Understanding taste: how we offer our customers what they did not yet know they wanted
Always delivering an amazing experience is at the core of everything we do at Delivery Hero. In order to achieve this mission, we strive to ensure a great and seamless online ordering process for all our customers. One key element of a seamless experience is showing customers content that best matches their interests or needs, enhancing their online experience and making sure that they want to come back for more.
Aleksandra Kovachev, Senior Data Scientist in the Recommendations and Ranking team, walks us through how our Product and Tech teams work relentlessly to show our customers what they did not yet know they wanted.
Shortcuts through the decision making process
Each day, Delivery Hero processes up to 4 million orders and partners up with hundreds of thousands of restaurants and a fantastic fleet of riders. Through our local apps, we have the widest and most diverse selection of dishes and restaurants available globally outside of China, and are also branching out in other verticals, delivering household goods, as well as pharmacy supplies or even flowers to our customers’ doors.
What this means for the online ordering experience, is that our customers open an app that contains a plethora of vendors (restaurants, grocery stores, flower shops, etc.), but also an incredibly lengthy list of items. How do we reconcile offering a seamless ordering experience to our customers with offering an overwhelmingly large selection of products?
Through our recommender systems, our teams aim at offering our customers shortcuts through the decision making process, so that they can find the items that best fit their needs and cravings as fast as possible.
What’s a recommender system?
The major goal of recommender systems is to help users discover relevant items (such as movies to watch, news to read, or products to buy) so as to create an enjoyable and seamless experience. Recommender systems present an alternative to search engines: they reduce the individual effort of proactive searches by surprising users with offers that they never searched for, and therefore did not know they wanted.
How do we reconcile offering a seamless ordering experience to our customers with offering an overwhelmingly large selection of products?
Today, recommender systems are among the most powerful machine learning systems that online marketplaces implement in order to drive incremental revenue.
There are two types of recommender systems: either based on mathematical logic, or based on the data they process. Following the mathematical approach, there are:
- Memory-based systems: use simple mathematical formulas to calculate similarities between users or items based on the previous ordering behaviour,
- Model-based systems: complex algorithms that predict the most likely item a customer might want to purchase.
Based on data, the following logics can be applied:
- Content-based filtering: showing you italian cuisine if that is what you have previously ordered,
- Collaborative filtering: recommending items that other customers showing similar purchasing patterns have bought in the past.
All these approaches can be mixed and combined together to bring state-of-the art models that tune the recommendations to different aspects of our customers’ behaviours.
Providing a brick and mortar experience
Providing personalized and relevant recommendations to our customers is the focus of our Recommendations and Ranking team. As Delivery Hero’s services grow and expand, we distinguish between different levels and categories of recommendations.
The main distinction we make is between recommending vendors (particular restaurants, grocery shops, flower shops, etc.) and recommending products (specific dishes, food categories, etc.). Whether we want to recommend a restaurant or a specific dish, we develop and serve different types of recommendations following different logics, such as: past orders, popular items, similar items, complementary items, etc.
Recommender systems present an alternative to search engines by surprising users with offers that they never searched for, and therefore did not know they wanted.
For each strategy, we have built different models that are constantly upgraded and served to our brands. Even the simplest logic, such as ‘popular near you’, is constantly adapted based on user behaviour in different countries.
Moving into our other verticals, such as Dmarts (small warehouses located within cities and optimized for fast delivery) and shops, we are recommending items based on their similarity, e.g. Coca-Cola and Pepsi, but also suggesting items that are usually bought together as bundles, like toothbrush and toothpaste. This gives our customers the impression that their full brick and mortar experience has been moved online.
Delivering food is truly the core of Delivery Hero’s operations. Therefore, we are on a constant journey to try and understand the taste, preferences and behaviour of our customers. And this endeavour comes with its own challenges:
- When you think about it, each and every dish is totally different depending on the restaurant you order it from. A burger bought in a big fast food chain does not necessarily contain the same ingredients as the one you would get from that fancy fusion restaurant around the corner.
- What’s more, they do not bear the same names at all, and we all know how creative these can get.
- To top it all off, taste is a very personal thing. What tastes spicy to me will not necessarily taste spicy to you.
Therefore, we are putting a lot of work into mapping each user’s taste preference through diverse machine learning methods, by striving for a thorough understanding of food types. We achieve this by analysing our data and using this knowledge to allow an automatic categorization and classification of dishes, creating product descriptions based on images, dietary tagging, and estimation of nutritional facts.
Today, recommender systems are among the most powerful machine learning systems that online marketplaces implement.
Let’s take a dish called Pizza Margherita: it would be categorized as Italian and as Pizza. Then, from the picture of the pizza its ingredients would be extracted, e.g. flour, yeast, cheese, tomatoes, etc. The Pizza Margherita would then be assigned with a dietary tag: vegetarian, and based on the size/weight of the pizza, an approximate estimation of nutritional facts would be made.
All of this information will be propagated further to match the dynamic portrait of a given customer and their preferences. Our ultimate goal is to generate so-called `Taste Communities’, where we combine, analyse and serve our customers not only through their unique tastes but also through the tastes of the communities that they belong to. This information will not only help us improve our recommendation models, but also enable company-wide learnings exchange in service of the customer.
Localization, localization, localization
When it comes to personalizing the online ordering experience, localization is also key (food and taste are also very much engrained in cultural habits). Through analysis and experimentation, we have learned that countries have unique preferences and behaviour in food ordering, and therefore we strive to make each recommendation not only adapted to a customer but also to a culture/region as a whole.
This is why we rely on our local teams to know what is best for their customers – whilst supporting them with a strong global tech and product backbone. If you want to know more about a practical application of recommendations locally, stay tuned for an upcoming article about Swimlanes implementation for our brand Talabat (and if you do not know what a swimlane is, don’t worry, we’ll tell you all about it).