@Delivery Hero The Recipe of Delicious Recommendations · @Delivery Hero. Delivery Hero GMV: EUR...
Transcript of @Delivery Hero The Recipe of Delicious Recommendations · @Delivery Hero. Delivery Hero GMV: EUR...
GTEC Berlin, 05.06.2018Gugu Ncube - Director of Search & Discovery, Delivery Hero
The Recipe of Delicious Recommendations @Delivery Hero
Delivery Hero
GMV: EUR 3,8 bn, 14.000+ Employees
30+ Brands (Foodora, Lieferheld, Pizza.de etc)
20 MM Active Customers
40+ Markets
1 MM Orders per Day
Recommendation GoalsDiscover New Restaurants
Recommendation Algorithm
Foodrank - Content Based
Ranking: How GoodRestaurants are at your Favorite Food
Algorithms
Content Based - Foodprint
Collaborative Filtering
Foodprint Features
Key Factors:- Recency of orders- Customer Loyalty- Cuisines- Taste- Ingredients- Budget
Food Taxonomy:- Dish title & description- Menu category- Ingredients
Restaurant level- Reorder rate- Primary cuisine- Secondary cuisines- Budget- Day of week- Time of day- Payment methods
Dish level- Ingredients- Taste (spicy, sweet, umami)- Dietary (lactose free, vegan)- Preparation style (baked)- Menu category
Features Recommendation
Content based Algorithm
ETL
Architecture
OrderSystem
Enrichment& Tagging
Orders
Restaurant Rank
Foodprint
API
Indexer
Aerospike
Recommendation Service
Mobile & Web Frontend
Technology Stack
Hosting- Cloudflare- Google GCP- Kubernetes
Infrastructure- New Relic- SysDig
Code- Scala, Golang- Aerospike- Google PubSub- Cloud SQL
Other- Github- Altassian JIRA & Confluence
Challenges
Data Sparsity -> CF Fail
Quality Metrics (5/5 Ratings) -> Focus on Orders
Missing Data and Mislabeled -> Re-Classification and Cleaning
Offline Popularity of Restaurants -> See McDonalds’
Presentation Bias -> Top listed Restaurants -> Top Ordered Restaurants
Interpretation of KPIs -> Cultural Aspects, Historical Factors