Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
SlideShare a Scribd company logo

DataScience Lab 2017_Графические вероятностные модели для принятия решений в проектном управлении_Ольга Татаринцева

•
•
GeeksLab Odessa
GeeksLab Odessa

Графические вероятностные модели для принятия решений в проектном управлении Ольга Татаринцева (Data Scientist at Eleks) Как часто вам приходится принимать решения, используя знания в определенной предметной области? На сколько хороши такие решения? А теперь представьте, что вы собрали знания лучших экспертов в предметной области. Похоже, что ваши решения, основанные на этих знаниях, будут куда более взвешенными, не так ли? Мы будем говорить о системе ProjectHealth, которая была построена на основе опыта лучших экспертов в проектном управлении в компании Eleks. Для реализации поставленной задачи была использована графовая вероятностная модель, а именно байесовская сеть, имплементированная на Python. За время работы над проектом мы прошли шаги от извлечения требований, поиска данных и построения модели с нуля до реализации BI дашборда с возможностью углубиться в детали, доходя до сырых данных. Сейчас ProjectHealth экономит большое количество времени для топ менеджмента и ресурсов компании, так как мониторит состояние бизнеса в малейших деталях ежедневно и делает это как настоящий эксперт. Все материалы: http://datascience.in.ua/report2017

1 of 32
Download to read offline
REASONING WITH PROBABILISTIC GRAPHICAL MODELS IN A PROJECT BASED BUSINESS BY OLGA TATARYNTSEVA, DATA SCIENTIST
• Neural networks • Deep learning • Natural language processing • Image processing • Big data • Bioinformatics Today we are NOT about
• Neural networks • Deep learning • Natural language processing • Image processing • Big data • Bioinformatics • Project based business • Company organization • Project management and project success • Applying the DS methods to business problems • Small example • Results we gained Today we are NOT about We are about
Chief Technical officer ... Delivery Director ... PM ... Architects Business Analysts Developers QA Data scientists ... ... ... Simplified structure of delivery organization
Chief Technical officer ... Delivery Director ... PM ... Architects Business Analysts Developers QA Data scientists ... ... ... Simplified structure of delivery organization PROJECT
Scope Resources Quality Time Project management triangle
ProjectTeam Direct contacts with the customer Requirements management Solution design Product development Project budget planning Working on improvements and up-sale opportunities Management of internal activities DeliveryDirector Summarized projects’ budget and cash flow results Calculated CSAT Calculated ESAT General view of project activities Information from reports built by project team CTO Budget flow for vertical Tendencies of overall CSAT Tendencies of overall ESAT Information from reports built by DDs Information transfer
Chief Technical officer Delivery director PM PM … Delivery Director PM PM … ... PM ... 1000+ employees Simplified structure of delivery organization
Don’t panic We have PGM
CTO at ELEKS has a BI Dashboard
• Declarative representation of our understanding of the world • Identifies the variables and their interaction with each other • Sources: • experts’ knowledge • historical data • Representation, inference, and learning • Handles uncertainty Probabilistic graphical models
«As far as the laws of mathematics refer to reality, they are not certain, as far as they are certain, they do not refer to reality» Albert Einstein, 1921 Uncertainty
• Causal inference • Information extraction • Message decoding • Speech recognition • Computer vision • Gene finding • Diagnosis of diseases • Traffic analysis • Fault diagnosis PGM applications Daphne Koller
Observations: None Project health: success = 68.91%
Project health Time: In time Time: Exceed Budget: Match Budget: Exceed Budget: Match Budget: Exceed Success 95 50 40 10 Fail 5 50 60 90 Success = 𝑃 𝑆 𝑖𝑛_𝑡𝑖𝑚𝑒, 𝑚𝑎𝑡𝑐ℎ + 𝑃 𝑆 𝑖𝑛_𝑡𝑖𝑚𝑒, 𝑒𝑥𝑐𝑒𝑒𝑑 + 𝑃 𝑆 𝑒𝑥𝑐𝑒𝑒𝑑, 𝑚𝑎𝑡𝑐ℎ + 𝑆 𝑒𝑥𝑐𝑒𝑒𝑑, 𝑒𝑥𝑐𝑒𝑒𝑑 = = 0.95*0.6659*0.8071 + 0.5*0.6659*0.1929 + 0.4*0.3341*0.8071 + 0.1*0.3341*0.1929 = 68.91% F𝐚𝐢𝐥 = 𝑃 𝐹 𝑖𝑛_𝑡𝑖𝑚𝑒, 𝑚𝑎𝑡𝑐ℎ + 𝑃 𝐹 𝑖𝑛_𝑡𝑖𝑚𝑒, 𝑒𝑥𝑐𝑒𝑒𝑑 + 𝑃 𝐹 𝑒𝑥𝑐𝑒𝑒𝑑, 𝑚𝑎𝑡𝑐ℎ + 𝐹 𝑒𝑥𝑐𝑒𝑒𝑑, 𝑒𝑥𝑐𝑒𝑒𝑑 = 31.09% Node name: Project health States success fail Depends on Time, Project Budget Conditional probability distribution
