Realtime Next Best Action Recommendation at Scale

Transcript

1. Realtime Next Best Action Recommendation at Scale Kexin Xie, PhD Salesforce
2. We Are All Connected Welcome to the Fourth Industrial Revolution
3. “Artificial Intelligence is the new Electricity” -- Andrew Ng
4. “The performance of learners can benefit significantly from much larger training sets” -- Banko and Brill ’01 [Proceeding of ACL '01 Proceedings of the 39th Annual Meeting on Association for Computational Linguistics Pages 26-33]
5. ...
6. Data 4.3 M user match queries 8.8 M data capture events 1.6 M page views 0.7 M ad impressions 3 B email sends Every Minute Every Day User Attributes Page Attributes Emails SMS Push Notifications Tweets Facebook Updates User Segments ...
7. Einstein for Marketing 13 GA Features +6B predictions/day Kathryn Frederick EVP, Growth & Insights Ticketmaster Einstein Recommendations Einstein Splits Einstein Engagement Frequency Optimize Performance & Productivity Einstein Engagement Scoring Einstein Vision for Social Studio Einstein Social Insights Einstein Journey Insights Einstein Segmentation Einstein Cross-Device Identity Management Einstein Content Tagging NEW Einstein Lookalikes Personalize Every Interaction Einstein Send Time Optimization NEW NEW Einstein Marketing Insights (Datorama) NEW Know, personalize, and engage across every consumer touchpoint
8. personalization
9. INTERACTION ACTION ACTIVITY
10. https://blog.datarobot.com/delivering-next-best-action-with-artificial-intelligence awareness consideration purchase service loyalty Touchpoint 1 Touchpoint 2 Touchpoint 3 Touchpoint 4 Touchpoint 5 Touchpoint 6 Touchpoint 7 Touchpoint 8 Touchpoint 9 Touchpoint 10 Touchpoint 11
11. Einstein works within Interaction Studio to recommend next best action for customer engagement Reporting on insights, trends and impact of using AI-driven recommendations. Einstein for Interaction Studio Automate Next Best Action (NBA)
12. Analyze User Behavior Explore Action Space Deploy and Apply Try Different Options
13. User Context Action Rewards Maximized => + : Problem Setting
14. Incomplete Data - Feedback for other predictions are not available Sampling Bias - Predictions favored by the historical algorithm are over-represented
15. Analyze User Behavior Explore Action Space Deploy and Apply Try Different Options
16. Context Explore Exploit Learn
17. https://lilianweng.github.io/lil-log/2018/01/23/the-multi-armed-bandit-problem-and-its-solutions.html 50% 70% 25% 45% Which machine to pick next? Machine 1 Machine 2 Machine 3 Machine 4 Reward probabilities are unknown
18. Next Best Action - Contextual MAB Environment NBA Engine [Activity, Context] [Activity, Context] [Activity, Context] Action Action Action Action 1 Action 2 Action 3
19. Next Best Action - Contextual MAB Challenge 1 - how to reward multiple objectives? Single Ticket Purchase Increase Engagement Stage Progression One mixed Goal? How??? Seasonale Ticket Purchase
20. Next Best Action - Contextual MAB Challenge 2 - how to minimize counterfactual regret? Goal B Goal C Goal A
21. Next Best Action - Advanced Approach Propensity Scorer Hybrid MAB Offline Learning Online Learning Intent Predictor Policy Selector Contextual Action Engine
22. Next Best Action - Advanced Approach Profile Affinities Behavior Patterns Intent Propensity Scorer Input Multi-class Classification (I-Prop) Hybrid Contextual Multi-armed Bandit (H-Bandit) Output
23. Next Best Action - Evaluation train predict train predict train predict adjustable window Rolling Evaluation Predictions Hits Observations weighted confidence score + expected overall gain expected overall reward observed overall reward
24. explore exploit stable reactive