Offer Data AISummit

1 .Offer Recommendation System with Apache Spark at Burger King Luyang Wang Director of Machine Learning Engineering & Data Science, Restaurant Brands International Kai Huang Software Engineer at Intel Corporation 

2 .Agenda ▪ Use Case Overview ▪ DeepFlame Offer Recommendation System ▪ End-to-end offer recommendation system with Apache Spark 

3 .Offer Recommendation Use Case ▪ Offer is the main sales driver in QSR industry ▪ Different customers have different offer needs ▪ Different locations serving different pricing tier offers ▪ Some offers can be time sensitive. e.g., can only be redeemed during breakfast hours 

4 .Offer Recommendation Use Case • 1:1 Recommendation • Customer Segmentation ▪ Common Approaches: ▪ Common Approaches: ▪ Collaborative Filtering ▪ RFM ▪ Wide and Deep / Neural Collaborative Filtering ▪ K-Means / K-Mode / DBSCAN ▪ Challenges: ▪ Challenges: ▪ Low interpretability ▪ Segmentations won’t tell marketers directly ▪ Not flexible to maximize on multiple goals what offers to assign to each customer ▪ User embedding can be huge to manage segmentation 

5 .New Offer Recommendation System Goals ▪ An interpretable recommendation system ▪ A system that can consistently track customer’s movement across different segments ▪ Marketers can leverage this system to maximize on different goals to different customers ▪ Fast deployment, easy to maintain 

6 .Burger King’s Offer Recommendation System: DeepFlame • BERT • ResNET50 • TxT* • K-Means ▪ K-Means Clustering based on customer’s behavior data such as average spend, primary service channel, average ticket GPM, and visit frequency, etc. *https://arxiv.org/abs/2010.06197 https://github.com/intel-analytics/analytics-zoo/blob/master/pyzoo/zoo/models/recommendation/txt.py 

7 .DeepFlame Overview – Model Training A hybrid approach that allows SME to easily maintain and modify offer rules based on segmentations while still allowing DL models to automatically pick the best offers according to preset offer rules. TxT 

8 .DeepFlame Overview – Model Inference A hybrid approach that allows SME to easily maintain and modify offer rules based on segmentations while still allowing DL models to automatically pick the best offers according to preset offer rules. 

9 .Key Advantages of DeepFlame Goals DeepFlame Solutions ▪ An interpretable recommendation ▪ The embedded customer segmentation system can be used to explain the logic behind offer assignments ▪ A system can consistently tracks ▪ Behavior segmentation only uses behavior customer’s movement across different level feature which won’t be affected by segments offer or user pool changes ▪ Marketers can leverage this system to ▪ Offer rules are generated based on segmentations and the DL recommender maximize on different goals for picks the best offers under the rules different customer segments ▪ Unified pipeline built on a single Xeon ▪ Fast deployment, easy to maintain cluster using Apache Spark and Analytics Zoo 

10 . End-to-End Recommendation System on Big Data Prototype on laptop Experiment on clusters Production deployment w/ using sample data with history data distributed data pipeline Big Data Pipeline It would be beneficial to have a unified recommendation solution that: ▪ Supports easily prototype end-to-end recommendation pipelines which can apply AI models to big data ▪ Needs “Zero” code change from laptop to a distributed environment ▪ Can be seamlessly deployed on production Hadoop/K8s clusters 

11 .Analytics Zoo: Software Platform for Big Data AI Verticals (Feature Engineering, Models, Use Cases) ML Workflow (Automate tasks for building end-to-end pipelines) End-to-End Pipelines (Seamlessly scale AI models to distributed Big Data) Laptop K8s Cluster Hadoop Cluster Cloud Compute Environment DL Frameworks Distributed Analytics Python Libraries (TF/PyTorch/BigDL/OpenVINO/…) (Spark/Flink/Ray/…) (Numpy/Pandas/sklearn/…) Powered by oneAPI https://github.com/intel-analytics/analytics-zoo 

12 .Offer Recommendation System In Production ▪ Only a single cluster is needed for storing data, Offline Training (Yarn Cluster) performing data analytics and distributed training. Maintain training code ▪ POJO-style API for real-time inference with low ResNET50 latency. BERT SQL, MLlib TxT Output model files Online Serving (Kubernetes Cluster) Model Registry Request Load/update model files Offer Recommendation Client Collect new clickstream data InferenceModel https://github.com/intel-analytics/analytics-zoo 

13 .Feedback Your feedback is important to us. Don’t forget to rate and review the sessions.