Market Basket & Advanced Analytics at Dunkin Brands

Confidential information: Copying, dissemination or distribution of this information is strictly prohibited.

Market Basket & Advanced Analytics at Dunkin Brands

Mahesh Jagannath, Prasanna Palanisamy

Oct 1, 2014


Agenda

• About Dunkin Brands Inc.

• BI Program at Dunkin Brands

• BI Architecture at Dunkin Brands

• Advanced Analytics Architecture & Methodology

• Advanced Analytics Use Cases at Dunkin

• Market Basket

• Customer Analytics

• Q & A


3

Disclaimer

All data used is sample data for presentation purposes only and is not actual corporate sales or consumer data


About Dunkin Brands


BI Program At Dunkin Brands

5

• First launched at DBI in 2007

• 1350 BI users today with role based access to 504 dashboard pages

• Mature governance process

• Domestic POS sales analysis to increase comparable store sales and profitability of DD and BR in U.S.

• Store development dashboards to identify opportunities to continue DD U.S. contiguous store expansion

• International reported sales analysis to drive accelerated international growth across both brands.


BI/DW Architecture at Dunkin Brands

6

Franchisees(above store)

RPS Archive

Oracle BI

Enterprise Data Warehouse

RPS BluecubePAR

Oracle EBS

DBI Corporate Users

StetonSMG

Other DBI Data

Intl. POS

Hyperion Users

PAR Terminals

Social Media

Loyalty / CRM

Hyperion

Radiant Sales Data

Exadata Exalytics

R


Agenda






• Market Basket


• Q & A


Advanced Analytics platform• Products Considered

• Oracle Advanced Analytics / Oracle R Enterprise (ORE)

• Open Source R

• IBM SPSS

• Chose Oracle Advanced Analytics

• Excellent fit with existing analytics infrastructure

• All the benefits of Open source R

• Scalability of Oracle 11G on engineered systems

Strengths

• Powerful & Extensible

• Graphical & Extensive statistics

• Free—open source

Challenges• Memory constrained

• Single threaded

• Outer loop—slows down process

• Not industrial strength

R environment

R—Widely PopularR is a statistics language similar to Base SAS or SPSS statistics


Oracle Advanced Analytics Oracle R Enterprise Component Architecture


11

Oracle Advanced AnalyticsOracle R Enterprise Compute Engines


12

Advanced Analytics Methodology

Identify Business Objective

Understand Data

Prepare data

Develop modelTest Model

Deploy Model

Monitor Performance &

re-calibrate


ORE Advanced Analytics Framework


Agenda






• Market Basket


• Q & A


15

Market Basket AnalysisIdentify

Business Objective

Understand Data

Prepare data


Deploy Model


re-calibrate

•Understand role of category and purchase behavior

•Identify category marketing opportunities

•Get richer insight into behavioral changes from promotions

•Apply data validation rules

•Transform POS data into MB input format

•Pairwise association model similar to Apriori, custom SQL implementation

•Output to Star schema suitable for OBIEE consumption


Market Basket Business Questions

Choose a Category: (Sub Category Level)

Answer the following questions for that Item in a particular region last week.

• What % of all transactions include [Product]?

• What related items are sold most frequently with [Product]?

• What is the average ticket $ amount when [Product] is present?

• On Average how many [Product] are sold in each transaction?

• What beverages are consumers buying most with [Product]?

• In what % of [Product] transactions is [Product] the only product purchased?

16


Data Analysis & Design Considerations

• 8 M daily transactions, ~25M transaction detail lines

• 20 TB data warehouse size, sales data about 10 TB

• Hierarchies: 5 level Product, 2x4 level Org, 4 level regional ~1000 SKUs @Item Group/Size level

• Exponential growth in combinations with each hierarchy

• 2 years of pre-computed Market Baskets and associated sales measures for reporting

• Nightly compute within ETL window data with 1 day latency

• Measures are non-additive along the Product Hierarchy

17


Design : Approaches considered

1.Use Oracle Data Mining / Oracle R Enterprise Association Rules

2.Use Frequent Itemset table function in Oracle 11g to compute Item-sets

3.Custom SQL Development

• Approach Chosen

• Oracle Advanced Analytics for exploration / Ad-Hoc

• Custom SQL for repeatable basket computation

• OBIEE for reporting

18


19

High-level Design

Transaction Data

Data Model/ Pre-processing

Rule Development Measure

Calculation

UI / Reports

20

4 Key Reports

Transaction Detail: Product of

Interest

Related Product Pairings

Single Item Transactions: % of transactions when

products are purchased alone.

