Employing traditional approaches to analyzing customer behavior graphs and event sequences requires data simplifications, generalizations, and segmentations that severely degrade prediction accuracy and can lead to the loss of valuable information.
But there’s an alternative approach that retains the fidelity of customer profile data, the full sequence of events data, and enables powerful business intelligence and predictive analytics. In this webcast, join Apigee’s Joy Thomas, Chief Scientist, and Sanjeev Srivastav, VP Data Strategy, as they explore the superiority of new methods of behavior graph analysis over the simplifications required for traditional data storage and classical predictive algorithms.
Join to learn:
• How shortcomings of traditional approaches make it difficult to build and analyze complex behavior graphs
• Why GRASP—graph and sequence processing technology on Hadoop presents a different and more effective approach to behavior graph analysis
• How descriptive analytics using GRASP uncover hidden patterns in historical customer journey data
• How Predictive analytics that use behavior graphs and Bayesian algorithms, along with machine learning, ensure model performance over time
Download Video: http://youtu.be/CwxvlgW9aZY
Download Podcast: https://soundcloud.com/apigee/analyzing-complex-behavior-graphs-in-hadoop-at-scale
At Apigee,
Sanjeev leads the Data Strategy team
Joy leads the Data Science team
Multichannel interaction between customers and companies results in event streams that can be stored in the GRASP data structure, which is built on Hadoop HDFS
Analysis of large event stream data sets provides detailed patterns of customer behavior
Customer behavior, represented by sequences of event paths, is a very effective means of building predictive models
Customers engage with companies through email, web sites, mobile apps, social networks, telephone and physical stores. Such engagement involves activities that can be mapped on a timeline, with consideration of the duration of engagement for each activity. External contextual data such as location and weather can also be recorded or used as a real-time input.
From a customer’s point of view, as they interact with your business in the digital world, they go through a journey that involves a number of interactions sequenced in by time and spanning multiple channels. They might start their journey on one channel, then move to another, and so on. For example, Sally might receive an offer via email. She views product details from her laptop. She then tweets about it to check if any of her friends like it. Later she might be near your store, and it appears to be a nice, sunny day, so she pulls out her smartphone, checks directions and walks in. However, she doesn’t buy anything, but comes home and follows up with more questions on online chat or call center. And the journey continues.
As customers get more accustomed to digital interactions from Amazon, facebook, google, and more, their expectations of how they are treated and served is also increasing. You’ve living through this. For example, We expect a journey and experiences that are easy, convenient, relevant, and contextual, and individualized to our specific needs. If we don’t find it, we tune out, we start looking at other brands, and so on. So for a business to be relevant to their customers, they have to understand the customer journey so they can optimize it to meet the needs of their customers.
Insights generates the customer journey for each customer
Insights uses GRASP to provide a comprehensive understanding of customer behavior across channels and gives you a comprehensive understanding of customer behavior across your entire customer base. With this understanding, so can discover common interactions & influences that lead to successful customer journeys.
Now you can answer these questions:
What journeys are most successful?
What actions did customers take after receiving an offer?
What actions did customers take before canceling service?
Where are you failing to convert? What might be causing that?
What can you change to improve conversion?
How do you allocate sales credit or contribution across your channels?
A behavior graph captures the sequence and duration of events, while a social graph is used for storing neighborhood relationships.
Insights starts with all your customer interactions via web, mobile, store, call center, and other channels….and applies GRASP - Graph and Sequence Processing, a unique time sequenced graph analytics on Hadoop…
Insights enables you to analyze the customer journey using a unique big data graph structure that is purpose built to help detect hidden patterns by finding strong signals in a sea of weak signals. Insights starts with fine grained event data - individual interactions that occur at specific points in time across different channels and devices – each product page viewed on web, each offer received on mobile, each call into the call center. Each interactions when viewed in isolation might not signal much. However, when you stitch them together using the unique Big Data structure that Insights offers – a temporal graph structure built directly on Hadoop that is built to analyze such time based sequence of interactions, you start to identify hidden patterns and relationships. In other words, Insights is able to looks at all the weak and strong signals to find strong signals that otherwise would lay dormant. Unlike social graphs which connect nodes based on relationships, GRASP connects events in time – it is ideally suited for analyzing customer interactions. This approach is also different from traditional predictive analytics which summarizes data into a tabular format – you lose precision by summarizing the data.
Key points:
Focus is on Flow & Behavior
Flexible data model : Nodes, Users, Events are heterogeneous in the sense of not requiring a homogeneous attribute structure and permitting easy addition or dropping of attributes. For example, the attributes for an event may depend on its type and new types/attributes may be easily added.
Customers data may need to be transformed: Customer may have modeled data differently and will need to be transformed to this model. For example, customer’s events may contain the structure of dynamically generated content (attributes that are not metrics) which will end up corresponding to nodes in the model.
Sequential model of actions for a user : Concurrent actions are forced into a sequence.
Scale - Implementation at scale that comes with analysis of consumer behavior
Characteristics
Represents flow & behavior of all users
Automated construction from event logs
Information preserving
Aggregated representation
Permits drill-down
Useful for reasoning about traffic 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
Powerful functionality - multiple applications
Patterns for Machine Learning
Event sequences with time distances
Profiles and Sequence combinations
Social Scoring
Exploit graph based relationships
Segmentation & Profiling
Flows/Metrics for segments of consumers
Profiles of flows at nodes/edges
Path/Tree/Funnel Analysis
Flows by source or destination
Flow analysis for paths
Flows & Metrics
Consumer/Session flows on nodes/edges
Behavior metrics on nodes/edges
Insights can help business users visualize the customer journey across all interactions
Conduct ad-hoc analysis of customer journey
View customer path across all interactions
Eg: You can see customers who visited your website, then purchased a product, then called customer service, and then cancelled their order
Visualize multiple paths – roll back or roll forward
Visualize any path in the data
Start with a specific event and view future events along the path
E.g.: start with customers who visited website and determine the number who purchased that product
Start with a specific event and view preceding events along that path
E.g.: start with customers who canceled their contract, and determine the number who called customer service prior to canceling their contract
Drill down into analyze additional details
View additional details for each node
Insights can help data scientists build sophisticated models faster using R
Insights provides a R-based modeling environment for data scientists & statisticians
Data Scientists can build models faster by using R, a popular tool among data scientists
Data Scientists can Create, score, and test new models that leverage machine learning on GRASP
The models they build can incorporate unstructured and structured data. The ability to include unstructured data allows them to models that result in higher precision
They can experiment with different model configurations
Background: Model Configuration = data + parameters
Insights provides support for training and test environments
They can measure model performance (lift curve)