Deliver Dynamic Customer Journey Orchestration at Scale
1 like681 views
Publicis Epsilon offers advanced customer journey orchestration leveraging extensive data insights, helping brands enhance customer experiences through personalized interactions. With over 200 million unique IDs and a machine learning-driven approach, they optimize marketing strategies and improve customer retention by processing vast transactional data efficiently. Key outcomes include increased revenue for clients, improved customer satisfaction, and critical automation in data handling.
![Model Parameters – 344 million
Layer Type Output Shape Param # Conncted to
all_inputs(Input Layer) [(None, 2)] 0
tf_op_layer_Slice_1 (Tensorflow[(128, 1)]) 0 all_inputs[0][0]
tf_op_layer_Slice_2 (Tensorflow[(128, 1)]) 0 all_inputs[0][0]
cust_behv_embedding
(Embedding)
(128,1,100) 342224350
tf_op_layer_Slice_1
FlatterCusts(Flatten) (128,100) 0 cust_behv_embedding
first_hidden(Dense) (128, 64) 8256 FlatterCusts[0][0]
sec_hidden(Dense) (128, 32) 2080 first_hidden[0][0]
Prediction(Dense) (128, 1) 33 sec_hidden[0][0]](https://image.slidesharecdn.com/412krishkuruppathsharadvarshney-200707154120/85/Deliver-Dynamic-Customer-Journey-Orchestration-at-Scale-23-320.jpg)
Deliver Dynamic Customer Journey Orchestration at Scale
- 2. Deliver Dynamic Customer Journey Orchestration at Scale Krish Kuruppath SVP Technology, Publicis Epsilon @krishkuruppath Sharad Varshney VP Data Science, Publicis Epsilon @varshnes
- 3. Agenda Who we are Journey Orchestration What, why, and how Enablement Model Building and Training Foundation Decisioning Process Model Performance and Results Results Key Takeaways
- 4. Who we are
- 5. 200M+ unique IDs in the U.S. 7,000+ person-level attributes 600M loyalty accounts 56% of all U.S. non-cash transactions LEADER IN DATA DRIVEN MARKETING ID-based insights platform to plan & optimize media Discovery Clean room to analyze customers and acquire new ones Prospect Performance-based media across mobile, desktop & video Digital Media Solutions Customer data hub to power your cross-channel marketing Customer The leading loyalty platform, managing complex programs Loyalty The leading email and digital messaging solution Messaging
- 6. HOW DO WE DRIVE THE CUSTOMER SUCCESS? DATA ACTIVATION MEASUREMENT ID MACHINE LEARNING Highest-performing identity 96% accuracy, 80%+ match, 80% persistence #1 transactional & behavioral data in the industry High-definition view of 200M active consumers Fastest, most efficient AI 1B+ model updates every 5 minutes Driven by performance Independent, unbiased, focused on your outcomes with truth, proof and transparency PRIVACY AT THE FOREFRONT
- 7. Journey Orchestration - What, Why, and How
- 8. The Marketing Industry IS AT A KEY Inflection Point Consumers: Rising expectations Landscape: Increasingly difficult Brands: Marketing goals Drive growth Reduce costs Partner wisely Prove performance Hyper-connected Convenience with quality Want to be recognized, respected and protected Media fragmentation Partners’ conflicting interests Tech stack underdelivery Incomplete customer views Emerging privacy laws
- 9. Source: PwC Future of Customer Experience Survey 2017/18 “One in three consumers (32%) say they will walk away from a brand they love after just one bad experience.”
- 10. Make every Customer interaction personal and purposeful
- 11. To deliver the best customer experience, you need to know who the consumers are, where they are, what they want and when they want it.
- 12. ...and then Deliver a Personalized Experience Business Goals Conquest Brand Affinity Customer Lifetime Value Reduce Churn Operational Efficiency Customer Goals Know Who I Am Respect My Time Make it Easy and Fun Anticipate My Needs Give Me the Most Cost Effective Options Best Customer Experience Compelling Creative Channel Effectiveness EmpathyData driven insights
- 13. Our solution optimizes the journey by detecting the micro-moments and delivering the right call to action
- 14. Customer Journey Awareness Consideration Purchase Service Loyalty
- 15. Customer Journey Acquisition Purchase Decision New customer Loyal Customer Micro-moment: Runs into a sports store to pickup baseball cleats for her son Micro-moment: Browsing online at work looking at new tights Micro-moment: Buying boating gear and sailing clothes before summer Micro-moment: A sales associate asking her about the store card Micro-moment: Considers whether the card would benefit her Micro-moment: Decides to apply online first, then walks into the store applies in person Micro-moment: Uses the card for purchase in the store Micro-moment: Places the card in her purse Micro-moment: Receives a welcome email with some coupons Micro-moment: Back-to-school shopping for kids online Micro-moment: Purchase soccer cleats and ball for her daughter Micro-moment: Buying holiday gifts for the entire family Potential Touchpoints: 1.Geotargeted ad 2.In-store signage Potential Touchpoints: 1.Retargeted digital display ad 2.Triggered email Potential Touchpoints: 1.Retargeted digital display ad 2.Triggered email 3.Checkout notification Potential Touchpoints: 1.Employee communication 2.In-store signage Potential Touchpoints: 1.Brochure 2.Email Potential Touchpoints: 1.Email 2.Application experience Potential Touchpoints: 1.Checkout 2.Physical card or e-card Potential Touchpoints: 1.Brochure 2.Card Potential Touchpoints: 1.Welcome email Potential Touchpoints: 1.Triggered 2.Notification email Potential Touchpoints: 1.Push notification 2.Exclusive cardholder benefit Potential Touchpoints: 1.Triggered email 2.Push notification 1 2 3 4 5 6 7 8 9 10 11 12
