.065855 100.5727 3 193 2013-12-25 2015-03-28 1.88463382 4.331642 105.6887 4 150 2013-03-27 2013-09-19 3.81625316 7.363454 110.6842 5 31 2015-03-25 <NA> 5.17392728 11.824601 110.0590 Simple Requirements Just hire.date and term.date is all you need from HRIS Leave term.date as NA if the employee haven't terminated yet Merge with predictive input (independent) variables e.g. assessment results, experience, CV detail, social media, semantic Here we use factor.x, factor.y and factor.z Are these variables available pre-hire? www.talentanalytics.com © 2017 Talent Analytics, Corp 28 / 41
31.
▸ ▸ ▹ ▹ ▸ ▹ Calculate Fields for
Survival > attr.data %>% dplyr::select(-dplyr::contains("factor")) emp.id hire.date term.date is.term end.date tenure.years <int> <dttm> <dttm> <lgl> <dttm> <dbl> 1 52 2016-02-23 <NA> FALSE 2017-05-01 1.1854894 2 261 2014-10-30 <NA> FALSE 2017-05-01 2.5023956 3 193 2013-12-25 2015-03-28 TRUE 2015-03-28 1.2539357 4 150 2013-03-27 2013-09-19 TRUE 2013-09-19 0.4818617 5 31 2015-03-25 <NA> FALSE 2017-05-01 2.1026694 Simple Derivations is.term: Have they terminated? Is term.date NA? end.date: Either term.date, or censor date, depending on is.term Could put transfer.date here with no is.term The last date we know the employee was at work tenure.years: What is employee tenure from hire.date to end.date? All we know about the employee's tenure, as of censor date www.talentanalytics.com © 2017 Talent Analytics, Corp 29 / 41
32.
▸ ▸ ▸ Dataset is Ready >
glimpse(attr.data) Observations: 400 Variables: 14 $ label <fctr> a, a, a, a, a, a, a, a, a, a, a, a, a, a, a, a, a, a,... $ hire.date <dttm> 2013-04-20, 2016-01-14, 2013-09-20, 2014-05-17, 2016-... $ term.date <dttm> NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, 2016-... $ is.term <lgl> FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE... $ end.date <dttm> 2017-05-01, 2017-05-01, 2017-05-01, 2017-05-01, 2017-... $ tenure.years <dbl> 4.02975576, 1.29469635, 3.61029191, 2.95626651, 1.1463... $ factor.x <dbl> 7.44846858, 1.08376058, 2.81983987, 0.56880027, 0.7868... $ factor.y <dbl> 16.838313, 9.335225, 9.519951, 12.697147, 12.912898, 1... $ factor.z <dbl> 103.10732, 100.37171, 99.76382, 95.88645, 97.61509, 10... $ emp.id <int> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,... $ scale.x <dbl> 2.41351313, -0.24754913, 0.47829958, -0.46285224, -0.3... $ scale.y <dbl> 1.31773198, -0.57718839, -0.53053544, 0.27187182, 0.32... $ scale.z <dbl> 0.3407242, -0.1849069, -0.3017096, -1.0467221, -0.7145... $ is.training <lgl> TRUE, TRUE, TRUE, FALSE, FALSE, FALSE, TRUE, TRUE, TRU... Example is 400-row simulated dataset Scaled the input factors to consistent z-scale Separated into 80% training and 20% validation datasets www.talentanalytics.com © 2017 Talent Analytics, Corp 30 / 41
33.
▸ 2) Calculate Baseline
Survival ( ) survival::Surv is a single object to hold Tenure and Turnover at once > surv.obj <- survival::Surv(training.data$tenure.years, training.data$is.term) Needs “time variable” tenure.years and “event variable” is.term survival::survfit is the Kaplan-Meier Estimator of > surv.fit <- survival::survfit(surv.obj ~ 1) > summary(surv.fit) Call: survfit(formula = surv.obj ~ 1) time n.risk n.event survival std.err lower 95% CI upper 95% CI 0.000 320 1 0.997 0.00312 0.991 1.000 0.249 308 1 0.994 0.00448 0.985 1.000 0.298 303 1 0.990 0.00554 0.980 1.000 0.320 298 1 0.987 0.00644 0.974 1.000 www.talentanalytics.com © 2017 Talent Analytics, Corp (t)S0 (t)S0 31 / 41
34.
