Professional Documents
Culture Documents
Artificial Intelligence and Machine Learning in the Travel Industry: Simplifying Complex Decision Making Ben Vinod full chapter instant download
Artificial Intelligence and Machine Learning in the Travel Industry: Simplifying Complex Decision Making Ben Vinod full chapter instant download
Artificial Intelligence and Machine Learning in the Travel Industry: Simplifying Complex Decision Making Ben Vinod full chapter instant download
https://ebookmass.com/product/analytics-enabled-decision-making-
vinod-sharma/
https://ebookmass.com/product/risk-modeling-practical-
applications-of-artificial-intelligence-machine-learning-and-
deep-learning-terisa-roberts/
https://ebookmass.com/product/artificial-intelligence-and-
machine-learning-for-edge-computing-1st-edition-rajiv-pandey/
https://ebookmass.com/product/sustainable-developments-by-
artificial-intelligence-and-machine-learning-for-renewable-
energies-krishna-kumar/
Visualization Techniques for Climate Change with
Machine Learning and Artificial Intelligence Arun Lal
Srivastav
https://ebookmass.com/product/visualization-techniques-for-
climate-change-with-machine-learning-and-artificial-intelligence-
arun-lal-srivastav/
https://ebookmass.com/product/machine-learning-for-decision-
makers-cognitive-computing-fundamentals-for-better-decision-
making-2nd-edition-patanjali-kashyap/
https://ebookmass.com/product/machine-learning-for-decision-
makers-cognitive-computing-fundamentals-for-better-decision-
making-2nd-edition-patanjali-kashyap-2/
https://ebookmass.com/product/predicting-heart-failure-invasive-
non-invasive-machine-learning-and-artificial-intelligence-based-
methods-1st-edition-kishor-kumar-sadasivuni/
https://ebookmass.com/product/automation-and-autonomy-labour-
capital-and-machines-in-the-artificial-intelligence-industry-1st-
ed-2021-edition-steinhoff/
Artificial Intelligence and
Machine Learning in the
Travel Industry
Simplifying Complex
Decision Making
Edited by
Ben Vinod
Artificial Intelligence and Machine Learning
in the Travel Industry
Ben Vinod
Editor
Spin-off from Journal: “Artificial Intelligence and Machine Learning in the Travel Industry” Volume 20, Issue 3, March 2021
ISBN 978-3-031-25455-0
© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned,
specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter
developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific
statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of
publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for
any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional
affiliations.
This Palgrave Macmillan imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Contents
v
vi Contents
EDITORIAL
Over the past decade, Artificial Intelligence has proved efforts. Companies invest broadly across different platforms
invaluable in a range of industry verticals such as auto- such as Google or Facebook, and it is important to under-
motive and assembly, life sciences, retail, oil and gas, and stand the marginal impact of a specific marketing program
travel. The leading sectors adopting AI rapidly are Finan- or initiative on revenue. He uses a novel approach by casting
cial Services, Automotive and Assembly, High Tech and the well-known Koyck distributed lag model in state space
Telecommunications. Travel has been slow in adoption, but form to analyze the effectiveness of each marketing channel
the opportunity for generating incremental value for AI over and subsequent allocation of marketing budgets.
other analytics is extremely high (Chui et al. 2018). The paper by Ravi Kumar, Wei Wang, Ahmed Simrin,
In September 2019, Ian Yeoman and I discussed creat- Sivarama Krishnan Arunachalam and Bhaskar Rao Gun-
ing a special issue for the Journal of Revenue and Pricing treddy and Darius Walczak on competitive revenue manage-
Management on Artificial Intelligence in Travel. Information ment models is a collaboration between PROS and Etihad
from airlines and vendors on AI in travel has been sporadic, Airways. Their paper proposes a demand model that captures
usually discussed at industry conferences. Yet it was abun- realistic competitive dynamics by considering two types of
dantly clear to me based on my interactions with travel sup- customer behaviors: airline’s loyal customers who prefer to
pliers, software vendors, OTAs and GDSs that they were lev- buy from the airline even if their price is not the lowest in the
eraging core concepts in Artificial Intelligence and Machine market and fully flexible customers who buy the lowest fare
Learning to create new value propositions or improve on in the market. They develop a Bayesian machine learning-
existing applications related to travel. This was an opportu- based demand forecasting methodology for these models in
nity to showcase in a single issue the breadth and scope of both class-based and class-free settings that explicitly con-
what individuals in these organizations were focused on with siders competitive market information.
