We all need to make a good first impression. Even hotels.
That’s why travel giant Expedia is using AI to help hotels put the right photos in front of the right people. Travelers spend less than a second deciding what they think of a place. So those pictures better be Instagram-worthy.
“Your brain is wired to process images first,” said Nuno Castro, director of data science at Expedia. “You’ll know right away if you like an image or not.”
Castro works on the Expedia Affiliate Network, a business-to-business partnership brand of Expedia, Inc. He leads a team that’s using GPU-accelerated deep learning and image recognition to develop a system that automatically chooses the most alluring hotel photos and decide which you’ll see when.
“The order of images is important, especially if you’re on a mobile device and have only a few minutes,” he said. “If all the pictures are of the lobby, you may end up booking another hotel.”
First the Gym, Then the Pool
Expedia has 10 million pictures of the 300,000 hotels in its affiliate network. The company manually selects the first photo for each hotel, but it’s not practical do that for every image for every hotel. Other photos are randomly sorted or grouped by what they picture, which may yield less than ideal results, Castro said.
“Some people want to look at the bathroom to make sure the room is clean, but it may not be the best image to show on top,” he said.
With AI, there’s a better chance of showing shoppers more appealing pictures like a hotel room with a window view. Travelers feel better about hotels that showed a window view, according to a 2014 study conducted for Expedia.
But Castro wants his AI to do more than show beach views before bathrooms. He’s working on deep learning models that will tailor images to customer types. Business travelers might prefer photos of the gym instead of pool pictures, while the opposite is true for vacationing families.
Scoring the Best Photos
Expedia trained and deployed its neural network using our GeForce GTX GPUs locally and NVIDIA Tesla GPU accelerators in the Amazon Cloud. Castro and his team first built a dataset — crowd-sourcing ratings for 100,000 Expedia hotel images. Testers rated each image six times to produce half a million image scores.
Researchers then taught a pre-trained convolutional neural network called VGG16 to classify each image into one of 1,000 image categories. The network also predicts what objects are likely to appear together — a drink, a table and a restaurant, for example. On top of that, the
team trained a deep learning model to match the original crowd-sourced ratings.
The team executed the model in a single day.
Easier Hotel Search
Castro said his team needs to experiment with more data, fine-tune the model and test it before the company can roll it out. In addition to the Expedia-branded sites and its affiliate network, Expedia owns Hotels.com, trivago, Orbitz and Hotwire, among others.
“Navigating hotel photos is very time-consuming,” he said. “We want travelers to be able to find the hotel that’s best for them, in the most efficient way. While this is still just an experiment, it is an example of our commitment to that goal.”
For more information, watch the video below of Nuno Castro’s talk at PyData London 2017.
Feature image credit: Mandarin Oriental Hotel Group