This document provides details on the dataset that Expedia Group released to the RecTour community at the 15th ACM Conference on Recommender Systems. This dataset is based on real traveler lodging searches and bookings on Brand Expedia websites, which have been anonymized to protect identities of consumers and suppliers. The intention is to provide the recommendation system research community, and more specifically travel researchers, an open and rich dataset for their work. The motivation for this dataset was multiple requests originating from Expedia Group-sponsored competitions, where participants wanted to use the data that was provided for research purposes. This dataset was designed to meet that specific demand while preserving confidentiality.
The Expedia Group dataset consists of global lodging shopping and purchase data from consumers in multiple countries across tens of thousands of destinations. The data are organized around a set of ”search result impressions”, i.e. the ordered list of properties that a consumer sees after a lodging search at one of the Brand Expedia websites. The user response is provided as a click on a property or/and a purchase of a property room. Only clicks and purchases that occurred after a search and before the next search within a 180 minute time limit are attributed to a search.
A property refers to one of over a million hotels, vacation rentals, apartments, B&Bs, hostels and other properties appearing on Brand Expedia’s websites. Room types are not distinguished and the data can be assumed to apply to the least expensive room type.
The data span a period from 2021-06-01 to 2021-07-31 and contain searches for a random sample of consumers who made at least one click during the above time frame. Consumers who booked more than 4 distinct properties during ∗Both authors contributed equally to this research. this period are excluded. The data span more than 800k unique users and approx. 2.5M searches and include desktop and mobile device trafic. The data include traveler inputs such as adding filters and selecting specific sort types, such as price ascending.
Figure 1 outlines the relationship between the search and property data in the dataset with the values impressed on the Brand Expedia site. Figure 2 outlines the click and purchase pathways on Brand Expedia’s site. 2.1
Several steps have been taken to anonymize the data and obfuscate the true data distribution to protect users and commercial sensitivities.
First, the point_of_sale, geo_location_country and destination_id columns were mapped to frequency based indexes. The prop_id column was indexed based on a random order. Next, distributions of the following categorical attributes were obfuscated by randomly changing proportions of users: • point_of_sale • geo_location_country • destination_id • sort_type • is_mobile
For example, the proportion of mobile searches (given by the is_mobile column) is similar but not identical to the ”true” proportion. Finally, we changed proportions of the num_clicks and is_trans ”label” attributes at the property (prop_id) level. In other words, the click through rate (CTR) and conversion rate (CVR) at the property level computed based on the above attributes do not exactly match the ”true” CTR and CVR values.
In this section we provide a detailed list of attributes.
point_of_sale Integer geo_location_country Integer is_mobile destination_id checkin_date checkout_date adult_count child_count infant_count room_count sort_type applied_filters Boolean
Date Integer Integer Integer Integer String
impressions List[Impr]
ID of the Expedia point of sale (i.e. Expedia.com, Frequency based indexExpedia.co.uk, Expedia.fr, ...) ing. Obfuscated true distribution.
The ID of the country the consumer is located Frequency based indexing. Obfuscated true distribution.
Obfuscated true distribution.
Obfuscated true distribution.
Whether the search was made from a mobile device ID of the destination where the hotel search was performed Stay start date Stay stop date The number of adults specified in the search The number of children specified in the search The number of infants specified in the search Number of rooms specified in the search Sort type Hotel position on Expedia’s search results page.
The ID of the property. It matches prop_id from Table 2.
If the impressed property is a travel ad (labelled "Ad", pay per click advertisement).
bution.
Anonymized Property Name and Point of Interest filters.
Integer
The mean customer review score for the property on a scale out of 5, rounded to nearest integers. A 0 means there have been no reviews, null that the information is not available. The number of reviews for the property rounded to the nearest 25.
The star rating of the hotel, from 1 to 5. A null indicates the property has no stars, the star rating is not known or cannot be publicized.
If a booking can be cancelled without extra fees.
If the property had a discount price reduction
specifically displayed ("strikeout" price).
Price bucket (
2.2.1 Property amenities. In addition to the main dataset from Table 1 we also released a property amenities dataset described in Table 2. This dataset spans approximately 1.5 million properties. Properties from the main table which cannot be matched with properties from the amenities table can be assumed to have missing amenities. 3
Expedia Group has provided a dataset based on real traveler behavior specifically for academic researchers and students. This dataset should address the demand that has been expressed in the past for it during competitions and events. This dataset can also be used by instructors for courses. Feedback is welcome on how we can improve this dataset in the future, and what other datasets may be useful for the RecTour and recommendation system research community. 4
We would like to acknowledge Julia Niedhardt for her initiative with the idea of creating an industry-based real world dataset for recommendation system and tourism researchers. And for her eforts to make it a reality at RecTour 2021.
We also thank Dr. Wolfgang Wörndl for his contribution to this project.
prop_id AirConditioning AirportTransfer Bar FreeAirportTransportation FreeBreakfast FreeParking FreeWiFi Gym HighSpeedInternet HotTub LaundryFacility Parking PetsAllowed PrivatePool SpaServices SwimmingPool WasherDryer WiFi
Long Boolean Boolean Boolean Boolean Boolean Boolean Boolean Boolean Boolean Boolean Boolean Boolean Boolean Boolean Boolean Boolean Boolean Boolean