Balancing Machine Learning And Human Intuition In The Travel Industry

Travel planning is incredibly stressful. Between researching options, paying for bookings, and organizing your itinerary, you may also have to contend with the risk of being beaten and dragged off planes.

Machine intelligence can alleviate some of the pain points for both you and the travel companies you book with. Perhaps no one knows this better than Giorgos Zacharia, CTO of Kayak and holder of a Ph.D. in artificial intelligence and machine learning from MIT. “AI is kind of a fashionable domain at present,” he says, amused by the recent hype, “but we’ve been doing machine learning and AI at Kayak for a long time.”

Almost every aspect of your digital user experience is improved with AI. Your preferences towards specific seasons, hotel styles, and price parameters are carefully monitored so that you can be served results you’re likely to book. The photos you see on hotel listings are run through thousands of split tests in which users rank different versions and the results are optimized for mass appeal. Turns out we prefer our hotel photos to be clean and pristine and dislike when they feature other people.

Have you ever gone through an entire hotel or flight booking process, only to be told at the end the item was unavailable? Like many industries, travel companies suffer from inconsistency in data. Due to a slew of legacy systems, changes in hotel and airline databases might not fully propagate in time to booking providers to reflect real-time supply. To combat this problem, Kayak’s algorithms analyze a wide variety of historical sources to generate a more accurate forecast of availability.

Another common data challenge is handling duplicates. “With all those records coming from different systems, you can have misspellings, different word orders, and other issues that could cause a system to create duplicate records,” explains Zacharia. For example, a single listing could be titled differently as the “Boston Marriott Hotel” or the “Marriott Hotel In Boston.” To save time, the de-duping process is largely automated by machine intelligence. Only low-confidence records, where the algorithm isn’t sure of a prediction, are escalated to human staff for analysis. Records from different sources may even disagree about basic facts, such as whether a hotel has a pool, but Zacharia assures that “these algorithms can rationalize that data very, very quickly.”

Machine analysis can yield surprising learnings that contradict your intuition. In a previous role before Kayak, Zacharia built systems to predict corporate bankruptcy filings. One month before a bankruptcy, a company’s credit score often sees a dramatic improvement. The revelation led to further investigation. Turns out CFOs of at-risk companies desperately start paying back overdue bills in the hopes of getting another loan, but typically fail.

Similar findings occur in the travel space. For example, users care less about the average review score of a hotel and more about the number of reviews. A hotel with fewer than 24 reviews is far less likely to be booked even if the comments are overwhelmingly positive. Users also have a sophisticated nose for spotting good deals. Broadcasting a clear discount typically results in higher conversions, but even when a hotel discounts rooms without revealing original prices, buyers intuitively flock to the deal.

Airbnb, which matches hosts and travelers needing a place to stay, analyzes over a hundred signals to personalize guest search experiences. Beyond the obvious parameters of location, length of stay, and what you click on, algorithms track subtle cues about your preferences. “What kind of setting is it in? What kind of decor is in the house? These are things Airbnb can use to feed into the model to come up with a better prediction of which listings to show you first,” explains Mike Curtis, VP of Engineering.

Travelers are not the only ones to benefit from Airbnb’s investments in machine learning, which started in 2014. Hosts often struggle to determine the right pricing for their properties and typically resort to searching for comparable hotel and vacation rental listings in their area. Curtis and his team built a price prediction engine that calculates the probability of a listing being booked on any given night up to a year into the future and generates suggested prices based on expected occupancy rates.

These two data products are major contributors to Airbnb’s business growth, but also drive improved user experiences for both sides of the market. Curtis emphasizes that “the real product is out in the world, human interaction in the physical world, and everything that we do in technology and online is to facilitate more great offline experiences and connections between people.”

Kayak and Airbnb are not the only travel companies  leveraging machine learning. Booking.com, helmed by CEO Gillian Tans, prides itself on international reach, listing properties in over 225 countries in 43 languages.

Many don’t realize that “Booking.com is one of the biggest translation companies in the world,” according to Tans. Headed to a foreign country where you don’t speak the language? No problem. Booking’s cross-platform chatbots allow guests to connect with overseas hotels and hosts and perform real-time translations for all of their supported languages.

In addition to translating human-to-human conversations in real-time, Booking.com bots act as 24-hour customer service agents able to answer most simple travel questions. Kayak also has bots on Facebook Messenger, Amazon Echo, and Slack, in order of popularity. While the natural language processing (NLP) behind the bots are the same, Zacharia notes that users approach different platforms with different intents. “On Alexa, we get more aspirational queries, such as how much a flight is to Hawaii. On Facebook, we get more lower-funnel queries after a user has already booked, such as where they should eat,” he reveals. Complex questions still need to be handled by humans, but Tans remarks “it is surprising how much you can do with machine learning and how good it is getting.”

While the improvements brought by machine learning are impressive, the travel industry must overcome many barriers to reach Tans’ ultimate vision of AI being a “fully-functional digital travel assistant that can proactively solve potential issues before you even know they exist.”

“Booking travel is not like shopping, or groceries or booking a restaurant. It’s much less frequent, so understanding what works just takes a lot more time.” Tans also emphasizes that we must aim for the correct balance between human interaction and sufficient automation.

“Travel is a combination of the personal and the emotional,” she says. “Every customer is different and the travel experience is completely fluid, but the end goal is to find the best solutions.”