Developing Relevant Dining Visits with Oracle Advanced Analytics Olive Garden s transition toward tailoring guests experiences Matt Fritz Senior Data Scientist
Business Challenge Darden comprises several premiere full-service dining brands Olive Garden s 830 restaurants flourished with a strong brand and standardized process since 1982 One-size-fits-all marketing and operations sometimes scrapped worthwhile programs that didn t fit at all locations and guests Over time, initiatives became constrained with this consistent, national approach 2
Competitive Landscape The casual dining and fast-casual spaces have already begun to localize and tailor occasions Applebee's franchise model affords managing partners the flexibility to locally tailor restaurants Chili s tablet menus enable rapid, dynamic assortment and pricing by location and even table Panera s loyalty program directly links to in-restaurant experiences that adapt each visit to guests needs Starbucks formats and atmosphere reflect local architecture and culture Adaptive decision making capitalizes on untapped profits with: Atmosphere Communications Customer service Menu assortment Pricing Promotions 3
Analytic Opportunities Two major opportunities existed for Olive Garden Restaurant Usage Clustering Are guests using restaurants similarly among regions? Guest Behavior Prediction How will a specific guest behave next? 4
Analytic Solution Selection Balancing flexibility and complexity is key Unsupervised machine learning helped assign each restaurant to one of five distinct behavioral clusters 5
Analytic Tool Selection Analyzing the transaction-level database is challenging Scale Years of records Oracle Advanced Analytics offers scalable, detailed analysis Multiple brands will leverage the same solution Detail Novel, custom variables need to be defined Unknown variable significance and correlation Too many variables for manual cross-tabulation, exploration, and clustering 6
Oracle In-database Advanced Analytics All analytics are performed inside the database Behavioral variable aggregation translates individual transactions into restaurant-level data Principle Components Analysis both: 1. Explores variables relationships and importance 2. Removes unnecessary noise in the dataset K-means Clustering groups similar restaurants according to guests in-restaurant behavior Predictive Classification assigns new or changing restaurants to the appropriate cluster 7
Oracle In-database Advanced Analytics Oracle R Enterprise provides a single interface to utilize code from: Open-source R SQL Oracle s advanced algorithms Oracle R Enterprise s code-based framework versus visual interfaces best fit Darden s needs in allowing for the most flexible and iterative programming structure 8
Oracle R Enterprise Scripting # Directly extract data via SQL ore.exec( "CREATE TABLE OG_INPUT AS SELECT MP.Location, MP.Take_Out, MP.Daily_Specials, MP.Kids_Meals, NM.Sales_Avg, NM.Entree_Price_Index, NM.Lunch_Dinner_Ratio FROM V_MENU_PREF MP JOIN V_NON_MENU NM ON MP.Location = NM.Location") # Exclude specific locations via Transparency Layer PCAin <- OG_INPUT[!OG_INPUT$Location %in% c('1451 Times Square - New York, NY', '1547 I-Drive - Orlando, FL'),] # Execute open-source R s PCA via Embedded R Layer PCAout <- ore.tableapply(pcain, function(x) { library(factominer) PCA(x, ncp=10)$coord }, parallel=true) # Cluster locations via Oracle Advanced Algorithm clusters <- ore.odmkmeans(~., PCAout, num.centers = 5, distance.function = "euclidean") 9
Oracle In-database Advanced Analytics Compared to Darden s conventional methods, this solution: Reaches all available data instead of sampling Comprehensively explores the data s features & relationships Systematically identifies common restaurants across several variables Develops a robust clustering model for scoring restaurants 10
Architecture & Configuration Oracle Database 11g Enterprise Edition Release 11.2.0.3 64bit Red Hat Linux 6.1 Oracle Memory & Processing (QA) 4 x 10 Core CPU 2.9GHz 192GB RAM 5TB Memory (Flash) Business Intelligence MicroStrategy 9.0.1 11
Performance Oracle in-database analytics improves the aggregation of 115 million transactions (1.05 billion rows) by 2000% versus Darden s business intelligence tool Further time is saved by executing the analysis inside the database Task Business Intelligence Oracle Adv. Analytics Data Preparation Analysis Outputs spreadsheets that need manually joined Client laptops limit computing power Pre-formatted, joined, and ready to analyze Advanced algorithms run in the database 12
