©2015 Apigee Corp. All Rights Reserved. Preserving signal in customer journeys Joy Thomas, Apigee Jagdish Chand, Visa.

©2015 Apigee Corp. All Rights Reserved.

Preserving signal in customer journeysJoy Thomas, ApigeeJagdish Chand, Visa

©2015 Apigee Corp. All Rights Reserved. 2

Overview

• Customers journeys and event data• Customer Behavior Graph• Queries on Behavior Graphs• Predictive models on behavior graphs

© 2014 Apigee Confidential – All Rights Reserved

Customer View: A journey

3

Consumers interact with enterprises through multiple channels at multiple points of time

Each of these interactions is an event with a timestamp and the sequence of interactions is important

© 2014 Apigee Confidential – All Rights Reserved 4

A graphical structure can identify common interactions and influences

Common Interactions & InfluencesCustomer Journey

Customer behavior graphs vs. social graphs

5

Behavior Graph• Sequence of events:

– Actions experienced and taken

Social Graph• Links between people & activities

– At a point in time

Behavior graph

Social graph

©2015 Apigee Corp. All Rights Reserved.

Model for User Behavior

Users act on nodes in a temporal sequence of events

1

3 52

2 50

0

USER PROFILEUserID: U56Gender: MGeo: San FranciscoInterests: Bikes, Fashion

USER PROFILEUserID: U57Gender: FInterests: News, FinanceAge: 35-40

NODE PROFILEType: ContentPageID: P100Category: Product ReviewSubCat: Mountain Bike

NODE PROFILEType: CreativeID: Creative95Category: VideoAdAdvertiser: BikePros

EVENTType: PageViewUserID: U56PageID: P100TimeSpent: 180 seconds Scrolls: 3

EVENTType: AdViewUserID: U56AdID: Creative95PlayTime: 30 secRewinds: 1

©2015 Apigee Corp. All Rights Reserved.

Aggregated Behavior Graph (ABG)

0

1

2

3

5

0

1

3 52

2 5Impressions: 1TimeSpent: 20Clicks: 1

0

0Impressions: 4TimeSpent: 10Clicks: 0

Impressions: 5TimeSpent: 30Clicks: 1

Combine

Characteristics

• Represents flow & behavior of all users

• Automated construction from event streams

• Information preserving• Aggregated representation• Permits drill-down

• Useful for reasoning about customer flows• Count unique users at node/edge• Aggregate metrics at nodes/edges• Measure drop-offs on a path (funnel)• Profile traffic at a node or edge• Analyze flows for user segments

©2015 Apigee Corp. All Rights Reserved. 8

Examples of queries on Behavior Graph

• Count the number of users who went from A -> B -> C -> D

• Find the distribution of (Age, Gender) for the people who took the path P-> Q ->R

• Of all the females in California who went from C to D, what are the most likely nodes that they are likely to visit next

• Of the people who bought a computer 3 months ago and received an email offer for a discounted printer 1 month ago, what fraction of them have bought printers in the last month

• All these queries would be painful to express in SQL on a large event table

©2015 Apigee Corp. All Rights Reserved. 9

Predictive Analytics on Behavior Graphs

Past behavior of consumers is the best predictor of future actions

• The behavior graph allows one to search for patterns of consumer behavior that are correlated with responses of interest

• Using the patterns we can build a Bayesian model to predict what users will do next

• Use the predictive model for recommendations, targeting and churn prediction

©2015 Apigee Corp. All Rights Reserved. 10

Comparison with other Machine Learning algorithms • Most machine learning algorithms assume that the training

data for a learning algorithm is in a form of a large table of examples, with responses in one column, and features in other columns, e.g. Logistic Regression, Random Forest, etc.

• These algorithms are designed for profile attributes such as age, gender, country, etc.

• To handle event data, the data scientist typically creates aggregate features out of the event data (e.g. total purchases over the last year, total purchases over the last month, etc.)

• The behavior graph allows the data scientist to automatically search over a large space of aggregates to use in the predictive model

©2015 Apigee Corp. All Rights Reserved. 11

Summary

• Event data should be treated differently from profile data

• A graphical data structure designed for event data can efficiently answer queries on event based patterns

• Event based patterns can be used to build predictive models for targeting, recommendations and churn prediction

• There is a need for a common query language to express queries for event data

Thank you