applications and business process. Norbert Remenyi and Xiaodong Luo from Sabre discuss
The research papers… practical limitations of the choice-based demand models
An excellent contribution from Rodrigo Acuna-Agost, found in the literature to estimate demand from sales trans-
Eoin Thomas and Alix Lh’eritier from Amadeus, who action data. They propose modifications and extensions
propose a new method to estimate price elasticity for deep under partial availability and extend the Expected Maximi-
learning-based choice models with an excellent set of refer- zation (EM) algorithm for nonhomogeneous product sets.
ences. The insights they provide are particularly relevant for The data preprocessing and solution techniques are useful
airline offers based on customer segment and context. for practitioners.
Ahmed Abdelghany and Ching-Wen Huang from Embry- The practice papers…
Riddle University and Khaled Abdelghany from Southern The paper on recommender systems by Amine Dadoun,
Methodist University propose a novel reinforcement learning Michael Defoin Platel, Thomas Fiig, Corinne Landra and
approach to calibrate itinerary choice models and measure Raphael Troncy from Amadeus highlight the central role of
schedule profitability. recommender systems to create personalized offers and its
Melvin Woodley from Sabre solves the attribution prob- growing importance with IATA’s New Distribution Capa-
lem of associating sales or revenue to individual marketing bility messaging standard. It is a well-researched paper and
truly relevant to the future of airline retailing.
Michael Byrd from Yum! and Ross Darrow from Charter
* B. Vinod and Go make the case for contextual bandits, a reinforce-
benvinod@yahoo.com ment learning technique for personalizing offers in retailing.
1
Southlake, USA They provide insights into the use of Thompson sampling, a
Chapter 1 was originally published as Vinod, B. Journal of Revenue and Pricing Management (2021) 20:211–212. https://doi.org/10.1057/s41272-021-
00307-0.
popular exploration heuristic and how they can be deployed. Foremost on the minds of corporations as they leverage
They discuss the step improvement that can be achieved with AI for competitive advantage is how to scale AI across the
contextual bandits, despite greater computational complexity organization. Deborah Leff and Kenneth Lim from IBM
incurred when contextual features are included in the model. draw upon their extensive experience working with many
Tomasz Szymanski from Nordea Bank and Ross Darrow companies to provide insights into the various organizational
from Charter and Go discuss the important topic of shelf barriers to scale AI, the importance of executive sponsorship
placement on agency storefronts. While airlines focus on and recommend best practices. This paper is a “must read”
offer creation, the GDS desktop must display non-homoge- for anyone who is a practitioner of AI.
nous content that is addressed in this paper. A shelf product The futures article…
assortment method is proposed for categorizing airline offers Ross Darrow’s future’s article is thought provoking.
into utility levels, thus facilitating the itinerary selection pro- “The Future of AI is the Market” paints a picture of how the
cess for travelers. future travel distribution landscape will be influenced by
Jian Wang from Realpage outlines a practical application interactions in the marketplace and less on targeted one-off
of reinforcement learning used to determine reference rents solutions.
for apartments. He demonstrates how the new approach out- I would like to take this opportunity to thank all the anon-
performs the traditional rules-based approach. ymous referees I reached out to over the past few months
Shriguru Nayak, Nitin Gautam and Sergey Shebalov from to provide feedback on the submitted papers. This special
Sabre apply machine learning models to estimate market issue would not have been possible without your feedback
size and market share from competitive future schedules and and requests for revisions.
augmented data sources, a key component for developing
airline schedules. They also discuss how revenue manage-
ment practices can be improved with access to data from
network planning. Reference
The paper by Rimo Das, Harshinder Chaddha and Som-
nath Banerjee from LodgIQ focuses on forecasting market Chui, M., R. Chung, N. Henke, S. Malhotra, J. Manyika, M. Miremadi,
and P. Nel. 2018. Notes from the AI Frontier: Applications and
demand considering seasonality and market events. They Value of Deep Learning. McKinsey.com, April 2018.
examine a variety of machine learning techniques that the
data were calibrated upon and report on the accuracy of the Publisher’s Note Springer Nature remains neutral with regard to
forecasts. jurisdictional claims in published maps and institutional affiliations.