Bayesian Model with pgmpy c_maturity_cpd = TabularCPD(variable='Customer maturity', variable_card=2, values=[[0.4, 0.6]], evidence=[], evidence_card=[]) ... pr_health_cpd = TabularCPD(variable='Project health', variable_card=2, values=[[0.95, 0.5, 0.4, 0.1], [0.05, 0.5, 0.6, 0.9]], evidence=['Project budget', 'Time schedule'], evidence_card=[2, 2])
Bayesian Model with pgmpy pr_health_cpd = TabularCPD(variable='Project health', variable_card=2, values=[[0.95, 0.5, 0.4, 0.1], [0.05, 0.5, 0.6, 0.9]], evidence=['Project budget', 'Time schedule'], evidence_card=[2, 2]) print pr_model.get_cpds('Project health') +------------------+-----------+------------+------------+-------------+ | Project budget | match | match | exceed | exceed | +------------------+-----------+------------+------------+-------------+ | Time schedule | in_time | exceed | in_time | exceed | +------------------+-----------+------------+------------+-------------+ | success | 0.95 | 0.5 | 0.4 | 0.1 | +------------------+-----------+------------+------------+-------------+ | fail | 0.05 | 0.5 | 0.6 | 0.9 | +------------------+-----------+------------+------------+-------------+
Bayesian Model with pgmpy pr_model = BayesianModel([('Customer maturity', 'Requirements elicitation'), ..., ('Time schedule', 'Project health'), ('Project budget', 'Project health')]) pr_model.add_cpds(c_maturity_cpd, pr_complexity_cpd, time_cpd, req_elicitation_cpd, req_management_cpd, res_availability_cpd, c_budget_cpd, pr_budget_cpd, pr_importance_cpd, pr_health_cpd)
Bayesian Model with pgmpy res = BeliefPropagation(pr_model).query(variables=["Project health"]) print res["Project health"] +------------------+-----------------------+ | Project health | phi(Project health) | |------------------+-----------------------| | success | 0.6891 | | fail | 0.3109 | +------------------+-----------------------+
Observations: Customer maturity: mature Customer budget: large Project health: success = 73.82% (↑4.91%)
Bayesian Model with pgmpy res = BeliefPropagation(pr_model).query( variables=["Project health"], evidence={'Customer budget':1, 'Customer maturity':0}) print res["Project health"] +------------------+-----------------------+ | Project health | phi(Project health) | |------------------+-----------------------| | success | 0.7382 | | fail | 0.2618 | +------------------+-----------------------+
Observations: Customer maturity: mature Customer budget: large Resource availability: available Project complexity: small Project importance: very important Project health: success = 85.73% (↑11.91%)
Observations: Customer maturity: mature Customer budget: large Resource availability: available Project complexity: complex Project importance: very important Project health: success = 79.45% (↓ 6.28%)
Maturity level Lead time Team composition Staffing time Story point time variance Risk management Granularity of stories Stakeholders substitution Acceptance criteria rating RotationsLegal risks
DataScience Lab 2017_Графические вероятностные модели для принятия решений в проектном управлении_Ольга Татаринцева
Events: None Project health: success = 87.36%
Cooperation model: fixed bid Project stage: stable Cycle time: decreasing Number of the opened and reopened bugs: decreasing Finance: fit the company’s KPI values Environment: tools are set and in-use Soft- and hardware: no blocking requests Human resources: no open vacancies Customer satisfaction index: high Number of tasks in the in-progress state: increasing Predicted release date: out of the schedule Hypothetic Project observations
Observations: Listed on previous slide Project health: success = 12.71% (↓ 74.65%)
Latest date: 2017-02-14 Project health: success = 53.18% Major degradation happened on: 2017-02-10 Reason: Outage in Quality Reason: Descending of Project Effectiveness Reason: Cumulating of tasks in Progress Real project
• New communication tool for all levels of the company which is already in use on every day basis • For C-level provides understanding: • Of the department • Of each project in particular • For Project Manager it is an instrument: • For project organization and control • For the Client: • Fair presentation of the project work Benefits
• Improve metrics for measuring project performance • Build even more intuitive dashboard • Introduce the model to our customers • Improve the reflection of the domain by fitting the model to data Future work
Inspired by Technology. Driven by Value. eleks.com

