Author: Ivaylo Bahtchevanov

Delivering a Delightful Customer Experience

Online food delivery is a massive market worth over $150B annually and growing at an exponential pace. With so much opportunity, competition is fierce – the market is incredibly fragmented, with new entrants coming in every year.

From the customer’s perspective, switching costs are low and the delivery is a commodity. The decision-making of which service to use comes down to whatever service offers a better selection, how quickly can the food be delivered (or at least within a very precise time window), and cost-efficiency. Customer loyalty is earned by providing a convenient, delightful, and highly personalized experience. To do this well, platforms need to have a very nuanced understanding of each customer’s unique preferences and tastes, and to correlate these tastes with users’ location as well as order and lead times.

Delivering personalization and operational efficiency at scale is very challenging for traditional machine learning platforms – the problem formulation of correlating user preferences combined with driving routes and lead times in order to recommend new options is very complex. What’s more, when a new user joins the platform, or when a new restaurant or a new dish joins, there is little to no historical data to make predictions.

Graph learning becomes a very powerful tool for food delivery platforms to (1) fine-tune meal recommendations at scale and (2) optimize delivery routes and times to deliver the best possible experience across all users at scale, regardless of historical data. Graph learning can also infer future behavior and trends that will allow you to make the right decisions to retain and engage your users. In this blog, we’ll dive into each of these important problems.

Fine-Tuning Meal Recommendations

Personalized meal recommendations can boost customer satisfaction and encourage more frequent orders. Graph learning can effectively model relationships between users, restaurants, and specific dishes to draw connections across many variables. For example, nuanced taste preferences like ordering a spicy vegetarian dish that is always paired with a cold, refreshing tropical beverage is something a Graph Neural Network (GNN) can detect, but a traditional recommender system might miss. This enables a platform to make recommendations across new options that both delight and surprise the user.

Some of the most successful food delivery systems leverage graph learning to maximize user satisfaction. For example, Uber Eats’s homepage shows both restaurants and menu items, offering users a wealth of options informed by both historical data and inferred preferences.

Their use of GNNs improved user recommendations by over 12%. Learning relationships between users, meals, and restaurants allowed their ML platform to make connections that are not explicitly present in historical data in surveys, reviews, and purchase histories.

Similarly, Doordash’s machine learning powered recommendation system moved to graph-based methods that helped increase order volume by 25% and understand the features that explain a user’s ordering propensity.

Optimizing Delivery Routes and Times

Meal selection is important, but when that meal is delivered is just as important. A user will quickly move to another service and order an alternative less-optimal option if the delivery window isn’t ideal for their schedule. What’s more, an early dispatch leaves the delivery partner waiting for food, but a late dispatch means the food may not be as fresh when it comes to the customer.

Graphs are particularly useful in modeling a network of streets, delivery points, real-time traffic patterns and then combining it with additional information such as average preparation times for restaurants and specific meals. Presenting options that have optimal delivery times will ensure the user will select that option, while also reducing operational costs with the most efficient routes.

A good example of this is the Uber Eats platform, which leveraged graph learning to optimize the routing and dispatch for millions of drivers and over 300,000 merchants.

By leveraging GPS sensors with motion data and activity recognition, Uber understands exactly what stage of the delivery cycle the driver is in, and can make intelligent routing decisions and recommendations to optimize each step of the process and thus minimize wait times.

Shaving a few minutes off delivery times has a major impact on customer experience.

Predicting Customer Retention

Another major challenge for subscription-based meal delivery business models is customer retention. While delivering optimal delivery times and personalized recommendations is top of mind, it’s very difficult to understand whether these aspects are sufficient for customer loyalty until it is too late.

If a food delivery platform can pick up meaningful signal that will result in either diminished ordering activity or churn altogether, then it can find ways of re-engaging those users and getting them back to active and profitable customers.

Companies like HelloFresh use machine learning to predict churn risk among its customers. Their Customer Lifetime models (CLTV) allowed them to rethink the customer journey and have a tangible impact on retention and revenue.

Graph learning is particularly effective at analyzing broader relational data such as order history, feedback, views/clicks and predicting the likelihood of churn along with the key drivers for that behavior. This allows you to go one step further and determine what the best path forward is for retaining the customers (i.e. targeted offers, discounts, and notifications delivered at the optimal times).

The Kumo Approach

While graph learning is promising, building these capabilities in-house requires sufficient expertise, resources, and expensive talent.

Kumo brings these capabilities directly to your enterprise data and allows you to build GNN models directly on top of your raw data. Kumo will build your enterprise graph from the raw data, learn the relationships and interactions across all entities in your tables. Kumo provides a flexible no-code interface to tackle these problems. Predicting where your customers will order from and when is as simple as writing a SQL query – no feature engineering, no training data generation, and no feature experimentation required!

If you are interested in hearing more about Kumo, reach out to us directly!