Accommodation booking site Hotels.com helped attract more than 600 million monthly users and $8 billion (£6.2 billion) of revenue for its parent company Expedia in 2016. The travel platform is focused on creating repeat customers. Giving users what they want through personalisation is central to that strategy.
Hotels.com does this by compiling a customer's search criteria to present the most competitive offerings for their needs at the top of the list of results when they make a travel query.
It sounds like simple ecommerce practice, but the quantity of data that's involved makes this more complex than meets the eye.
"Typically we will have more than seven terabytes of compressed data that feed into this model," Matthew Fryer, the chief data science officer at Hotels.com, told Computerworld UK.
The algorithms used to determine rankings have 105 billion combinations of factors and include more than ten million different sort models. They are continuously monitored and adapted to changes in the market and customer behaviour to improve conversion rates.
The average user visits the site four to nine times before making a booking, and the search results have to adapt and be updated each time that they return.
"It's got to be right four to nine times in a row, which creates an extra data science problem because I also need to understand what happened before to make sure I'm becoming even more relevant as you give me more signals," says Fryer.
Search visibility is determined by a combination of three factors: "Offer Strength", which benchmarks the price relative to value, "Quality Score", which compares the hotel with others in the market, and "Compensation", a measure of how much hotels pay the company for each booking. The latter only boosts a hotel's rankings over those of an equivalent Offer Strength and Quality Score.
The major factors behind a hotel's ranking include the type of accommodation, location, facilities, price, popularity, guest review score, the details on any special offer, and increasingly subtle customer signals.
The typical performance based-advertising model would charge hotels for every click on the ad, a model known as cost per click (CPC). Hotels.com, however, uses an approach called cost per action (CPA), which charges hotels based on the conversion rates.
Hotels can temporarily enhance their visibility compared to similarly ranked properties by paying for a feature called Accelerator if, for example, they need to quickly fill rooms following a cancellation. This allows them to boost their ranking on searches which are already relevant to the user.
To use Accelerator, hotels bid for a better visibility by paying Expedia an additional commission on top of their normal payment. The higher the bid they offer, the greater ranking position they receive.
For every booking that is then made during the period that they specify, the hotel pays Expedia an extra commission. For example, if a hotel bid five percent, they would pay an extra £5 on every £100 stay. The supplier only pays for the value they receive, so cancellations are excluded from the cost.
How searches adapt to customer needs
A data tool called Travel Graph analyses previous purchases, review and search history, to make personalised recommendations based on historical information about users with similar traits and travel objectives.
The algorithms have to pick up signs of very specific needs. Some customers only want to see a very small subset of hotels, while others will just want to book the place where they previously stayed. Then there are the users with more segmented needs, whether it's a hostel, a five-star hotel, or multiple rooms all on the same floor.
"Every trip is different," says Fryer. "Whether you're going for a holiday in the summer, you're treating your significant other to a honeymoon or an event. Or you're going away with your mates. Different use cases have very personal characteristics."
[Read next: How to get a job as a data scientist]
Numerous cities are known by the same name, which can add further complexity. Most customers who search for a hotel in Geneva will be looking for the Swiss city, but those living in the states around its namesake in New York may prefer something closer to home.
Hotels.com helps each of them navigate to a selection of accommodation that has been personalised to their needs, using methods such as market-based analysis and collaborative filtering. The former makes predictions by comparing the user behaviour to that of previous customers, while the latter forecasts future purchases based on their connections to the products that the customer previously bought.
Comparing an individual user to similar customers can establish patterns that are fed back into the algorithm to further tailor the results.
The success of the algorithms can be judged by conversion rates, time spent looking at a property, which ones users click on and where the customer's choice appears on the list of recommendations. If a hotel is being chosen more frequently than its ranking would suggest, the team can analyse the factors behind that deviation and promote them into the main algorithm.
Data science developments
In the future, Fryer hopes to add a more conversational approach to personalisation by harnessing the power of natural language processing (NLP).
"The nice thing about voice is that you are not limited to actions that you can display in the UI," he says. "Voice is more of an unrestricted search rather than filling out a form. Users can ask for 'family friendly hotels' or 'hotels near X street, close to the beach, etcetera' and our challenge is to accept all of that signal and show the user what they are looking for.
"Voice technology will convert the user's request to a string, but natural language processing will decompose that string into actions that we can take in our application. As we allow users to perform more detailed or 'long tail' searches, we will have to get better about enabling our recommendation systems for these signals."
Fryer believes the company has only scratched the surface on the impact data can have on its business.
"We've moved out of the baby phase of this, which was really about collecting data," he says. "We're now at the toddler phase, where we're starting to use it to help people and starting to really make some big breakthroughs. Where we go from there is a tricky futurologist piece, but it's moving so fast and that makes it incredibly exciting."