More Related Content

DataScience Lab 2017_Графические вероятностные модели для принятия решений в проектном управлении_Ольга Татаринцева

  • 1. REASONING WITH PROBABILISTIC GRAPHICAL MODELS IN A PROJECT BASED BUSINESS BY OLGA TATARYNTSEVA, DATA SCIENTIST
  • 2. • Neural networks • Deep learning • Natural language processing • Image processing • Big data • Bioinformatics Today we are NOT about
  • 3. • Neural networks • Deep learning • Natural language processing • Image processing • Big data • Bioinformatics • Project based business • Company organization • Project management and project success • Applying the DS methods to business problems • Small example • Results we gained Today we are NOT about We are about
  • 4. Chief Technical officer ... Delivery Director ... PM ... Architects Business Analysts Developers QA Data scientists ... ... ... Simplified structure of delivery organization
  • 5. Chief Technical officer ... Delivery Director ... PM ... Architects Business Analysts Developers QA Data scientists ... ... ... Simplified structure of delivery organization PROJECT
  • 7. ProjectTeam Direct contacts with the customer Requirements management Solution design Product development Project budget planning Working on improvements and up-sale opportunities Management of internal activities DeliveryDirector Summarized projects’ budget and cash flow results Calculated CSAT Calculated ESAT General view of project activities Information from reports built by project team CTO Budget flow for vertical Tendencies of overall CSAT Tendencies of overall ESAT Information from reports built by DDs Information transfer
  • 8. Chief Technical officer Delivery director PM PM … Delivery Director PM PM … ... PM ... 1000+ employees Simplified structure of delivery organization
  • 10. CTO at ELEKS has a BI Dashboard
  • 11. • Declarative representation of our understanding of the world • Identifies the variables and their interaction with each other • Sources: • experts’ knowledge • historical data • Representation, inference, and learning • Handles uncertainty Probabilistic graphical models
  • 12. «As far as the laws of mathematics refer to reality, they are not certain, as far as they are certain, they do not refer to reality» Albert Einstein, 1921 Uncertainty
  • 13. • Causal inference • Information extraction • Message decoding • Speech recognition • Computer vision • Gene finding • Diagnosis of diseases • Traffic analysis • Fault diagnosis PGM applications Daphne Koller
  • 15. Project health Time: In time Time: Exceed Budget: Match Budget: Exceed Budget: Match Budget: Exceed Success 95 50 40 10 Fail 5 50 60 90 Success = 𝑃 𝑆 𝑖𝑛_𝑡𝑖𝑚𝑒, 𝑚𝑎𝑡𝑐ℎ + 𝑃 𝑆 𝑖𝑛_𝑡𝑖𝑚𝑒, 𝑒𝑥𝑐𝑒𝑒𝑑 + 𝑃 𝑆 𝑒𝑥𝑐𝑒𝑒𝑑, 𝑚𝑎𝑡𝑐ℎ + 𝑆 𝑒𝑥𝑐𝑒𝑒𝑑, 𝑒𝑥𝑐𝑒𝑒𝑑 = = 0.95*0.6659*0.8071 + 0.5*0.6659*0.1929 + 0.4*0.3341*0.8071 + 0.1*0.3341*0.1929 = 68.91% F𝐚𝐢𝐥 = 𝑃 𝐹 𝑖𝑛_𝑡𝑖𝑚𝑒, 𝑚𝑎𝑡𝑐ℎ + 𝑃 𝐹 𝑖𝑛_𝑡𝑖𝑚𝑒, 𝑒𝑥𝑐𝑒𝑒𝑑 + 𝑃 𝐹 𝑒𝑥𝑐𝑒𝑒𝑑, 𝑚𝑎𝑡𝑐ℎ + 𝐹 𝑒𝑥𝑐𝑒𝑒𝑑, 𝑒𝑥𝑐𝑒𝑒𝑑 = 31.09% Node name: Project health States success fail Depends on Time, Project Budget Conditional probability distribution
  • 16. Bayesian Model with pgmpy c_maturity_cpd = TabularCPD(variable='Customer maturity', variable_card=2, values=[[0.4, 0.6]], evidence=[], evidence_card=[]) ... pr_health_cpd = TabularCPD(variable='Project health', variable_card=2, values=[[0.95, 0.5, 0.4, 0.1], [0.05, 0.5, 0.6, 0.9]], evidence=['Project budget', 'Time schedule'], evidence_card=[2, 2])
