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.
8. The complete personalization solution
Segmentation
Predictive Analytics on
Big Data
Real Time Interaction
Platform
Personalization:
Right Person
Right Offer
Right Time
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
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
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
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
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