OpenTalks.AI - Эмели Драль, PA и RS - обзор текущего применения в бизнесе​

Transcript

1. Predictive analytics and recommendation systems: overview of business applications Emeli Dral Data Science Expert
2. About me • Co-founder Mechanica AI • Ex Chief Data Scientist at Yandex Data Factory • Co-founder of Data Mining in Action, largest offline data science course in Russia • Co-author of two Coursera specializations in data science with > 50K students • Lecturer at Harbour.Space University, GSOM MBA https://www.linkedin.com/in/emelidral ?! 50+ Industrial applications of machine learning
3. vs Predictive analysis Recommender systems
4. Technologies Future trends Interpretable predictions for robust decision support Optimization of key business KPIs, full process automation Question What will happen in the future? Which action to take? Parametric models Machine learning Physical models Statistical models Expert systems …. Predictions VS recommendations
5. Trends in predictive analysis
6. Trend 1: from standard use cases EXISTING BUSINESS PROCESS BETTER PREDICTION QUALITY + Image source: https://getwallpapersinhd.com/images/medium/a-hard_desicion-969070.jpg
7. Trend 1: to new applications and product features NEW PREDICTIONS + NEW BUSINESS PROCESS / FEATURE
8. Trend 2: from black box WHY? WHY? WHY?!! W H Y? WHY?!!
9. Trend 2: to explainable models root cause analysis ≠ explainable model How the model works? How the model made this specific prediction? Which factors are used?
10. Trend 2: explainable models Predict which steel coils are likely to have a defect of each specific group For highly suspicious coils – indicate top factors in prediction
11. Trend 3: from straightforward predictions So what? Answer: 42
12. Trend 3: to scenario analysis What if?
13. Trend 3: scenario analysis Predict which steel coils are likely to have a defect of each specific group …for each possible production route the engineer has available Route 1: 42% Route 2: 60% …
14. Summary: trends in predictive analysis New use cases + Explainability Scenario analysis + Making predictive models usable by business and domain experts
15. Trends in recommendation systems
16. Trend 1: from item/content recommendation
17. Trend 1: to process/decision optimization Traditional decision support (parametric, rule-based models, technological maps…) Now with machine learning! The “Sniper” service was introduced as a pilot program in the production process, optimizing the consumption of ferroalloys and additional materials in steelmaking.
18. Trend 2: hybrid models machine learning + human in the loop machine learning + first-principle models Content moderation Industrial production optimization
19. Trend 3: from recommendations to automation
20. Summary: key trends
21. Key trends 1. From standard to novel use cases and product features based on ML 2. From black box to explainable AI 3. Scenario analysis with ML 4. From item/content recommendations to process/decision optimization 5. Hybrid applications based on combination of ML with other tools 6. From recommendations to automation of action
22. Questions? Emeli Dral