Native Survival Plot
Crafted ggplot2 Plot Plot Baseline Survival www.talentanalytics.com © 2017 Talent Analytics, Corp 32 / 41
35.
▸ ▸ ▸ ▸ ▸ 3) Back to
Proportional Hazard Consider a multiplier x that centers on 1 Multiply for proportional (linear) changes to Hazard Survival Curve rotates in response: Beware - inverse relationship: x ← 1.1 will increase , decrease x ← 0.9 will decrease , increase So we just need to predict x to predict Survival One number drives the entire curve, based on baseline www.talentanalytics.com © 2017 Talent Analytics, Corp (t) = (t)xHx H0 (t) =Sx e − (t)xH0 (t)Hx (t)Sx (t)Hx (t)Sx (t)H0 33 / 41
36.
Model Proportional Hazard with
Cox Regression > cox.model <- survival::coxph(formula = surv.obj ~ scale.x + scale.y + scale.z, data = training.data) > cox.model Call: survival::coxph(formula = surv.obj ~ scale.x + scale.y + scale.z, data = training.data) coef exp(coef) se(coef) z p scale.x -0.4298 0.6506 0.0918 -4.68 0.0000028 scale.y -0.3204 0.7258 0.0957 -3.35 0.00082 scale.z -0.1967 0.8215 0.0984 -2.00 0.04558 Likelihood ratio test=38.2 on 3 df, p=0.000000025 n= 320, number of events= 125 www.talentanalytics.com © 2017 Talent Analytics, Corp 34 / 41
37.
▸ ▹ ▹ ▸ ▹ ▹ 4) Predict validation.data with
New Cox Model > cox.pred <- predict(cox.model, newdata = validation.data, type = "lp") Note: we built model with training.data, now we predict against validation.data Our model has never seen anything in validation.data Great test to see how closely our predictions match known outcomes The predict.coxph() function returns 5 different types of predictions We want type = "lp" which is log(multiplier) to baseline Other types won't work with survivalROC() www.talentanalytics.com © 2017 Talent Analytics, Corp (t)H0 35 / 41
38.
▸ ▸ ▸ ▸ How does ROC
Work with Survival? > roc.obj <- survivalROC::survivalROC(Stime = validation.data$tenure.years, status = validation.data$is.term, marker = cox.pred, predict.time = 1, lambda = 0.003) > roc.obj$AUC [1] 0.7102442 ROC typically evaluates classification methods So, we make this a classi cation problem Classify: Was the employee terminated at t = 1 or not? Compare actual vs. predicted classi cations at all cutpoints Our AUC is 0.71 Not bad for a censored attrition prediction! www.talentanalytics.com © 2017 Talent Analytics, Corp 36 / 41
39.
Plot AUC with
ggplot2 www.talentanalytics.com © 2017 Talent Analytics, Corp 37 / 41
40.
▸ ▸ ▸ ▸ Don't Get Fooled
by Randomness We are predicting AUC, too Let's try predicting 1000 random samples of the same underlying pattern > test.repl <- replicate(1000, demoPrediction(verbose = FALSE)) > summary(test.repl) Min. 1st Qu. Median Mean 3rd Qu. Max. 0.3470 0.6191 0.6833 0.6771 0.7356 0.9446 AUC generally fell between 0.619 and 0.735, but 50% were outside this range The likely AUC is closer to 0.68 - still good What if you got a lucky (or unlucky) 80/20 validation split? www.talentanalytics.com © 2017 Talent Analytics, Corp 38 / 41
41.
▸ ▸ Multiple Cross-Validation Raises the
Bar Against Getting Fooled Typically run 20-100 CV iterations Can consume lots of core-hours if nested for hyperparameters www.talentanalytics.com © 2017 Talent Analytics, Corp 39 / 41
42.
▸ ▸ ▸ ▸ ▸ Download and Experiment
With this Code These slides will be on PAW website Download demo code and doc from https://github.com/talentanalytics/class_survival_101/ Clone it, fork it, run it Send bug xes Let me know how it goes www.talentanalytics.com © 2017 Talent Analytics, Corp 40 / 41
43.
Questions and Discussion Pasha
Roberts pasha@talentanalytics.com Follow Demo Code on GitHub: https://github.com/talentanalytics/class_survival_101/ www.talentanalytics.com © 2017 Talent Analytics, Corp 41 / 41
Descargar ahora