Clustering Model Assignments Great Lakes West Coast Northeast Notes: Results show generalized learnings from the 100+ variable clustering No geographical variables were used, yet the clustering still shows regional differences Other Clusters: Rural South Suburban South
Operationalization Today, Olive Garden has a comprehensive understanding of different restaurant types through the lens of clustering Demographic and operational performance data is layered into the analysis, providing further explanation of the clusters 14
Operationalization Commonplace reports can now be broken down into five views to uncover unique differences across the five restaurants clusters Menu item preference, market basket affinity, and loyalty reports reveal the most differentiation 15
Operationalization New or changing restaurants can be instantly reassigned to the appropriate clusters according to the data These predictions allow Olive Garden to nimbly read and react to changes in customer behavior from pointed tests or broad initiatives 16
Financial Success Developing the solution internally formed a competitive strength through architecture simplification, performance, and scalability Darden saved several hundred-thousand dollars and two months of development time compared to third-party bids Other Darden brands can leverage the same solution in the future for continued cost avoidance Restaurant clustering will develop into helping: Better inform Olive Garden s nationwide remodel campaign Identify millions in profit opportunities by optimizing pricing, menu assortment, and marketing efforts across the five clusters It will be an ongoing staple as Olive Garden transitions into increasingly more localized marketing and operational strategies 17
Analytic Opportunities Two major opportunities existed for Olive Garden Restaurant Usage Clustering Are guests using restaurants similarly among regions? Guest Behavior Prediction How will a specific guest behave next? 18
Analytic Solution Selection Now, complexity is necessary to add the full value of analytics Supervised machine learning helped predict each guest s next action 19
Parsimonious Data Transformation Transaction data was structured to track guests across visits, but lacked information to condense & measure habitual behavior over time Time Check & Guest Details Guest A Guest B Day Part & Sales Area Habits Preference Across Brands Party Details Across Locations Demographics, Emails, etc. Summarized Habits Guest A Guest B Always dine in at lunch Sometimes bar, else take out Buys higher price entrees High alc bev spend at every brand Family dinner at home, solo dinner on the road Always visits the same locations >$100K income; jdoe@gmail.com 123 Main St. Orlando, FL 20
Habitual Behavior Repeat guests exhibit habitual behavior over time Each segment has a consistent propensity to act unique from other guests E.g., Segment A is eight times more likely to eat at the bar than others First grouping guests by habits eliminates noise that would otherwise obstruct the prediction algorithm 21
Performance Oracle in-database processing enables advanced analysis across millions of guest records Client machines memory would be exhausted at this scale of data Distributed processing accelerates runtime for more real-time analysis 10MM records in 2 5 minutes Predictions can be directly scored to each guest record without an ETL process 22
Prediction Lift Several prediction processes leveraging both Oracle s algorithms and open-source R evaluated each guest s most likely behavior Prediction rates increased several times over compared to initial accuracy 23
Operationalization Potential Understand Guest Patterns Track patterns & changes linked to marketing activities Segment guests based on similar behavioral patterns over time Estimate Guest Value Identify the most valuable guests and what behaviors are most correlated with value Undercover new ways to drive acquisition, frequency, and retention Predict Guest Behavior Forecast what guests are likely to do next, when guests might churn, and what will likely happen in response to a specific action Optimize direct, digital, and email marketing at the guest level 24
Special Thanks Ongoing analytic support ensured these projects successes: Oracle Advanced Analytics Mark Hornick, Director Product Development Charlie Berger, Sr. Director Product Management Vlamis Software Dan Vlamis, President Tim Vlamis, Consultant Questions and comments: Matt Fritz, Senior Data Scientist mfritz@darden.com 25