Real-Time Decisions Using ML on the Google Cloud Platform
•
1 like•249 views
Ocado Technology is providing a full solution to put the world’s retailers online using the cloud, robotics, AI and IoT. Processing tens of thousands of orders every day, we generate millions of events every minute, leading to huge amount of data to be managed. We will present how this Big Data is handled in Google Cloud Platform to build a end-to-end machine learning pipeline: how data is stored and processed in BigQuery, post-processed and copied with Dataflow, then used to train Deep Neural Network models with TensorFlow, how all this is orchestrated using our in-house scheduling software called Query Manager, and how predictions are finally run in real-time using Cloud ML Engine and Datastore.
Recommended
Recommended
More Related Content
What's hot
What's hot (19)
Similar to Real-Time Decisions Using ML on the Google Cloud Platform
Similar to Real-Time Decisions Using ML on the Google Cloud Platform (20)
Recently uploaded
Recently uploaded (20)
Real-Time Decisions Using ML on the Google Cloud Platform
- 1. Real-Time Decisions Using ML on the Google Cloud Platform Przemysław Pastuszka & Carlos Garcia QCon London 7th March 2018
- 2. How many of you are interested in machine learning?
- 3. but… how many of you are running real-time machine learning in production?
- 4. Who is Ocado? Ocado is the world’s largest dedicated online grocery retailer We have 645,000 active shoppers And 49,000 SKUs in our webshop Three highly-automated fulfilment centres 263,000 orders a week ‘picked’ 3 million routing calculations per second
- 6. What Ocado Technology does (1) Cloud and AI (2) Automation and robotics (3) Big Data (4) Web and app development (5) IoT
- 7. Fraud: An ML journey
- 10. But then… what is fraud? • Mainly chargebacks • Other types of fraud? • Learn from the actual outcome
- 11. Do I really need to do any ML?
- 12. Know your target • Do you need ML? • What do you want to predict? • How good are you at predicting that?
- 13. Cost of mistakes • False positives and false negatives • How expensive are they?
- 14. Start with heuristics • Ask domain experts • Derive rules from expert knowledge • “If more than 80% of order is alcohol, then classify as risky”
- 15. Heuristics are not enough
- 16. Motivations for Machine Learning Data-driven
- 17. Motivations for Machine Learning Data-driven Fraudsters learn
- 18. Motivations for Machine Learning Data-driven Fraudsters learn Customer patterns
- 19. Motivations for Machine Learning Data-driven Fraudsters learn Business changesCustomer patterns
- 20. Challenges • Fraud (human) agents • ML is affected by human decisions • Unbalanced classes (fraud / not-fraud)
- 21. What ML model do you choose?
- 22. “With great power there must also come… great responsibility” Spider-Man
- 25. Challenge your explanations “Why should I trust you?” 2016, M. Tulio, S. Singh, C. Guestrin
- 27. Criteria Online vs batch predictions Explainable predictions Cloud vs on-premise Programming language
- 28. Our ML choice
- 29. Criteria Online vs batch predictions Explainable predictions Cloud vs on-premise Programming language online preferable cloud Python / Java
- 33. Interesting alternatives Amazon SageMaker Google Cloud Machine Learning Engine
- 34. Problem #1 Not fast enough
- 35. Data exploration cycle State the hypothesis Act Validate 03 01 02
- 36. Big Query Validate your hypothesis - fast!
- 39. Problem #2 Data delivered too late
- 40. µService µService µService µService Amazon Web Services Google Cloud Platform Big QueryKinesis Data Flow Apache Beam +
- 41. List<String> strings = ... strings.stream().collect( Collectors.groupingBy( word -> word.charAt(0), Collectors.counting())); PCollection<String> pipeline = ... pipeline .apply(MapElements.via(row -> KV.of(word.charAt(0), word))) .apply(GroupByKey.create()) .apply(Count.perKey())
- 44. Missing data
- 45. Capture every change to the business state
- 46. Training
- 48. train(C1 , … CN , O1 , … ON , Y) = model C1 , … CN , O1 , … ON - customer and order features C1 - Average basket size for the customer O1 - % of alcoholic items in current order ... Y - Fraud or not fraud
- 52. train(C1 , … CN , O1 , … ON , Y) = model model(C1 , … CN , O1 , … ON ) = prediction C1 , … CN , O1 , … ON - customer and order features C1 - Average basket size for the customer O1 - % of alcoholic items in current order ... Y - Fraud or not fraud prediction - Probability of current order being fraudulent
- 53. Model µService O1 , … ON Machine Learning Engine
- 54. Model FeaturesEvents SQL Apache Airflow µService Datastore Custom App ID: C1 … ID, O1 , … C1 …, O1 , … Machine Learning Engine
- 55. Model FeaturesEvents SQL Apache Airflow µService Datastore Custom App ID: C1 … ID, O1 , … C1 …, O1 , … Machine Learning Engine Training Serving C1 …, O1 , …
- 56. Architecting for the future
- 58. Know your target Keep It Simple Choose your model wisely Google Cloud ML Engine for Neural Nets Have data and tools ready BigQuery is king Unified architecture for training and serving predictions
- 59. Thank you!