Data Science for Dinner

Building Gousto's recommendation (eco)system

Transcript

1. MARC JANSEN / RecSys NYC / April 25, 2018 Data Science for Dinner Building Gousto’s recommendation (eco)system
3. Simple model Leading proposition Fueling family life through good food • You pick your recipes from our menu • We deliver a box of wholesome ingredients in exact proportions with step-by-step recipe cards • No planning, no supermarkets, no waste • Most choice (30 recipes per week) • Most delivery options • From £2.98 per portion
4. Exponential growth is fun, but also very challenging. Especially, when you are pushing boundaries in customer propositions at the same time. An exciting journey so far 2018
5. 1B meals per week in the UK 70% are home-cooked photo credits by https://kulishana.wordpress.com
6. ~40 CTO Head of Data 2 Data Scientists Machine Learning Engineer 5 Engineering Squads + UX/UI Head of Digital Product Head of Engineering Data Scientist jr. Data Engineer GoustoTech & Data Science Q1 ‘17 Q3 ‘17 Q1 ’18 Q2 ’18 sr. Data Engineer jr. Data Scientist H2 ‘18 H2 ‘18
7. Collect Everything Store Everything Expose Everything * * Whilst maintaining data security apps web microservices 3rd party airflow (ETL) amazon Redshift (data warehouse) amazon S3 (data archive) (unified log) data scientists data products business users periscope (analytics and dashboards) Amazon DMS (data migration)
8. Demand Planning Personalised Choice Warehouse Optimisation historic orders data Stock Handling X total orders recipe A 3.4% recipe B 2.8% recipe C 4.1% … onions chicken garlic ... 10,000 8,000 7,000 ... orders routes 10012 10012 10012 10013 10013 ... S001 S004 S020 S005 S013 ... R1 R2 R3 R1 R2 R3 0.8 0.8 0.6 0.6 0.4 0.4 1.0 1.0 1.0 R2 onion R3 beef chicken parsley R1 defaults choosers station picks
10. Problem • Our menus are ‘moving targets’ • Boxes add implicit choice constraints Result • Hard to keep performance consistent from menu to menu • Forked version computationally heavy & complex to tune V0.1 collaborative filtering
11. Ingredients C ooking style Calculating recipe similarity Italian pasta bake French tray bake • Full recipe ontology • Jaccard-3 distance metric • Rank menu against user history
12. Measuring success historic live collaborative + curation content-based mc-precision@k* mc-precision@k ~0.55 ~0.70 ? % bought of top 5 TEST 39% % bought of top 5 CONTROL 25% * menu-conditional precision@k, where k = half of no. recipes on menu
13. DAB (Docker / Airflow / AWS Batch) stack deployment 1 2 3 4
14. Building the recommendation ecosystem Direct Indirect mail web voice subs