Every day, millions of travellers from all over the world make multiple decisions on Booking.com related to their upcoming travel plans. With all of these taps and clicks on property photos and scrolling through search results, Booking.com naturally has a wealth of data insights to help the company make changes on the platform to improve the customer experience. In addition to the responsibility of handling all of this information securely and ethically, how do you analyse all of this data properly and continue to make useful recommendations for customers? Is what works well for a Dutch traveller equally as relevant for a traveller from Japan? And how do you ensure that customers continue to receive interesting travel recommendations that are relevant to them without getting stuck in a filter bubble?

On the road to even better recommendations

One way to understand what constitutes a good recommendation is looking at what previous travellers have chosen and the experiences that their choices yielded. Machine learning techniques are well suited to learning such connections and preferences. However, the problem is that the connections and preferences found in the data are not only informed by the choices of other travellers, but also by the suggestions and selections the system showed them. In the Mercury Machine Learning Lab, researchers from the University of Amsterdam (UvA) and Delft University of Technology will work together with data scientists from Booking.com to develop methods that will ensure that this type of bias is avoided and that the learned connections remain accurate in a new or different context.

From the classroom to real-time e-commerce

Joris Mooij, scientific director of the Mercury Machine Learning Lab at the UvA: ‘It's a huge opportunity for us as researchers to have access to a live dataset of global data and be able to experiment on Booking.com's platform.’ Frans Oliehoek and Matthijs Spaan, scientific directors from TU Delft, agree. Oliehoek: ‘By testing AI techniques in the real world, we can better understand the limitations of current state-of-the-art methods in the field of reinforcement learning, as well as improve the application of AI in practice.’ Spaan: ‘The lab's focus on developing better algorithms for recommender systems is highly relevant to our society as these systems guide many of our digital interactions. By addressing fundamental AI challenges, the results of the lab will also be valuable in other domains.’

Onno Zoeter, Principal Data Scientist and Booking.com scientific director adds: ‘Unlike in a hospital scenario for example, where experimenting with different types of data-driven recommendation systems can have real life-and-death consequences for patients and their suggested treatment protocols, testing approaches, models and hypotheses with travel data from Booking.com doesn’t come with the same public health implications. This means we can safely test and develop new machine-learning methods together that also have a potential impact far beyond the trips booked on our platform.’

Learnings from other languages

Artificial intelligence and natural language processing are already used to perform many important tasks in different languages, such as categorising reviews and fraud detection. The researchers will look for ways of creating a system with multiple languages in which the smaller languages can benefit from what has been learned in models and rounds of experimentation with languages that are spoken more widely. This should enable Booking.com to support all 44 languages and dialects in which their platform is available to customers in various new contexts, even more quickly.

Nurturing talent

The Mercury Machine Learning Lab will be part of ICAI, the Innovation Center for Artificial Intelligence. The Lab will provide world-class opportunities for Machine Learning graduates to stay in the Netherlands and lead innovative research, keeping important talent connected to Dutch universities and industry.

In addition to the existing researchers, the Mercury Machine Learning Lab will comprise six PhD candidates and two postdocs who will work on six different projects related to bias and generalisation problems over the course of the next five years. They will spend two days a week in the office at Booking.com doing research and actively participating in related streams of experimentation to test their hypotheses.