Scalable User Interface Optimization Using Combinatorial Bandits

The mission of major e-commerce platforms is to enable their customers to find the best products for their needs. In the common case of large inventories, complex User Interfaces (UIs) are required to allow a seamless navigation. However, as UIs often contain many widgets of different relevance, the task of constructing an optimal layout arises in order to improve the customer's experience. This is a challenging task, especially in the typical industrial setup where multiple independent teams conflict by adding and modifying UI widgets. It becomes even more challenging due to the customer preferences evolving over time, bringing the need for adaptive solutions. In a previous work [6], we addressed this task by introducing a UI governance framework powered by Machine Learning (ML) algorithms that automatically and continuously search for the optimal layout. Nevertheless, we highlighted that naive algorithmic choices exhibit several issues when implemented in the industry, such as widget dependency, combinatorial solution space and cold start problem. In this work, we demonstrate how we deal with these issues using Combinatorial Bandits, an extension of Multi-Armed Bandits (MAB) where the agent selects not only one but multiple arms at the same time. We develop two novel approaches to model combinatorial bandits, inspired by the Natural Language Processing (NLP) and the Evolutionary Algorithms (EA) fields and present their ability to enable scalable UI optimization.