In January 2018, an AI initiative was established at Sabre
to identify industry-relevant problems suitable for AI-based
solutions, raise internal awareness and accelerate adoption. B. Vinod serves as Chief Scientist and Senior Vice President at
This initiative also led to the creation and distribution of an Sabre (2008–2020). Before rejoining Sabre in 2004, he was Vice
President at Sabre Airline Solutions, responsible for Pricing and Yield
internal AI newsletter, quarterly town halls to monitor pro-
Management.
gress and discuss use cases that I was responsible for. My
contribution to the special issue reflects this initiative and
steps taken to solve a range of problems in travel.
RESEARCH ARTICLE
Received: 10 May 2020 / Accepted: 4 September 2020 / Published online: 22 March 2021
© The Author(s), under exclusive licence to Springer Nature Limited 2021
Abstract
One of the most popular approaches to model choices in the airline industry is the multinomial logit (MNL) model and its
variations because it has key properties for businesses: acceptable accuracy and high interpretability. On the other hand,
recent research has proven the interest of considering choice models based on deep neural networks as these provide better
out-of-sample predictive power. However, these models typically lack direct business interpretability. One useful way to
get insights for consumer behavior is by estimating and studying the price elasticity in different choice situations. In this
research, we present a new methodology to estimate price elasticity from Deep Learning-based choice models. The approach
leverages the automatic differentiation capabilities of deep learning libraries. We test our approach on data extracted from
a global distribution system (GDS) on European market data. The results show clear differences in price elasticity between
leisure and business trips. Overall, the demand for trips is price elastic for leisure and inelastic for the business segment.
Moreover, the approach is flexible enough to study elasticity on different dimensions, showing that the demand for business
trips could become highly elastic in some contexts like departures during weekends, international destinations, or when the
reservation is done with enough anticipation. All these insights are of a particular interest for travel providers (e.g., airlines)
to better adapt their offer, not only to the segment but also to the context.
Keywords Price elasticity · Discrete choice modeling · Deep learning · Interpretability · Automatic differentiation · Travel
industry
Chapter 2 was originally published as Acuna‑Agost, R., Thomas, E. & Lhéritier, A. Journal of Revenue and Pricing Management (2021) 20:213–226.
https://doi.org/10.1057/s41272-021-00308-z.
Fig. 1 The proposed methodology to extract elasticity estimates from choice models. Note that the enrichment process is optional, and here it is
used to describe additional attributes of the data that are added for analysis, but not considered as features by the choice model
Natural evolutions of the MNL model deal with these (Lundberg and Lee 2017). A prediction can be explained by
limitations. For example, the mixed-MNL model allows assuming that each feature value of the instance is a player
grouping the decision makers by segments (Hayden in a game where the prediction is the payout. Shapley values
Boyd and Mellman 1980) and, similarly, the nested logit determine how to fairly distribute the payout among the fea-
model allows grouping alternatives by nests (McFadden tures. However, both LIME and Shapley values are relatively
1980). More recently, (Lhéritier et al. 2019) uses a machine expensive to compute, as they require a sampling of data
learning approach to allow an arbitrarily complex util- based on the decision to be explained.
ity function that can also depend on the decision maker’s Furthermore, the aim of this study is to extract mean-
attributes. ingful economic information, such as price elasticity1 from
Deep learning is a branch of machine learning that uses complex choice models, which are not provided by generic
artificial neural networks to model functions via an arbitrary model interpretation techniques. The proposed methodol-
number of composed transformations allowing to achieve ogy is summarized in Fig. 1. The first step is to estimate the
high performance in a large variety of tasks (see, e.g., Good- choice probabilities by the choice prediction model, then
fellow et al. 2016). Recently, some works used such approach elasticities are extracted for all the alternatives and choice
to build more flexible choice models. In Mottini and Acuna- situations. In order to give meaningful interpretation a filter-
Agost (2017), the authors propose a sequence transforma- ing step is needed to discard all non-relevant alternatives.
tion approach to define the choice probability allowing to Finally, the analysis is based on different aggregations using
condition on the decision maker’s attributes. More recently, a representative value (e.g., the median) per several dimen-
Lhéritier (2020) uses a deep learning approach to parameter- sions (e.g., customer segments and other features of interest
ize the flexible class of Pairwise Choice Markov Chains that depending on the application).
allows to escape traditional choice-theoretic assumptions The structure of the paper is as follows. We initially intro-
such as IIA, stochastic transitivity and regularity (Ragain duce the economic concept of elasticity and give an over-
and Ugander 2016). view of how it can be estimated for deep learning models.