% of Transactions

containing related items

Related Item

21

What beverages are sold most often with PM

Flats?

22

POI Transaction Detail

Transaction Detail: Product of

Interest

23

Related Purchases

Related Product Pairings

Related Transactions

Non-additive measures

5+3+3 Don’t Equal 11 in this case because some medium and small coffees might be sold in the same

transaction!

Single Item Transactions

Click on to drill down for more detail


Agenda






• Market Basket


• Q & A


27

• Customer Profiling

• Clustering / Segmentation

• Customer Churn Prediction

• Targeted Promotions

Current Areas Of Interest


28

Customer ProfilingIdentify

Business Objective

Understand Data

Prepare data


Deploy Model


re-calibrate

• Compute behavioral variables

• Create Customer record

• Data Exploration in R


Customer Profiling: Attributes

Descriptive Spend/ Check Transaction/Frequency

Store Features Historical Purchase

1. Customer ID2. City3. State4. DMA5. Age6. Profession

1. Min Check 2. Max Check3. Total Spend4. Average Weekly

Spend5. Total points earned6. % Points redeemed7. Total No. of coupons

redeemed8. Total discount

amount (Coupons)9. Avg weekly coupon

redeemed

1. Start Date2. Last transaction date3. Days since last

transaction4. Total

transactions/Visits5. Average weekly visits6. % discounted visits7. Top Day part8. Daypart - % Visits9. Preferred Store10. Multi Store flag11. Average DD Card

Recharge Amount12. Average DD Card

Recharge Frequency13. Days since last

recharge14. Current card balance15. Transaction Activity in

weeks

1. POS: drive thru or not2. Combo or not3. Wifi

1. Total Spend /Category

2. % spend on each Category

3. % spend Sub category

4. Average number of items per transaction

5. Preferred item combo

List of customer attributes used as-is or derived from their transactional history


30

Customer Segmentation / ClusteringIdentify

Business Objective

Understand Data

Prepare data


Deploy Model


re-calibrate

• To understand your customers

• Targeted Marketing• Design Promotions




• Identify variables for clustering,

• Normalize data for Clustering

• K-Means Clustering used to cluster Customers and find individual cluster characteristics

• Model displays cluster means – Cluster properties

• Number of Customers in a cluster

• Deployed for targeted Marketing and Monitoring Customer behavior

• Re-run the model periodically to update the new clusters

• Indicates any shift in the customer behavior


Customer Segmentation / Clustering

Clustering Algorithm

Customer Data Profiles

Analyze Cluster means to Derive Cluster Properties

- Regulars – avg weekly visits are 5- 78.2% visits in morning

- Mostly coffee drinker, but 25% times food buyers

- Coffee Regulars - Avg weekly visits are 5.45

- Avg coffee transactions 80.29%

- High Spenders, Frequent visitors- Avg weekly spend ($35.12)- Avg. weekly visits (7.44)

- Coffee and Food in basket (Avg items per transaction 2.4


32

Customer Churn AnalysisIdentify

Business Objective

Understand Data

Prepare data


Deploy Model


re-calibrate

• Define Churn & Active Customer

• Identify Churn Customer patterns

• Is the churn pattern localized or National?




• Create Training data set• Should have equal

distribution of churn and usual customers

• Model to derive churn risk score.

• SVM• Logistic

regression• Naïve Bayes

Classifier

• Test the model on test data set, for which outcome is known

• Select threshold for model selection

• Confusion Matrix for the best Model

• Model will calculate the churn score for existing customers

• Flag customers with high risk, low risk based on churn score

• Monitor the response and re-calibrate by updating training data or model parameters

• Calculate the metrics for model evaluation

Class Active Churn

Active 71.93% 28.07%

Churn 15.37% 84.63%


33

Possible Future initiatives

• Periodic Churn Rate Modeling – measure churn over time

• Customer Segments based on buying pattern – what they buy, when they buy?

• Identify customers who are more likely to respond to offers

• Personalized promotions for retention

• Customer Lifetime value

• Customer Sentiment Analysis

• Enrich customer profiles with modeling scores


Q & A

34

Documents

Market Basket & Advanced Analytics at Dunkin Brands