=Paper=
{{Paper
|id=Vol-2974/invited1
|storemode=property
|title=Presentation of the expedia group RecTour Research Dataset
|pdfUrl=https://ceur-ws.org/Vol-2974/invited1.pdf
|volume=Vol-2974
|authors=Adam Woznica,Jan Krasnodebski
|dblpUrl=https://dblp.org/rec/conf/rectour/WoznicaK21
}}
==Presentation of the expedia group RecTour Research Dataset==
https://ceur-ws.org/Vol-2974/invited1.pdf
Expedia Group RecTour Research Dataset
ADAM WOZNICA∗ and JAN KRASNODEBSKI∗ , Expedia Group, Switzerland
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.
Additional Key Words and Phrases: datasets
1 INTRODUCTION
Expedia Group is the world’s travel platform that offers consumers a broad selection of travel products across brands
such as Expedia, Hotels.com and Vrbo. 2019 bookings were over $107 billion while serving hundreds of millions of
travelers [4].
To foster research in recommendation systems for travel, Expedia Group has provided a real world dataset that
consists of lodging shopping and purchase data. This builds upon Expedia Group’s previous efforts in the area of sharing
data for recommendation system and tourism researchers via competitions [6, 7] and educational challenges [1, 3].
Participants were often interested in using the data from the contest for additional research of their own. However,
datasets from contests are not directly fit for general research as they are designed for the smooth operation of a
specific competition. This places various requirements on them not related to research uses such as doctorate theses or
academic research. The authors consulted with leading researchers from the RecTour community [2] to create a dataset
inspired by these competitions that was oriented towards research use. There was also a perceived desire within the
wider RecSys community for datasets similar in concept to MovieLens [5] in other fields, in order to provide diversity
and additional avenues for recommendation research.
The dataset is available under a Creative Commons license, subject to appropriate acknowledgement.
2 DATASET
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.
Authors’ address: Adam Woznica, awoznica@expediagroup.com; Jan Krasnodebski, jkrasnodebski@expediagroup.com, Expedia Group, Rue du 31
Décembre 40-42, Geneva, Switzerland, 1207.
Copyright © 2021 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
A. Woznica and J. Krasnodebski
Fig. 1. Data labels as seen on Brand Expedia sites.
this period are excluded. The data span more than 800k unique users and approx. 2.5M searches and include desktop
and mobile device traffic. 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 Data Anonymization and Resampling
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
Copyright © 2021 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
Expedia Group RecTour Research Dataset
Fig. 2. Representation of click and purchase on Brand Expedia sites.
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.
2.2 Attributes
In this section we provide a detailed list of attributes.
Table 1. Attribute description.
Attribute Name DataType Description Comments
user_id String Unique user id (i.e. browser cookie)
search_id String Unique search id
search_timestamp Timestamp Date and time of the search Rounded to minutes
Copyright © 2021 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
A. Woznica and J. Krasnodebski
point_of_sale Integer ID of the Expedia point of sale (i.e. Expedia.com, Frequency based index-
Expedia.co.uk, Expedia.fr, ...) ing. Obfuscated true dis-
tribution.
geo_location_country Integer The ID of the country the consumer is located Frequency based index-
ing. Obfuscated true dis-
tribution.
is_mobile Boolean Whether the search was made from a mobile Obfuscated true distri-
device bution.
destination_id Integer ID of the destination where the hotel search was Obfuscated true distri-
performed bution.
checkin_date Date Stay start date
checkout_date Date Stay stop date
adult_count Integer The number of adults specified in the search
child_count Integer The number of children specified in the search
infant_count Integer The number of infants specified in the search
room_count Integer Number of rooms specified in the search
sort_type String Sort type Obfuscated true distri-
bution.
applied_filters String Pipe delimited list of applied filters. Each filter is Anonymized Property
identified by its name and value. Sample value: Name and Point of Inter-
STAR:4.0|LODGING:HOTEL est filters.
