Deep-Dive: Predicting Customer Behavior with Apigee Insights
7 likes3,197 views
The document discusses using Apigee Insights to enable personalized experiences through predictive analytics. Insights uses a GRASP technology to analyze sequential customer behavior patterns at scale from big data sources. This allows building predictive models to anticipate customer needs and adapt interactions in real-time across channels. The platform provides segmentation, predictions, and an interaction layer to deliver the right offer to the right customer at the right time.
Deep-Dive: Predicting Customer Behavior with Apigee Insights
- 1. Deep-Dive: Predicting Customer Behavior with Apigee Insights Anticipate and adapt to each customer’s journey
- 3. Today’s Speakers Yong Kim Alan Ho
- 5. Consumers will tune you out if you are irrelevant! 5
- 6. A push toward personalization 6 Vs.
- 7. A perspective from Amazon 7
- 8. The complete personalization solution Segmentation Predictive Analytics on Big Data Real Time Interaction Platform Personalization: Right Person Right Offer Right Time
- 9. The Promise of Personalization 9
- 10. Reality for Most Enterprises 10 Estimated 70% to 75% of enterprises struggle to deliver personalized experience
- 11. Insights platform for personalization 11 Consumer profile Consumer behavior • Targeting via Self-Service Behavior Segmentation • Behavior Predictions at Scale • Real-Time Interaction Layer Offers Shopping Purchases Usage Reviews Social + Right Offer + Right Customer + Right Time
- 12. Demo: Real time personalization web and mobile app 12
- 13. Apigee Platform for Developing and Deploying Personalized Apps 13 Big Data Analytics Integrated Platform for Intelligent Apps Insights API BaaS + Edge • What happened? • Why did it happen? • What will happen next? • What is happening now? • Where is it happening? • How should I interact? • At scale • Real time • Multiple channels and devices Into Action
- 14. Insights Demo: Platform Overview 14
- 15. Past behavior is best predictor of future behavior: Use past purchase transactions with contextual information to provide most relevant results for customer up-sell. Apigee Insights Approach
- 16. Insights Demo: Data to Recommendation API 16 Real Time Interaction • Right Offer • Right Member • Right Time Member ID Location Context /Recommendations /MerchantOffers API BaaS Node.js
- 17. The Value Chain, Enhanced by Machine Learning and Human Discovery 17 Developer API API Team Backend Predictive Analytics Hadoop Data Warehouse AppApp Data Scientist/ Analyst
- 18. GRASP: Graph and Sequence Processing on Hadoop 18 Time-sequenced graph analytics on Hadoop
- 19. How: Insights GRASP technology ? Innovative machine learning approach for automatically detecting complex, hidden patterns in consumer behavior at scale
- 20. Our View of Big Data 20 Sequence of interactions across time, channel, and location. Behavior Data: ~95% of Big Data Profile Data: ~5% of Big Data (Age, Income, Gender, etc.)
- 21. Behavior data is complex 21 Behavior graph visualization from a web log http://www.cnaa.acad.md/en/
- 22. Most models are mainly profile based • User behavior is summarized as a set of features that are aggregated as frequencies and broken out into a set of dummy variables • Order and sequential patterns are limited at best, and most often not considered 22
- 23. Challenge of Tool Bias and Feature Selection Bias 23 Traditional tools/approach forces summarization and is craft-dependent • Mainly rely on profile data • Summarize behavior as set of features to fit into columns and rows
- 24. Challenge: Are you answering the right question? 24 What product will this customer purchase next? • What product will this customer also purchase? • What is the likelihood to purchase this product? (repeat for each product, or product category) Traditional approaches require modifying the business question and extending existing algorithms ?
- 25. 25 Insights 2 2 1 1 2 2 1 1 Without Insights Uncover sequential patterns that help predict what will happen next. Sequential patterns are lost and hard to predict what will happen next. Challenge of losing sequence of interactions?
