Vincent Morder Analytics Manager, Loyalty New Zealand Insights on Loyalty’s Approach To Big Data SUNZ February 2013
Oct 19, 2014
Vincent Morder
Analytics Manager, Loyalty New Zealand
Insights on Loyalty’s Approach
To Big Data
SUNZ February 2013
Loyalty Vision: To create, maintain and
motivate loyal customers for our Participants
End to End Marketing
4
History of data and analytics In the beginning….
• Data • Small, manual, fixed
• Systems • Static, relational
• Analytics • Sophistocated statistics and mathematics
• Retrospective
• Hand- crafted variables
• Static view
• Long time lags
5
History of data and analytics Over time….
• Data
• Bigger data, coming from various systems, more timely
• Systems
• Operational systems
• More storage required
• Relational databases
• Computing power increased
• Analytics
• Applying stats to business, banks, retail, telecommunications, etc…
• Same statistical principles applied, but now on hormones with software
• Sample size no longer an issue.
6
Today: Big Data
• Sheer magnitude of data captured by digital world:
• Web
• Social
• Free form text
• Retail Transactions
• Mobile device activity
• Software logs
• Cannot be managed by traditional data management tools
• Unstructured to a large degree
• Data is distributed, so difficult to perform traditional queries for analysis
• Applications and needs are for real time.
7
Big sets of Data
• Historically loyalty has a lot of data that is big
• Transactions
• SKU Transactions
• Campaign history
• But it has not been Big Data
• Here are some examples….
8
Transactions
• SKU transactions still fit into structures and can be managed
• But it’s big – over 1,000,000,000 items and 100,000 products.
• Challenged to make data manageable
• Models help to key dimensions to make actionable
• E.g., Clusters using Ideal Dimensions (Morder, SUNZ 2012)
• Boil down products and customers to 20 key dimensions.
9
Campaigns
• Campaign history still fit into structures and can be managed
• But it’s big – over 500,000,000 records and 000’s of campaigns.
• Models on campaign history have to boil down data to key dimensions to make actionable.
• E.g., SAS MO uses these models to optimise campaign performance.
• Boil down 70 response models (one per partner).
10
Where the world is heading
• Digital data will continue to grow faster than traditional data
• Digital representation of all our transactions and activity
• Retail environment craves more info to keep up with competition
• ……And our customers expect us to use it instantly
11
Big Data at Loyalty
• Last 12 months we have put a new web-server called Harry.
• Postgres System
• Captures web activity
• Integrates with Service Centre.
• Plans are for Harry expand to incorporate
• Core Fly Buys rewards, transactions, and points processing
• Smart phone app behaviour
• Real time recommendations
• … And this is Big Data.
12
CAP Theorem or Brewer’s Triangle
• Consistency: all clients always have the same view of the data
• Availability: each client can always read and write
• Partition Tolerance: operation continues despite physical network partitions.
13 (P)artition Tolerance
Tradeoffs • A and C means you will not have P
• C and P means you will not have A
• A and P means you will not have C
14
(P)artition Tolerance
A+C
A+P C+P
Pick Any Two
Big Data at Loyalty
• We have systems all around the CAP triangle for different purposes.
15
Riak Mongo
Data warehouse
Pick Any Two
(P)artition Tolerance
Postgres
Pick
Any
Two
Implications for Analytics
• Some data and models are same old style: static, historical looking. • One model or segmentation applied across all customers
• Still very important to have these.
• More and more models using data based on current activity.
Examples
• Real time decisions are rule- based and trigger off of observed activity
• Recommendation engines use a model for every individual • Recommendation based on individual product preferences
16
Models
• Preferences and associations to be updated as data comes in.
• Model structure needs to be able to handle being updated as data is updated.
• Models need to be able to deal with eventual consistency.
• Information may not be 100% complete.
• E.g., Poisson Models coefficients to be updated additively.
• Good for billions of record that need to be processed online, like movie recommendations. 1
7
Recommendation Engines
• Two main types of recommendation systems:
• Content-based
• Collaborative filtering,
• Batch vs. real time algorithms
• Push and pull through all our channels
18
Design of data
• Combine big historical string with new information as it comes in….
• Final Prepare has helped us ‘prepare’ for the future,
• Unfortunately it’s not final.
19
Slow moving models and characteristics
(FINAL PREPARE)
Fast moving new data
Going from Unstructured to Structured data
• Modeling and analysis still requires that data has a structure.
• Find and define your associations. Applies to: • Text mining
• Web logs activity
• Smart phone behaviour
• Consumer behaviour.
Two words: Map and Reduce
• Software functions to go from unstructured to structured
• Tell your IT guys to code for the established mappings.
20
Example - Hypothetical
• Data on a customer and his history of buying patterns – historical data on CA server • Discover that he is online and looking to find out about specials – data feed to AP system. • RFI indicates he is entering the store X – another data feed but to CP server. • CP server triggers for an algorithm to calculate the best offer for that person.
• Proactive message sent by CP server to offer a discount on complementary product. • Capture record of actual spend at checkout – another data feed to CA server • Feedback survey send after leave store – capture on AP server. • Capture record of spend and feedback to perform post campaign analysis and improve models on CA
system. • Report campaign results online to retailer on CA system • Better shopping experience for customer, greater commitment, spreads the word. • Better models, more accurate offers next time. Continual improvement. • Applications are limitless.
21
SAS Products on the Big Data Front
• Real Time Decision Manager
• Social Network Analytics
• Text Miner
• Visual Analytics
22
The Future of the Analytics Industry and of LNZ
• More applications, more digital presence
• Analytics are being led by technology and data
• Implications is that models are changing • Hybrid models: slow models + real time data
• One model per customer
• Analysts will be shifting their focus • Build new types of models and real time solutions
• Unstructured to structured
• Tracking of models will change.
• Lab 360 will be there to provide technology and services for both internal partners and external clients.
23