How Expedia is using deep learning to improve the hotel booking process

Expedia Partner Solutions is using deep learning models to improve the hotel booking process for partners and - ultimately - the end consumer.

The company is employing a three-stage process with a deployment algorithm sorting properties for its Rapid API to deliver to different partners whether airline, retail store or perhaps corporate travel partner.

Describing it as an “advanced, cutting-edge deep-learning model,” Nuno Castro, chief data scientist for Expedia Partner Solutions (EPS), says:

“We have half a million properties, and we’re signing 15,000 every month, so how do we decide which to give to each partner given that adding them is very complex because we need to match with competitors and other suppliers? 

“It’s based on how well each property is going to work for them, for example, a property in central London or at the airport, which should they add to their system?”

The second stage is about real-time sorting of the properties according to who is searching on a partner's website so that relevant results are presented whether the customer is a corporate traveler or family or other traveler type.

The third stage, which Castro says EPS is launching, is a recommendation API, which kicks in with similar properties that might be a better fit, whilst the traveler is exploring a property they like.

Rapid results

The work seems to be reaping results for both EPS and its partners. The company claims typical conversion improvements of 8% using the ranking models.

One partner test conduced from January to May last year saw a 20% lift in incremental gross bookings.

EPS says another partner saw a 20% increase in conversions from August to December 2018 using deep learning to recommend properties on an email confirmation page.

Castro describes the improvement in the booking process for the end consumer as a win-win for Expedia and partners.

“They get what they want faster. If they can’t get what they like, they are going to move on to other sites, so we need to give them what they want, when they want it and at the right time. 

“Sometimes they don’t know what they want so we need to help them figure it out. It’s about saving them time, and it’s good for the customer and good for us as they will not abandon the site, they will be more loyal to our website.”

Human learning

While Expedia has been following a test-and-learn model for some years, it is not always in a live environment. Castro says:

“To get one idea right we need to try at least eight [tests], and we try to also test offline using deep-learning metrics. If you only use A/B testing you may end up showing something bad. So we do it offline using advanced evaluation metrics which act as a proxy of what the performance would be in terms of conversion for the customer.”

Acknowledging that it's not only the machines that are learning, he adds that data scientists can sometimes select the most advanced model to use without considering simpler ways of doing things.

Castro says that approaches such as a tree-based model can provide good accuracy with interpretable results for stakeholders and customers.

“In so many domains deep learning is a great approach. We all talk about images and voice and even in sort, it can be used in pockets where we can identify properties more accurately but sometimes going for the most complex approach is something we can learn from and maybe use that at a later stage when we have exhausted other models.”

He adds that often data scientists forget to put the human element into results and can get too caught up in the evaluation metrics and the improvements they are seeing.

Forward-thinking

Going forward, Castro says there will be more ways to book hotels such as via chatbot or voice - a development from the dates-and-destination approach today.

At WTM London in 2017, Castro presented a vision of how chatbots might be used to provide a warmer approach to hotel search in the future. 

He says there has now been an explosion in chatbots with major companies like Amazon, Apple and Google all using these systems.

Expedia tackles the development centrally so all elements can “leverage the wealth of data.”

HomeAway, for example, recently launched an AI-powered chatbot to answer basic booking questions for customers.

Shared knowledge

The travel industry is seen as behind other sectors when it comes to AI and machine learning and must look outside its own walls for learnings.

Castro says what’s out there is fairly transferable and that Expedia has learned, and leverages the experience of, Netflix and Google as well as academic research.

“If we only look at travel there is not much going on. Expedia is a tech company but most travel agents don’t define themselves as tech companies. 

“To be good in machine learning and AI, you need to have data but also the infrastructure and talent. It’s a transformational program for many traditional travel companies to have all three but when you have all three, they also need leadership buy-in because there is a risk using this in production. I see many companies struggling.”