Expedia: opening the data treasure-trove with gen AI and LLMs

It’s early spring, and Shiyi Pickrell has been caught up in the UK’s erratic weather. It was a few degrees warmer when she left the Pacific Northwest, but when we met for lunch in East London, it was chilly, grey, and windy.

We both agreed it would be nice to have a weather assistant or digital helper equipped with real-time info that could tell you whether to pack an extra jumper.

As senior vice president of data and AI at the Seattle-headquartered Expedia Group (she reports to CTO Rathi Murthy), Pickrell has become accustomed to mulling over use cases like this, ones that make the traveller’s life a little easier and more personalised.

Travel needs to act fast. It’s the sector with the largest number of variants: an online retailer such as Amazon has an estimated 700 billion different product choices (relating to product, colour size, etc.), whereas a travel platform like Expedia can manage 1.26 quadrillion (that is 15 zeroes!) product variations for hotel bookings alone.

The more data that can be gathered relating to customer preferences, searches, reviews etc. the more relevant recommendations firms like Expedia can offer its users.

At the heart of all this is the platform’s 70 petabytes of traveller information, stored on AWS cloud, which enables it to make hundreds of billions of AI predictions annually.

Pickrell, a former director of data science at Microsoft, joined Expedia in 2021, about a year before generative AI became a household name. However, part of her role has always involved developing machine learning models to power the end-to-end traveller journey.

Once AI exploded, however, the travel brand knew that it had to act fast.

“We were familiar with OpenAI and were playing with an earlier version of ChatGPT but by December 2022, it suddenly became big. Even my mum called me from Beijing, and she’s almost 80 years old, and she started to talk about it. So, we jumped on the tidal wave,” says Pickrell.

“We put it in our app in Q1 of 2023 – a very fast release — and we also worked closely with OpenAI to leverage the tech via a plug-in,” she added.

Travel GPT style

Expedia’s early forays with OpenAI’s ChatGPT were two-fold. First, there’s a planning experience powered by ChatGPT released in the Expedia app, which allows travellers to start a conversation and get recommendations on places to go, where to stay, how to get around, and what to see and do based on the chat which you might otherwise have had with a travel agent.

Any hotels discussed in this open-ended conversation are automatically saved in the app, making it easier for travellers to start choosing dates, checking availability, adding on flights, etc.

Expedia’s second move was to build a plugin for ChatGPT so a traveller can start a conversation directly on the OpenAI app and select the Expedia plugin to bring their trip to life. When they’re ready to book, they can do so in Expedia.

According to Pickrell, the travel platform is also using ChatGPT to summarise trip reviews and report on things like noise levels or breakfast choices, should people ask.

Pickrell adds that her team was careful not to just create “a bunch of one-off solutions”: these genAI activities needed to be scalable.

“We were mindful that this thing may be bigger, so we wanted to make sure the back end was able to scale – so that later we can have genAI powering more features for the traveller as well as the partner as well as in terms of internal employee productivity.”

A year on, what kind of results is this partnership with OpenAI yielding?

“When we launch something, we want to learn from it and measure it. We have iteration cycles. So, we watch how people engage with genAI. I’d say that it’s early on. But we know they eventually add something to the trip board. Based on the learning, we are continually improving on it to make the experience richer,” she says.

GenAI is also being used internally too: for auto document summarisation in call centres and even for productivity as a co-pilot coder in Pickrell’s own department, although she insists the technology will not replace the work of developers.

She reasons, “We use copilot to help our developers, data scientists, and engineers increase their coding productivity. Ultimately, we’re paying them not to code but to solve a problem — so the more they can think about a problem, the better.”

Innovation lab

For Expedia’s more hardcore, travel-loving technophiles – and ones based in the North American markets (as it’s not available anywhere else yet) — there’s also a new ‘Innovation Lab’ section on its site that allows consumers to access beta features and leave feedback.

The first available option sounds like an advanced version of the original GPT chatbot experience, albeit one that gives users more guidance. It will curate trips based on different budget tiers, location, time of year, and interest.

According to Pickrell, the idea is to try to engage people and get their ‘buy-in’ much earlier during the ideation stage — when they are mapping out their dream holiday.

“Most people only have a vague idea of a holiday before they book it. They might only know that they want to go on a family holiday somewhere in August. We can offer them a discovery tool and put together an itinerary. The app really gets to know you,” she explains.

OpenAI’s new text-to-video tool Sora, which the world had a sneak peek of last month via the release of an internal beta, may soon also help holidaymakers visualise their dream destinations. Is this something Pickrell is trialling?

“Sora is quite stunning, really, and we have been looking for something like that. The content looks quite high quality; it’s slick, it helps generate more personalised content. So, it’s a very cool technology — but it’s too early on in its release now, but we’d love to play with it,” Pickrell says.

She emphasised that OpenAI is just one of a range of tech companies that Expedia is in GenAI talks with: “We also talk with AWS and Google. We want to make sure that we leverage all of them because they all have different trade-offs in terms of speed, volume, pricing, SLAs…”

“We’ve built our platform in a way where we can leverage all of them. And switch back and forth to leverage them fully.”

LLMS

The engineer adds that Expedia is also using some Large Language Models (LLMs) in-house, “especially in instances where the firm is dealing with sensitive information that it doesn’t want to pass on commercially.”

Expedia’s ML and AI active models help slice up the data to serve customeres in different ways. The travel platform currently has 350 active models that power the end-to-end traveller journey. 

One machine learning technique involves turning behavioural data (browsing history, questions asked, etc) into ‘embeddings. Pickrell explains:

“This is an ML technique that helps our model to learn about travellers and destinations or connect the tastes of a traveller to a particular destination or a property so that we can serve more relevant recommendations to them.”

One new feature to emerge from this technique Pickrell refers to as ‘themes’, which can break the site down thematically (into beach, sea, convention centres or hotel themes, for instance). “Everyone sees different ones they are all personalised based on browsing and purchase history,” she says.

Read how Accenture is using LLMs to help with product discovery

As the firm experiments with AI, it has created a cross-discipline Responsible AI Council: “We have legal involved, cyber security involved — so with every use case, we evaluate what data is being used,” Pickrell says.

She adds that Expedia Group’s central legal team is based in London, where it is attuned to the more stringent EU AI and GDPR policies.

“AI is it to be the benefit of the traveller so we’re transparent about how this model is working, and to comply with privacy and data laws,” she adds.

Internal transformation

The Expedia Group is a vast organisation boasting more than 50,000 B2B partners, who connect them to more than 3 million hotel and rental properties and more than 500 airlines, cruise companies, and car rental agencies.

Pickrell adds that AI and ML are also being applied to spot internal synergies. Machine learning is being used to recommend new deals to supply partners or to survey optimal pricing for different times of year to help them stay competitive.

Pickrell points out that the company has been able to react to developments with GenAI and LLMs so swiftly in part thanks to a three-year digital transformation journey internally, which has focused on shedding legacy tech and consolidating everything into a singular platform.

In 2023 alone, Expedia Group consolidated more than ten machine learning systems which it uses to train and deploy ML models into one platform. The group also consolidated 6 to 7 experimentation platforms (which design, set up and analyse a/b testing) into another single platform.

“Now that we have this nice tech stack, we are looking for ways to maximise the return on that to benefit, to really crunch out innovations much faster.”

Expedia wants to position itself as the travel industry’s go-to platform for innovation, where other stakeholders can innovate, and smaller businesses and travel agents can digitise their assets and join them online.

“There are lots of smaller players in this space,” adds Pickrell, “and we want to provide something for them too so that we have a really slick space at the back end which is leveraging the platform, and we have the white label template.”