Optimizations Towards ML-based Travel Recommendation through More Explicit User Preference Data

Abstract

Despite the travel industry’s colossal size and the large amount of available data on travel destinations, there is no prevalent ML-based travel recommendation service deployed in industry. Using a novel dataset collected from over 60,000 users of the travel recommendation website, Stravl, launched by myself and Alexis Sursock in 2022, this paper outlines the process of training and optimizing recommendation models for the relaunch of this site. We evaluate state-of-the-art recommendation techniques in search for methods which may best serve the challenges of this dataset. Building on the two-tower neural architecture, we propose several additions and modifications to the model architectures to improve recommendation quality. Finally, we assess the effects of these model changes on recommendation quality by relaunching Stravl and conducting A/B tests on thousands of users. *Please note that this project is completed in conjunction with Sursock 2023, which focuses on the design, development, and deployment of Stravl’s website; however, all work related to the development of recommendation models belongs completely to this paper.*