Airbnb, online booking platform, has over 5 million property listings and tens of thousands of hikes and tours among other travel experiences on offer.
Since sifting through such details can be overwhelming for someone, the San Francisco firm is convinced that artificial intelligence can come in handy.
In a paper published on Arxiv.org dubbed “Applying Deep Learning To Airbnb Search”, Airbnb researchers give a detailed encounter of how in the past two years they have been able to implement a complicated neural network in Airbnb’s mobile and web applications in a bid to boost the importance of search results.
The report follows closely the in-house artificial intelligence (AI) system of Airbnb, which is known for converting design sketches into product source codes. On the company ’s machine learning-driven language system helps in translating listing reviews into native languages for guests.
“The application to search ranking is one of the biggest machine learning success stories at Airbnb. Much of the initial gains were driven by a gradient boosted decision tree model,” they wrote. “The gains, however, plateaued over time. This paper discusses the work done in applying neural networks in an attempt to break out of that plateau.”
According to the explanation given by researchers, most guests begin with a search at the company’s website, particularly for available homes in a given geographic area.
Originally, a scoring function that was “manually crafted” decided which rooms and homes would make their way to the top. In the end, a gradient boosted decision tree (GBDT), which is a model that helps in identifying and ranking predictive factors, replaced the scoring function. The researchers said this switch resulted in “one of the largest step improvements in home bookings in Airbnb’s history.”
Airbnb does not rely on a single artificial intelligence system. In fact, it leverages an “ecosystem” of algorithms that help in forecasting the probability that a host would accept a guest ’s booking request and that a guest would rate an experience or trip highly. User interactions aid in training the algorithms. After a new model is trained, it undergoes testing to evaluate whether it attains a statistically considerable rise in bookings.
The company ’s first artificial intelligence (AI) system set the right conditions for more advanced ones to come. The second system adopted LambdaRank, which is an algorithm that utilizes monitored machine learning technology in solving ranking issues. On the other hand, the final model entails a deep neural network, which considered approximately 195 aspects including historical booking count, amenities and the price.
The task was not seamless. In fact, the model training process involved trial and error. The initial version of the processing pipeline of the team utilized only a portion of graphics card processing power, approximately 25%.
Among the neural networks that the Airbnb team evaluated, one of them utilized the distinctive ID that corresponded with listings as a feature. The concept behind this involved the indexing of the IDs into an embedding, which would aid in encoding the unique properties of each listing. This works similarly to the recommender systems used by both Amazon and Netflix.
“Even the most popular listing can be booked at most 365 times in a year,” they wrote, “[and] typical bookings per listing are much fewer.”
To make matters worse, not every trend was obvious, especially at first. Bookings in testing appeared to correlate with long listings’ views.
However, when a model that forecasted the likelihood of long view and booking times at the same time was employed online, it did not lead to an uptake. The team thinks that long views could stem from various aspects including listings with long descriptions that are hard to parse, extremely unique and “sometimes and humorous” listings, or high-end but costly listings among others.
When it came to the feature engineering side, the team’s investigations resulted in a previous aspect, which had not been taken into account that influenced occupancy.
Despite the setbacks and roadblocks along the way, the team is convinced that all that was worth it eventually. “Feeding on the ubiquitous deep learning success stories, we started at the peak of optimism, thinking deep learning would be a drop in replacement for the GBDT model and give us stupendous gains out of the box,” the researchers wrote.
“A lot of initial discussions centered around keeping everything else invariant and replacing the current model with a neural network to see what gains we could get. Over time we realized that moving to deep learning is not a drop-in model replacement at all; rather it’s about scaling the system. As a result, it required rethinking the entire system surrounding the model.”