- 16. Customer Journey Data flow Normalization Segmentation Entity Resolution Scoring Data Sequencing Data Blocking Text Mining Unique ID RETAIL ECOMMERC E CONNECT Unique IDs Graph ID CustomerLTV Segmentation Graph Segmentation 1 PII Data Customer Engagements (Siloed Experiences) Profile Purchases Clickstream Analytics Commerce Data Mobile Data 1st Party Data (BIA, SAS, Profile ID) Transaction Data (ATG ID, USA ID) Clicks | Impressions | Pixels Emails | Social | Preferences (Email ID, Customer ID) Profile Data | Cookies | IP (D_Profile ID) Device IDs | OS | Location Social Data PartialViewoftheCustomer Same Customer (Data organized by Channel) Sentiment | Social IDs 2 3 4 5
- 17. End-to-end processing on Databricks Clusters Cross Cutting Services CxDB (Delta Lake) Data Hub (Batch & Real-time Fusion Processors) CxDB (Cosmos DB) API logs Inbound, Internal & Outbound Events Meta Data DBs, Logging Incoming & Outgoing API Calls Configurable Ingestion Engines Configurable MML Engines Customer Profile Customer Device Coupon Promotion Reference Data CxDM Customer Defined Purchases CLTV Churn Propensity Recommendation Identity Resolution Knowledge, Sentiment Video, Image, Face Configurable Outbound Engines Storage Adobe Campaign Adobe Target Braze BlueCore Certona API Framework Alert Processor Customer Defined Communications Async Logging, Alerting, Error Handling Analytics & ML Stack Speech Transcription… … … CxDM / Custom DBs Custom Formats ML Models
- 19. Customer Journey Models & Attributes Customer Propensity Customer Affinity Brand Product Product Category Customer Lifetime Value Customer Act-a-like Recommendation Engine Channel Customer Churn
- 20. Model Building and Training
- 21. Databricks Automated MML Pipelines Clickstreamand Customer dataTransaction Dataset ML Pipelines ML Pipeline Customer Demographics Purchases Returns Product Info Click Stream Data Services Data Aggregation Features Generator Pre-processing MML Automation Pipeline Aggregation Pipelines CUSTOMER DATA ENTITIES ER API CLV Propensity Affinity Churn Sentiments Act-a-like Channel Selection Recommender Look-a-like Lifetime Revenue Purchase Behavior Customers At Risk Entity Resolution Deterministic &Probabilistic Stitching
- 22. Aggregate Features Generation Now Predicted Value Time T- Q2T- Q1 T- Q3 T- Q4 Target Features Regular Customer Best Valued Customer Churn Customer
- 23. Model Parameters – 344 million Layer Type Output Shape Param # Conncted to all_inputs(Input Layer) [(None, 2)] 0 tf_op_layer_Slice_1 (Tensorflow[(128, 1)]) 0 all_inputs[0][0] tf_op_layer_Slice_2 (Tensorflow[(128, 1)]) 0 all_inputs[0][0] cust_behv_embedding (Embedding) (128,1,100) 342224350 tf_op_layer_Slice_1 FlatterCusts(Flatten) (128,100) 0 cust_behv_embedding first_hidden(Dense) (128, 64) 8256 FlatterCusts[0][0] sec_hidden(Dense) (128, 32) 2080 first_hidden[0][0] Prediction(Dense) (128, 1) 33 sec_hidden[0][0]
- 24. Model Architecture – Churn - LSTM CustBehavior Embeddings Store Model Arch for Training Product Features Softmax LSTM Layer Dense Layers(2-3) Flatten Layer User Latent Vector Embeddings Layer Product Features Softmax LSTM Layer Dense Layers(2-3) Flatten Layer De-stitched embeddings layer User Latent Vector Embeddings Layer
- 25. Scalability Issues with Inference - De-stitched embeddings layer to reduce model size from gbs to kbs - Cached embeddings in-memory database
- 28. Model Performance and Business Results
- 29. Model Performance ▪ 25 Million Customers ▪ 2.5 Billion Weekly aggregated Transactions ▪ 70+ Data sources ▪ Omnichannel campaign activation ▪ 88% Accuracy with 91% precision ▪ F1 score above 90% ▪ Hit rate >67% Model Performance MatrixData processing volumes
- 30. What did we achieve? ▪ Improved customer retention – able to increase revenue for a a retail client by 2.3 Million/per year ▪ Optimized marketing campaign dollars ▪ Cost optimization through on- demand autoscaled clusters ▪ 2.5 Billion Transactions processed in < 25% of time ▪ Full-scale automation ▪ Faster time to market Operational ExcellenceBusiness Results ▪ Personalized Recommended Products ▪ Better promotional offers and deals based on the life time value ▪ Higher customer satisfaction Customer Benefits
- 31. 3 Key Takeaways ▪ Large volume of customer behavioral and transactional data gives better accuracy and precision ▪ Ability to handle real-time and batch is critical Processing SpeedComprehensive Data ▪ Data Pipeline automation is essential for the success Automation
- 32. Feedback Your feedback is important to us. Don’t forget to rate and review the sessions.