• Dânia Meira

Marketplace recommendation systems. How Delivery Hero shows customers what they want

Dr. Aleksandra Kovachev, Data Science Manager, Delivery Hero

Varun Chitale, Senior Data Scientist, Delivery Hero

#datalift use case in production

Presented live at #datalift No 1 on 25 November 2020

Dr. Aleksandra Kovachev, Data Science Manager:

"Building a model often is specific for a particular country or region, and depends, for example, on population density and typical behavioral patterns."

Dr. Macarena Beigier, Data Scientist:

"Delivery Hero business success is based on AI models specifically designed for particular countries and regions, giving the company a great potential for expansion to new markets."

Delivery Hero is a platform network of 30+ brands operating in 40+ countries, partnering with 600k+ restaurants and a fleet of riders. Delivery Hero is also a pioneer of Q-commerce (quick commerce) for ultra-fast deliveries from grocery stores, pharmacies, pet shops, and more. A particular highlight of the company culture is that people from 100+ nationalities work at the Berlin headquarters.

The Recommendations and Ranking Team at Delivery Hero discovers where, when, and what the customers want. The team couples scalable and efficient data pipelines with fast and reliable APIs to serve recommendations for >6 million transactions daily.

Watch the recording on YouTube: https://www.youtube.com/watch?v=TcyQonAmSGY

Productionizing memory- and model-based recommender systems

Delivery Hero utilizes memory- and model-based recommendation systems in its marketplace. It helps users find relevant items, saves search time, and enhances the experience with new or unexpected offers. Recommender systems improve the customer experience while driving incremental revenue.

The memory-based recommender system is based on a non-parametric mathematical approach computing similarity between users and items using heuristics. We use content-based filtering for discovering similarities among items and collaborative filtering for finding similarities among users. These filters may be combined in a hybrid model.

Our model-based recommender system uses an algorithm to predict the next item a user wants to see and order, again using content-based and collaborative filtering to arrive at a hybrid model.

Users give us explicit feedback on the recommender system through purchases, ratings, and recommendations. We also analyze the implicit feedback deriving from the way users interact with the platform content.

How we evaluate the dynamic recommender system: In addressing the challenge that training data may be biased by previous recommendations made, we favor online over offline evaluations. A/B testing gives us a more objective perspective if we invest the necessary time for the experiment to run before comparing results.

Collaborative filtering for recommending items and vendors

For collaborative filtering, Delivery Hero analyzes transactions based on user similarities and preferences or based on item similarity. A key advantage of collaborative over content-based filtering is that it achieves independence from the context, making the recommendations more practical (e.g., what customers buy together). As we retrain the model over time, we capture trend evolution too.

This type of filtering works particularly well for repeat customers, whereas for new customers, one might start by, e.g., showing the most popular items. We also need to protect the algorithm. For example, so-called shilling attacks might occur where vendors or product owners seek to influence recommendations and ranking (e.g., fake reviews).

Types of recommendations

Delivery Hero distinguishes between vendor- and product-level recommendations. We have different categories, like restaurants, supermarkets, or pharmacies, which define the marketplace. Typical offers then include past orders, the most popular items, similar items, or personalized recommendations.

Data sources and the pipeline

Delivery Hero uses a DataFridge for collecting, unifying, and standardizing all data, whether explicit data (e.g., vendors, products, orders, reviews) or tracking data (e.g., clicks, interactions). From analyzing this data, we extract the meaningful features for the recommender system.

Want to know more about how data is leveraged at Delivery Hero?

Meet one more use case in production at #datalift summit from 22 to 24 June 2022 in Berlin!

Dish catalogue optimization: unstructured data to knowledge graph


About #datalift

Organized by the AI Guild, #datalift is about data analytics and machine learning use cases in production.

From September 2020 to June 2021, we hosted Season 1 entirely online. The #datalift No 1 to No 5 events had 40+ AI Guild members showing best practices from 12 industries to over 5.2k registered attendees.

We are now in the middle of Season 2, which runs from November 2021 to July 2022 and is a hybrid experience: Online + 3-day Summit.

Invitation to #datalift summit

#datalift Summit is in Berlin from 22 to 24 June 2022.

The confirmed speakers and partners are up on the main page www.thedatalift.eu

Get your Early Bird ticket with a discount until 31 January!

8 views0 comments