  • 17. Bayesian Model with pgmpy pr_health_cpd = TabularCPD(variable='Project health', variable_card=2, values=[[0.95, 0.5, 0.4, 0.1], [0.05, 0.5, 0.6, 0.9]], evidence=['Project budget', 'Time schedule'], evidence_card=[2, 2]) print pr_model.get_cpds('Project health') +------------------+-----------+------------+------------+-------------+ | Project budget | match | match | exceed | exceed | +------------------+-----------+------------+------------+-------------+ | Time schedule | in_time | exceed | in_time | exceed | +------------------+-----------+------------+------------+-------------+ | success | 0.95 | 0.5 | 0.4 | 0.1 | +------------------+-----------+------------+------------+-------------+ | fail | 0.05 | 0.5 | 0.6 | 0.9 | +------------------+-----------+------------+------------+-------------+
  • 18. Bayesian Model with pgmpy pr_model = BayesianModel([('Customer maturity', 'Requirements elicitation'), ..., ('Time schedule', 'Project health'), ('Project budget', 'Project health')]) pr_model.add_cpds(c_maturity_cpd, pr_complexity_cpd, time_cpd, req_elicitation_cpd, req_management_cpd, res_availability_cpd, c_budget_cpd, pr_budget_cpd, pr_importance_cpd, pr_health_cpd)
  • 19. Bayesian Model with pgmpy res = BeliefPropagation(pr_model).query(variables=["Project health"]) print res["Project health"] +------------------+-----------------------+ | Project health | phi(Project health) | |------------------+-----------------------| | success | 0.6891 | | fail | 0.3109 | +------------------+-----------------------+
  • 20. Observations: Customer maturity: mature Customer budget: large Project health: success = 73.82% (↑4.91%)
  • 21. Bayesian Model with pgmpy res = BeliefPropagation(pr_model).query( variables=["Project health"], evidence={'Customer budget':1, 'Customer maturity':0}) print res["Project health"] +------------------+-----------------------+ | Project health | phi(Project health) | |------------------+-----------------------| | success | 0.7382 | | fail | 0.2618 | +------------------+-----------------------+
  • 22. Observations: Customer maturity: mature Customer budget: large Resource availability: available Project complexity: small Project importance: very important Project health: success = 85.73% (↑11.91%)
  • 23. Observations: Customer maturity: mature Customer budget: large Resource availability: available Project complexity: complex Project importance: very important Project health: success = 79.45% (↓ 6.28%)
  • 24. Maturity level Lead time Team composition Staffing time Story point time variance Risk management Granularity of stories Stakeholders substitution Acceptance criteria rating RotationsLegal risks
  • 27. Cooperation model: fixed bid Project stage: stable Cycle time: decreasing Number of the opened and reopened bugs: decreasing Finance: fit the company’s KPI values Environment: tools are set and in-use Soft- and hardware: no blocking requests Human resources: no open vacancies Customer satisfaction index: high Number of tasks in the in-progress state: increasing Predicted release date: out of the schedule Hypothetic Project observations
  • 28. Observations: Listed on previous slide Project health: success = 12.71% (↓ 74.65%)
  • 29. Latest date: 2017-02-14 Project health: success = 53.18% Major degradation happened on: 2017-02-10 Reason: Outage in Quality Reason: Descending of Project Effectiveness Reason: Cumulating of tasks in Progress Real project
  • 30. • New communication tool for all levels of the company which is already in use on every day basis • For C-level provides understanding: • Of the department • Of each project in particular • For Project Manager it is an instrument: • For project organization and control • For the Client: • Fair presentation of the project work Benefits
  • 31. • Improve metrics for measuring project performance • Build even more intuitive dashboard • Introduce the model to our customers • Improve the reflection of the domain by fitting the model to data Future work
  • 32. Inspired by Technology. Driven by Value. eleks.com