Improving Airbnb Search Ranking Diversity with Neural Networks

Airbnb connects millions of guests and hosts daily, and its search results are driven by a neural‑network‑based ranking algorithm. The original model was designed to select a single listing for a guest; the latest improvement extends the network to generate results containing multiple listings, thereby enhancing diversity.

The ranking neural network predicts the probability that a listing will be booked by comparing listings pairwise and assigning weights to various attributes such as price, location, and reviews. By training on search logs and rewarding listings that were actually booked, the model learns preferences—for example, lower‑priced listings receive higher scores.

Beyond price, the model also captures signals like distance to the search location, number of reviews, bedroom count, and photo quality. Balancing these factors across cities and seasons is a major challenge.

The article examines whether a universal “majority principle” (favoring the preferences of most guests) can be applied everywhere. Using data from Rome, it illustrates a Pareto‑type distribution where 20 % of orders generate 50 % of booking value, highlighting heterogeneity in guest preferences.

To increase diversity, Airbnb introduces a secondary similarity network that estimates how similar a candidate listing is to those already shown in the result list. Training data are constructed from search logs by pairing the first displayed listing (treated as a “reference”) with subsequent unbooked listings.

During inference, the similarity score is subtracted from the booking probability predicted by the primary ranking network. The system iteratively builds the result list, selecting at each position the listing with the highest adjusted probability while keeping similarity to previously shown listings low.

Deploying this approach led to a 0.29 % rise in non‑canceled booking rate and a 0.8 % increase in booking value, with a 0.4 % uplift in five‑star ratings, indicating higher overall guest satisfaction.

The article concludes that reducing similarity among displayed listings, backed by solid machine‑learning techniques, improves search quality and sets the stage for further research on result diversity.