How machine learning boosts personalisation in travel

Models can predict what sort of customer is using a travel website after just a few clicks

She opens the browser, puts the cursor in the search bar, types "cheap flights from Boston to London," and up pops the first ten links from Google's results page. After some surfing, she lands on Skyscanner filtering the flights by date and cost, and selects the cheapest deal from Norwegian Airlines. As Skyscanner aggregates the offers from providers, she again bulk opens multiple links leading to deals by online travel agencies. She glances over the pages, never staying more than 30 seconds on any of them. At one of the online travel agencies, she opens flight details, hastily closes a pop-up window without reading its contents, and continues searching. In two days, she returns to the same agency closing the top deal from the search feed.

The behaviour of this imaginary user is quite common. Data scientists from AltexSoft, a travel tech provider, call this type of ticket surfer an "economy buyer". The economy buyer accounts for about a half of airfare searches. They look for the most affordable deals, don't spend too much time exploring flight details, and don't care about long layovers or seating.

Back in 2012, Amadeus published a research called Who Travels with You. The study outlined five main segments of travellers: digital natives, young adults, family travellers, empty nesters, and golden oldies. While digital natives and young adults combined are only 22 per cent of the entire travel market, they are the most active web users who prefer booking flights separately from accommodation and leisure activities. But there's a tangible difference even between digital natives and young adults in their behaviour. For instance, digital natives usually belong to the economy buyers group, while young adults aged 25-44 with no kids can afford a more preferential manner of choosing their travel services.

The data science team from AltexSoft has proved that pattern by actively gathering user interaction data working with their OTA clients. To collect all records linked to user behaviour, the team devised a user behaviour analysis engine (UBA). It consolidates data and allows data scientists to build prediction models around it. Today, the machine learning model trained on this data starts predicting the likelihood of a visitor's conversion after just a couple of clicks. Economy buyers belong to the main category assigned to a new user by default. Once they begin checking amenities and looking at other details, the model gains confidence about whether the user is likely to buy and learns their value needs.

The team tracks practically every little interaction aspect: destination searches, chosen travel dates, clicks, and even the ways users examine travel service providers.

The spectrum of this AI application is broad. "We can easily distinguish between business and leisure travellers," says Alexander Konduforov, head of data science at AltexSoft. "Now, it's a matter of value difference that we can suggest to these two groups of visitors." Data analysis can be more sophisticated than that. For instance, bleisure (business+leisure) travel is becoming increasingly popular and providers can suggest sightseeing and restaurant options to try in between meetings and conferences.

Economy buyers comprise half of the audience. The other half is more sophisticated in their travel preferences and you can't just suggest the lowest cost to them. Here's when the real personalisation starts as the system must account for things this person values most. Some don't like long layovers when choosing flights, some are picky about meal options, and some are fans of particular hotel chains. These insights lay grounds for making a customised search engine that will filter travel alternatives considering cost and value priorities for everyone typing in their destinations.

While 79 per cent of business executives surveyed by Forrester believe that personalization will help them achieve marketing and customer experience goals, the practice is still an investment in data science and the underlying technology itself. "The two main challenges we see today are data related," according to Alexander. "As we collect more data, we have to figure out how to efficiently store and further process it."

Another problem is the lack of individual user data. Although the dataset has enough records to build accurate predictions about incoming visitors, the machine still needs users to stay on a website a bit to define their group and tailor the offerings. A long-term user interaction history can provide better personalisation opportunities. Saving cookies allows the algorithm to recognise visitors who have been visited before, and that simplifies things. But people tend to block cookies. Even if a user had purchased travel services before and can be qualified as promising during the second or third visit, one browser cleanup rolls this person back to the unknown state. Now the system is dealing with a clean slate and is at square one in data collection on this individual.

A registered customer who regularly uses the account on all devices allowing the system to provide the best value options based on long-term and consistent data is the best-case scenario. It's possible with loyalty programme members and other MVCs (most valuable customers) and doesn't work with occasional buyers. Just a fraction of users is registered. Some login only on desktops, and most don't have accounts at all. That's why the general advice to embark on personalisation is to start simple and try personalising for MVCs, the people you know best.

However, there are even some ways to partially sidestep the cold-start problem. Although you may not have the behaviour data, you can make assumptions about users solely relying on metadata: Referral websites that people came from, devices, and browsers give some insight. For instance, the users coming from Skyscanner are more likely to buy than those coming directly from Google.

Maryna Ivakhnenko is a manager at AltexSoft