For the application of air itinerary choice prediction, there We then present the motivating application of air itinerary
is a high interest in better understanding the choices of trave- choice modelling. This is followed by the numerical analysis
lers. This application of choice models can have important of elasticities for flight choices, and finally, we present future
impacts on revenue from using the most accurate models, research directions and conclusions.
but also by being able to get actionable insights that travel
providers like airlines or travel agencies could exploit to
improve their business metrics and offers. Elasticity estimation from neural
The high predictive performance of neural networks network‑based choice models
comes at the cost of a difficult interpretation of the models,
which has sparked research into complementary techniques In this section, we first introduce the economic concept of
to some shed light on how inputs attributes influence the elasticity and then we show how it can be computed on com-
outputs of the model. Various methods for performing fea- plex deep neural network-based choice models.
ture importance have been proposed (Molnar 2019), as well
as more specific methods to explain individual decisions.
Local interpretable model explanations (LIME) (Ribeiro
et al. 2016) can be used to explain a specific decision, by
building a linear model from samples close in the feature
space to the target sample. Shapley values have also been 1
Elasticity relates to the relative change of one variable (e.g.,
proposed as a method inspired by coalitional game theory demand) to the relative change in another variable (e.g., price).
Airline itinerary choice dataset with price Price elasticity estimation from airline itinerary
elasticity estimation and trip purpose choices
segmentation
A passenger name record (PNR) contains relevant data
Nowadays travelers have higher expectations and choice regarding travel bookings, such as flight information of
than in the past. This is mostly driven by the experience each segment of a journey and information about the indi-
they already have in other industries, in particular online vidual, as well as information about ancillary services and
retail. Some examples of the new standards are: relevant special service requests. In order to obtain a full choice set,
and timely recommendations, fully customized products
and services, transparent pricing, modern search, shopping
cart functionalities on different channels (e.g., mobile and Table 1 Features of the airline itinerary choice dataset
desktop), and high flexibility (e.g., be able to cancel sub- Type Feature Range/cardinality
scriptions at any time or to send back products and being
fully reimbursed). Many, if not all, of these improvements Individual Cat.Origin/destination 97
observed in retail have been boosted by leveraging data and Search office 11
newer algorithms thanks to machine learning, and naturally Num. Departure weekday [0, 6]
the travel industry is following this trend too. Stay Saturday [0, 1]
Figure 3 summarizes the application of the proposed Continental trip [0, 1]
methodology for getting elasticities and business insights Domestic trip [0, 1]
from air itinerary choice models. The remainder of this sec- Days to departure [0, 343]
tion presents the methodology to obtain this enriched dataset Alternative Cat. Airline (of first flight) 63
of chosen alternatives with elasticity and trip purpose esti- Num. Price [77.15, 16,781.50]
mates. The aggregations and business insights are presented Stay duration (min) [121, 434,000]
in the results section of this paper. Trip duration (min) [105, 4314]
Number connections [2, 6]
Number airlines [1, 4]
Outbound departure time [0, 84,000]
(in s)
Outbound arrival time (in s) [0, 84,000]
data from PNRs are matched with search log activity, which interested in price elasticity, that is, the elasticity of the
shows all available options presented to the traveller prior demand, estimated by the probability of being chosen, with
to booking. respect to the price. Moreover, we consider only one alterna-
In this experiment, the dataset from Mottini and Acuna- tive per choice situation. For each choice situation, we take
Agost (2017) consisting of flight bookings sessions on a set the alternative with the largest probability of selection as
of European origins and destinations is used. Each choice estimated by the PCMC-Net model. It should be noted that
session contains up to 50 different proposed itineraries, this alternative does not necessarily match the alternative
one of which has been booked by the customer. There are that was chosen by the consumer. This decision was taken as
815,559 distinct alternatives among which 84% are single- our main goal is to give explainability to the neural network-
tons and 99% are observed at most seven times. In total, based model, rather than understanding individual choices.
there are 33,951 choice sessions of which 27160 were used
for training and 6791 for testing. The dataset has a total of Trip purpose segmentation: motivation
13 features, both numerical and categorical, corresponding
to individuals and alternatives, as shown in Table 1. The first step to better modeling the traveller decision-
Choice models are important in helping to select, high- making process is to understand the reason why the traveler
light and rank different offers. Several methods have previ- would like to travel, which we will refer to hereafter as the
ously been suggested, the performance of which is shown trip purpose. If travel providers could get this information
in Fig. 4 on a common training and test dataset. The metric accurately at shopping time (i.e., before the booking), they
used here is the Top 1 accuracy, which measures the per- could greatly improve the shopping experience: offer the
centage of sessions for which the most probable alternative best product, at the best price, at the best moment, to the
identified by the model is indeed chosen by the user. Each targeted customer.