”|” delimited list of impressions. Each impression
consist of the following ”,” delimited attributes:
• rank
• prop_id
• is_travel_ad
• review_rating
impressions List[Impr] • review_count
• star_rating
• is_free_cancellation
• is_drr
• price_bucket
• num_clicks
• is_trans
Impr.rank Integer Hotel position on Expedia’s search results page.
Impr.prop_id Long The ID of the property. It matches prop_id from Indexed based on a ran-
Table 2. dom order.
Impr.is_travel_ad Boolean If the impressed property is a travel ad (labelled
"Ad", pay per click advertisement).
Copyright © 2021 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
Expedia Group RecTour Research Dataset
Impr.review_rating Float The mean customer review score for the prop-
erty on a scale out of 5, rounded to nearest in-
tegers. A 0 means there have been no reviews,
null that the information is not available.
Impr.review_count Integer The number of reviews for the property rounded
to the nearest 25.
Impr.star_rating Float The star rating of the hotel, from 1 to 5. A null in-
dicates the property has no stars, the star rating
is not known or cannot be publicized.
Impr.is_free_cancellation Boolean If a booking can be cancelled without extra fees.
Impr.is_drr Boolean If the property had a discount price reduction
specifically displayed ("strikeout" price).
Impr.price_bucket Integer Price bucket (1-5) based on percentile of the
distribution of impressed prices; lower values of
price_bucket correspond to lower prices. A null
value means that the property was not available.
Impr.num_clicks Integer Number of clicks within 180 minutes Obfuscated true distri-
bution.
Impr.is_trans Boolean If there was a transaction within 180 minutes Obfuscated true distri-
bution.
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 CONCLUSIONS
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 ACKNOWLEDGEMENTS
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 efforts to make it a reality at RecTour 2021.
We also thank Dr. Wolfgang Wörndl for his contribution to this project.
Copyright © 2021 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
A. Woznica and J. Krasnodebski
Table 2. Property amenities table.
Attribute Name DataType Comments
prop_id Long It matches Impr.prop_id from Table 1.
AirConditioning Boolean
AirportTransfer Boolean
Bar Boolean
FreeAirportTransportation Boolean
FreeBreakfast Boolean
FreeParking Boolean
FreeWiFi Boolean
Gym Boolean
HighSpeedInternet Boolean
HotTub Boolean
LaundryFacility Boolean
Parking Boolean
PetsAllowed Boolean
PrivatePool Boolean
SpaServices Boolean
SwimmingPool Boolean
WasherDryer Boolean
WiFi Boolean
REFERENCES
[1] 2021. EXPEDIA GROUP X ENTER21 Data Science Competition Socially Responsible and Inclusive Tourism. https://enter-conference.org/compete/
expedia-group-x-enter21/.
[2] 2021. RecTour: Workshop on Recommenders in Tourism. https://recsys.acm.org/recsys21/rectour/.
[3] American Statistical Association. 2017. ASA DataFest 2017. https://www.dropbox.com/s/eafdup47fpcqvam/UofT%20Stats%20data%20than%20v5%20-
%20FINAL.mp4?dl=0.
[4] Expedia Group. 2020. Form 10-K. https://s27.q4cdn.com/708721433/files/doc_financials/2020/ar/Expedia-Group-Annual-Report.pdf.
[5] F. Maxwell Harper and Joseph A. Konstan. 2015. The MovieLens Datasets: History and Context. ACM Trans. Interact. Intell. Syst. 5, 4, Article 19 (Dec.
2015), 19 pages. https://doi.org/10.1145/2827872
[6] Adam Woznica and Jan Krasnodebski. 2013. Personalize Expedia Hotel Searches - ICDM 2013 Learning to rank hotels to maximize purchases.
https://www.kaggle.com/c/expedia-personalized-sort.
[7] Adam Woznica and Jan Krasnodebski. 2016. Expedia Hotel Recommendations. Which hotel type will an Expedia customer book? https://www.
kaggle.com/c/expedia-hotel-recommendations.
Copyright © 2021 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).