04-23, 14:45–15:30 (Europe/Berlin), B09
Wolt's Discovery page serves as the primary gateway for millions of weekly users exploring diverse cuisines and products. With over 130,000 merchants in 25 countries, presenting relevant content poses a unique challenge. In this presentation, we address the complexities of personalizing the Discovery page using a hierarchical multi-armed bandit (MAB) approach built on the Python ecosystem. We outline the challenges specific to an expansive online delivery platform, introducing our MAB solution that incorporates hierarchical parameters at user, segment, city, and country levels. Leveraging Thompson Sampling for exploration and exploitation, our approach accommodates data sparsity challenges. Evaluation results, both offline and online, showcase the effectiveness of our solution. The talk concludes with insights into the resilient, scalable, and adaptive architecture underpinning our approach, featuring open-source libraries such as mlflow, Flyte, and Seldon Core. Our learnings and future steps toward a personalized, context-aware Discovery page cap off the presentation. Join us as we navigate the intricacies of recommendation challenges in the dynamic world of quick commerce.
Wolt's Discovery page is the main entrance point for millions of weekly users seeking to explore new cuisines, order their favorite dish, or replenish their fridge's stock. The Discovery page is a vertical collection of multiple modules (carousels) which can stem from automatic and curated mechanisms. It features restaurants, retail venues, individual items and dishes along with a broad set of banners.
Wolt consumers have distinct tastes and preferences - all of which can change over time and vary with context. However, they expect Wolt to show what's relevant to them and to be able to discover - coupled with a frictionless experience. We want to satisfy our users, keep them engaged and grow our customer base around the world. Wolt delivery covers over 130.000 merchants in more than 500 cities across 25 countries, which results in a substantial variety and size of content Wolt has to offer its customers. Ranking the most relevant carousels at the top is a key challenge to solve so that our users find what they want fast. This renders personalizing the Discovery page as a key lever. Personalized carousel ranking presents a major recommendation challenge across many different domains like content streaming, ecommerce or quick commerce.
In our talk, we present a hierarchical multi-armed bandit (MAB) solution for personalizing the ranking of carousels on Wolt’s Discovery page which is built on top of the Python ecosystem. Therefore, we first illustrate the specific challenges of an (almost) everything online delivery platform and our goals for Wolt's Discovery page. Second, we present our MAB-approach which combines a novel hierarchical parameterization of bandits on user-, segment-, city- and country-level with classical Thompson Sampling for exploration and exploitation. This approach caters well to the challenge of data sparsity. We also share the offline and online evaluation results of our approach. Lastly, we illustrate the architecture to make this solution resilient, scalable and adaptive. Our architecture is built on top of well-known open source libraries. We’re leveraging mlflow for tracking and lineage, Flyte for ML workflows, Redis for serving features, and Seldon Core for serving user requests online fast and reliably. We will wrap up our talk with our learnings and an outlook for the next steps in our journey towards a personalized, context-aware, and controllable Discovery page.
Intermediate
Expected audience expertise: Python –Intermediate
Abstract as a tweet (X) or toot (Mastodon) –Personalizing Carousel Ranking on Wolt's Discovery Page with a Hierarchical Multi-Armed Bandit Approach
Senior Data Scientist in Wolt's Personalization Team working on Carousel Ranking and Personalized Recommendations. Show Host of Recsperts - Recommender Systems Experts, the Podcast Show with industry and academia experts in Recommender Systems. Building Recommenders and Personalization Solutions with Python for various industries since 7+ years as well as creator and instructor of Python RecSys Training.
After graduating with a business information systems degree from HS Fresenius, Steffen started working as a Data Engineer at Fielmann (Germany's biggest optician), building their internal data platform. In 2021 he joined Capgemini as an IT consultant working in various roles and projects for a multitude of clients - mostly enterprises. In 2023 Steffen joined Wolt as a Machine Learning Engineer in the personalisation team as one of Wolt's first embedded MLEs. Steffen is also a co-organiser of the ML Ops Community Berlin.