Digital and data journey demystified: how it all works
- 1. Digital & data journey demystified How It All Works AI & ML: Data-driven Strategies for the Power & Utilities Industry (& Crisis Impact) Michal Hodinka, 4.6.2020
- 3. A recent Gartner user survey shows that, while large IT projects are more likely to fail than small projects, around half of all project failures… https://www.gartner.com/en/documents/2034616/survey-shows-why-projects-fail
- 4. The Machine Learning is based on five pillars And two areas that you should evaluate when you build a machine learning workload Machine Learning Stack Phases of Machine Learning Workloads 4innogy · Název prezentace · DD měsíc RRRR
- 5. The end-to-end machine learning process 5
- 6. Variety of data science use cases are applied to retail topics within innogy Predicts which customers will not pay their bills on time or at all. Predicts if a customer will end his contract in the near future. Calculates the value of a customer, contract or lead Predicts how long a customer will ‘survive’ (= remain a customer) Predicts how much gas/electricity the customer will use in a certain period Predicts or calculates the uplift/conversion of a certain campaign time
- 7. High-level data processing architecture iDATA: LAB – Simpler, more automated ML on AWS Data catalog iDataHUB Data Science & Analytics Leonardo Lancelot S3 NetApp Redshift ETL Extract, Transform and Load Cisco IPCC SAP Analytics SAP BW4HANA SAP BW on HANA SAP Analytics Cloud SAP ESM IS-U, CRM SAP HR Reporting DataOps Retail Apps
- 8. CI/CD: Code processing architecture iDATA: LAB – Simpler, more automated ML on AWS Infrastructure as a code https://blog.alterway.fr/cetait-cette-semaine-sur-aws-lundi-7-mai-2018.html
- 9. Preventive retention project I Machine Learning & Multi-Armed Testing Cover a obrázek 1 Step 1: Identification of Customers@Risk • Which customer is more likely to churn? • What affects this risk? • Testing different algorithms (Neural Networks, Decision Trees) • Out-of-sample validation Pilot confirmed that Customers@Risk model is accurate in identifying future churn customers 9innogy · Název prezentace · DD měsíc RRRR AI Customer data Customers@Risk
- 10. Preventive retention project II Machine Learning & Multi-Armed Testing Cover a obrázek 1 Step 2: Multi-armed testing of campaigns and Next Best Offer recommendation • Which retention offering & channel works for which customer? • Optimization across possible offerings & campaigns for each customer First A/B tests confirmed that a reduction in churn coupled with reduction in in-bound calls was observed for some campaigns Next Best Offer to be deployed 10innogy · Název prezentace · DD měsíc RRRR Offer or treatment? Channel? What Message? Customers@Risk + Multi-armed testing
- 11. 1 Enable agility through the availability of high data quality datasets 2 Start simple and evolve through experiments 3 Decouple model training and evaluation from model hosting 4 Detect data drift 5 Automate training and evaluation pipeline 6 Prefer higher abstractions to accelerate outcomes Source: AWS Well-Architectured Framework
- 12. Thank you innogy · Název prezentace · DD měsíc RRRR