choice set contains 50 alternatives, thus a uniform sampling Business trips are driven by convenience and usually sub-
method achieves 2% Top 1 accuracy. Simple heuristics such ject to companies’ travel policies (i.e., the passenger do not
as selecting the cheapest offer can be used to give a baseline pay for this trip, but her company). It is also the case that
performance for the problem. As can be seen, non-linear sometimes the passenger has a less active role in the deci-
methods such as Deep Pointer networks, Latent Class MNL sion of the trip as the task is delegated to travel arrangers
and Random Forests give better performance than the linear such as travel agencies or assistants. The authors in Teichert
MNL model. However, PCMC-net results in the best Top 1 et al. (2008) confirm that people traveling for business have a
accuracy of all the methods tested. For more details on each strong correlation with these attributes: efficiency, punctual-
method, the reader is referred to Lhéritier (2020), Lhéritier ity, and flexibility.
et al. (2019) and Mottini and Acuna-Agost (2017). On the other hand, leisure trips are driven mostly by
Elasticity can be calculated for any reference value and price (Teichert et al. 2008). Although in some cases, the
for any alternative in the choice set. We are particularly passengers are sufficiently wealthy that they may give more
importance to comfort and efficiency. In practice, many trips are labeled as business travel. From this set of 400,000 book-
are not easily classified on exactly one of these two seg- ings, 40,000 random samples are held out as the test set for
ments as they can be both at the same time: bleisure trips the classification task.
(i.e., extending a business trip for leisure activities) (Vivion For each booking there are 48 features available corre-
2016). sponding to all non-sensitive attributes of the trip. These
As an important question for the industry, it is not surpris- relate to the origin, destination, route, carrier, various
ing that the problem has been addressed previously (Teichert aspects linked to the time and duration of travel along with
et al. 2008; Chatterjee et al. 2020; Tahanisaz and shokuhyar aspects of the booking such as the number of passengers
2020; Martinez-Garcia and Royo-Vela 2010; Jin-Long 2017; in the booking, the number of days prior to boarding that
Dresner 2006; Vinod 2008). Most of the previous work is the trip was booked, etc. A comparison of different models
based on stated preference surveys (Dresner 2006; Martinez- based on various feature subsets is provided in Appendix A.
Garcia and Royo-Vela 2010; Tahanisaz and shokuhyar 2020) In order to apply the segmentation of bookings to the
i.e., asking current or potential travelers about their prefer- choice dataset, only features common to both the bookings
ences and the reason of the trips. This kind of approach and choice datasets are selected. These are the origin and
brings a lot of flexibility in terms of the type of questions destination, international/domestic, stay duration, days to
as for example the analysts could even ask about hypotheti- departure, the day of the week for the booking, outbound
cal scenarios. It is well known that stated preference data flight and return flight and a stay Saturday feature.
present a series of inconveniences (Abdullah et al. 2011), A gradient boosting machine model (Friedman 2001)
for example, their incapacity to capture accurately all the is used as a classifier, with grid search and early stopping
market and personal limitations that occurs in the real world. selecting a maximum depth of 6 for 46 trees. On the hold-out
Another limitation of these works based on surveys is that test set, the model obtains 84.20% accuracy.2 Feature impor-
the conclusions are drawn based on relatively small amount tance analysis suggests that the stay duration, destination and
of data [around 3000 in Dresner (2006), 300 in Tahanisaz origin airports, stay Saturday, return day of week and days
and shokuhyar (2020), 808 in Martinez-Garcia and Royo- to departure are the most important features, respectively.
Vela (2010), and 5800 in Teichert et al. (2008)]. Note that this segmentation is not used as an input to the
Therefore, to adequately discuss the price elasticities choice model, but only to segment the dataset for analysis
obtained by the choice model, these should be done in the purposes (see Fig. 3).
context of business and leisure trip independently. However,
this information is not available from the choice dataset, and
as such must be inferred. Analysis and business insights
Trip purpose segmentation: dataset In this section, we analyze and discuss the elasticities
and experimental protocol obtained by the approach presented previously using differ-
ent aggregations.