- 26. Businesses need tools for analyzing behavior (event sequence) data • Discovering behavior patterns is very painful with traditional relational data structures. • Data scientists at some of the largest companies such as Expedia, AT&T, Pearson, Magazine Luiza, and Telstra agree. 26
- 27. Making Sense of Event Stream, Profile, and Unstructured Data 27 Text
- 28. Event and Profile Datasets Joined by Common User ID 28 Events Profile
- 29. Google Analytics Data Example 1) event_add -- All “Add to bag” events 2) event_remove -- All “Remove from bag” events 3) event_purchase -- All “Purchased product” events 4) event_viewprod -- All “Viewed product” events 5) event_other -- All other event hits not included in 1-4 6) item -- All items included in a transaction 7) page -- All page views 8) transaction -- All transaction events 9) social -- All shares on social media 10) visitor_profile -- Attributes of each visitor 29
- 30. event_viewprod 30 fullvisitoid,visitnumber,hitnumber,eventtime,country,hittype,eventaction,productid,category ,subcategory 179804623949526830,1,3,2014-05-21 00:46:34.974,us,e,Viewed product,37917731,Women,Sale 179885841781101277,1,5,2014-05-21 02:44:21.515,us,e,Viewed product,44985721,Women,Sale 179885841781101277,1,8,2014-05-21 02:45:13.181,us,e,Viewed product,44992241,Women,Sale 179885841781101277,1,11,2014-05-21 02:45:55.790,us,e,Viewed product,44985551,Women,Sale 179885841781101277,1,14,2014-05-21 02:46:27.730,us,e,Viewed product,44986041,Women,Sale 179885841781101277,1,17,2014-05-21 02:47:47.738,us,e,Viewed product,39047241,Women,Sale 179885841781101277,1,20,2014-05-21 02:49:52.539,us,e,Viewed product,39052051,Women,Sale 179885841781101277,1,23,2014-05-21 02:50:36.782,us,e,Viewed product,39044811,Women,Sale 179885841781101277,1,26,2014-05-21 02:57:23.268,us,e,Viewed product,39047951,Women,Sale 179885841781101277,1,29,2014-05-21 02:59:28.148,us,e,Viewed product,39056761,Women,Sale
- 31. Hotel Search Data Example 31
- 32. Retailer API Data Example 32
- 33. GRASP: Graph Database for Event Sequence On Hadoop Consumers act on nodes in a temporal sequence of events 1 2 4 3 3 4 0 0 CONSUMER PROFILE ConsumerID: U56 Gender: M Geo: San Francisco Interests: Bikes, Fashion CONSUMER PROFILE ConsumerID: U57 Gender: F Interests: News, Finance Age: 35-40 NODE PROFILE Type: Content PageID: P100 Category: Product Review SubCat: Mountain Bike NODE PROFILE Type: Creative ID: Creative95 Category: VideoAd Advertiser: BikePros EVENT Type: PageView ConsumerID: U56 PageID: P100 TimeSpent: 180 seconds Scrolls: 3 EVENT Type: AdView ConsumerD: U56 AdID: Creative95 PlayTime: 30 sec Rewinds: 1 Insights uses event time stamps to build a sequential view of all customer interactions across data sources.
- 34. GRASP: Aggregated Behavior Graph (ABG) 0 1 3 2 4 0 1 2 4 3 3 4 Impressions: 1 TimeSpent: 20 Clicks: 1 0 0 Impressions: 4 TimeSpent: 10 Clicks: 0 Impressions: 5 TimeSpent: 30 Clicks: 1 Combine Characteristics • Represents flow & behavior of all Consumers • Analysis of customer journeys • Predictive algorithms
- 35. Identify common interactions and influences 35
- 36. Machine learning automates science and removes bias 36 Automated feature selection from common behaviors (Micro-segments) • Drastically reduces time/effort of feature selection • Natural human bias removed from selection process • Machine Learning model, tuned to generalize well in production • Optimization Algorithms can match consumers with products/offers to maximize a metric (e.g. Margin) Micro-segments Predictors
- 37. Insights Streamlined Behavior Modeling Workflow 37 Data Extract Model Training Model Validation Extract profile features Join disparate event data Explore event sequence patterns Identify significant behavior patterns Summarize events as frequencies Data Extract Model Training Model Validation Extract profile features Identify event data Repeat for each product Traditional Workflow Insights Workflow Weeks Days
- 38. Behavior modeling for analysts with limited data science expertise 38 • Easy to use multi channel path exploration and visualization Replaces need to create complex data cubes • Simplified behavior based segmentation Replaces need for complex SQL like queries • Simplified model scripts in R Replaces need for machine learning scripting language expertise (Scala, Python, R) • Simplified model deployment Reduces need for engineering support
- 39. Deployed on modern infrastructure for delivering personalized real time interactions at scale 39 Node.js Controller Node.js Controller Node.js Controller Targeting Models Rec. Models Customer Journey GRASP Segmentation Speed Layer (Edge) Batch Layer (Insights) /predictions /activities (Push) / notifications Graph /datastore /segments
- 40. Insights Online Predictive Analytics Processing 40 • Customer Journey Analytics • GRASP Models • Recommendations • Targeting Storm Spark Kafka Insights Batch Processing Stream/Near-line Processing Component Algorithms • Fallback logic • Ensemble logic • Context injection • Rule based predictive models • Summary statistics API BaaS • Scores • Meta data • User information • Select transaction data Online Processing Layer Cassandra Node.js• Profile based models • Transaction data Other Batch Processing Mobile Web Workflow integration Apps APIs
- 41. Insights Architecture Customer Data R Data Scientist queries Graph Query Manager Business User Segments Manager Scores Propensity Upgrade 10% Off Churn User 1 0.72 0.68 0.33 User 2 0.56 0.23 0.55 User 3 0.32 0.45 0.67 User 4 0.20 0.32 0.18 User 5 0.44 0.69 0.22 Business User Real Time Serving Layer Analytics Engine Modeling Workbench Context
- 42. Summary of Benefits of Insights + Edge + API BaaS Edge: Integrated platform for data scientists and developers 42 • Rapid intelligent application development • Developer friendly experience • Deploy model output into production with limited engineering resources • Real time access to model output at scale API BaaS: Cassandra data store Insights: GRASP • Understand customer journey • Build behavior and profile based predictive models
- 43. Early bird ends May 31st! Use code: WEBCAST15 for 15% off
- 44. Thank you! Q&A 44 Time-sequenced graph analytics on Hadoop