In this section, we present an analysis of business vs leisure In order to aggregate the estimated elasticities, we use the
segmentation performed over a set of labeled bookings from median since it is a measure of central tendency robust with
a larger set of unlabeled data. The bookings correspond to respect to outliers and skewed data (see Fig. 5). Another
indirect bookings made by customers at traditional travel important element to remark is that the analyses are based on
agencies, online travel agencies and travel management com- the additive inverse of elasticity because the price elasticities
panies (among others) which are then processed by a GDS. are usually negative. This transformation helps to construct
We can consider the dataset as being partially labeled, as readable charts following the convention of economists that
most offices are identified as belonging to particular market are interested on the absolute values (magnitudes) instead
segments which deal almost exclusively with either business of the real number.
travel or leisure travel. Overall elasticity The median value for whole data is
The dataset used in this trip purpose model is a sam- − 1.73 which can be classified as elastic. Note that this value
ple of indirect bookings from the European market for the is in the same order of magnitude to the values published in
full year 2019. A balanced dataset is obtained by sampling previous works.
200,000 bookings made from offices tagged as “Retail- small Elasticity by trip purpose In order to understand better
medium enterprises” which are labeled as leisure travel, this value, we analyze the two main customer segments in
and sampling another 200,000 bookings made from offices
tagged as “Global Travel Management Companies” which
2
Accuracy: ratio of number of correct predictions to the total num-
ber of input samples.
Fig. 5 Distribution of elasticity values for both segments: Leisure additive inverse of the elasticity that is usually negative in its origi-
(left) and Business (right). The distributions are not symmetric, and nal form. It should be also noted the presence of few negative values
in both segments there is a peak of observations near to zero. Note (a) correspond to rare observations. These choices could be explained
we analyze one alternative per choice situation, the one with the high- by Veblen or Giffen behaviors of some consumers in certain circum-
est probability, (b) all charts (and this one in particular) present the stances
Fig. 7 Median elasticity (neg) for both segments (Business and Leisure) on different departure days of the week
in the near future (short-term) show very low price elas- Saturday and in blue for when the traveler did not stay at
ticities, probably explained as trips booked for closer dates destination a full Saturday. Note that the blue curve rises
are particularly related to urgent matters, in those cases the in a linear fashion, indicating that stay duration is directly
price is less relevant than other aspects like the schedule proportional to price elasticity for trips which return dur-
or the total trip duration. This is consistent with previ- ing the same week as departure (this also includes trips
ous research where elasticity was reported in the interval leaving Sunday and returning prior to the following Sat-
𝜖 ∈ {− 2.0, − 0.5} for DTD ∈ {2, 21} days (see Morlotti et al. urday). For trips including Saturday stays (in orange),
2017). Our results extend the previous results showing that all trips are highly price elastic, but it does appear that
the absolute value continues to increase to larger values until for stay durations between 1 and 3 nights are less price
100 days approximately. The chart also shows an increase on elastic. These short trips always contain a Saturday night
the dispersion of values for larger DTD, which is explained stay, and therefore often correspond to weekend geta-
mainly by the number of observations used to calculate the ways and possibly city breaks. Such trips are often to
median values (represented by the darkness of the line). geographically closer destinations, which are associated
Elasticity by stay duration An important criterion with cheaper prices overall, and thus price may be a less
in price elasticity is the stay duration, especially when important factor than for longer duration trips which can
weekend stays are factored into the analysis. In Fig. 10, be associated with a higher overall budget. Furthermore,
the price elasticity is shown as a function of stay dura- for such short trips, other factors such as time of arrival
tion in orange for trips where the traveler stayed an entire and departure may be more important to the travelers in
Fig. 9 Median elasticity (neg) for both segments: leisure (left) and business (right) as a function of different advance purchase days (a.k.a. days
to departure)
be applied to a choice dataset used in the main article. This Gradient boosting models have proven particularly adept
choice data profile segmentation model is effectively a com- at classification, here the H2O.ai library is used to train
promise in performance, due to only some features overlap- the models and determine the feature importance (Candel
ping between the bookings dataset and the choice dataset. and Malohlava 2020). The training phase uses a hold-out
Note that the H2O package was used to train the gradient boosting
machines in order to obtain best performance, whereas Scikit-learn
was used to train a decision tree that could be visualized and inter-
preted