Professional Documents
Culture Documents
Artificial Intelligence and Machine Learning in The Travel Industry Simplifying Complex Decision Making Ben Vinod Full Chapter
Artificial Intelligence and Machine Learning in The Travel Industry Simplifying Complex Decision Making Ben Vinod Full Chapter
Artificial Intelligence and Machine Learning in The Travel Industry Simplifying Complex Decision Making Ben Vinod Full Chapter
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
We note that the best single tree performance obtained with Chang, Li-Yen, and Pei-Yu Sun. 2012. Stated-choice analysis of
more complex trees was 82.4% accuracy. willingness to pay for low cost carrier services. Journal of Air
Transport Management 20: 15–17.
The resulting model is visualised in Fig. 11. It can be seen Chatterjee, Sujoy, Nicolas Pasquier, Simon Nanty, and Maria A. Zul-
that the first split reflects stay duration, indicating that travel- uaga. 2020. Multi-objective consensus clustering framework for
ers who stayed for more than 5.9 days at destination were flight search recommendation. arxiv:2002.10241.
primarily leisure travelers (87.7). Travelers who stayed less Dorman, Peter. 2014. Microeconomics—A fresh start. Cham:
Springer.
than 5.9 days, but whose destination was more than 1163 Dresner, Martin. 2006. Leisure versus business passengers: Simi-
km from the origin were also predominantly leisure travel- larities, differences, and implications. Journal of Air Transport
ers. For short duration and short distance trips, the main Management, 12 (1): 28–32. Leisure Traffic and Tourism: New
differentiator appears to be the number of passengers in the Strategies for Airlines, Airports and the Travel Trade.
Duncan Luce, R. 1977. The choice axiom after twenty years. Journal
booking, with solo travelers being more likely to be busi- of Mathematical Psychology 15 (3): 215–233.
ness. For most other conditions in the tree, we see the output Friedman, Jerome H. 2001. Greedy function approximation: A gradi-
is mostly business or for a small number of samples leisure ent boosting machine. Annals of Statistics 29 (5): 1189–1232.
but with low certainty. Based on the lowest Gini value, we Garrow, Laurie A 2016. Discrete choice modelling and air travel
demand: Theory and applications. London: Routledge.
can conclude that bookings will be classified as business for Ghoshal, Animesh. 1981. Price elasticity of demand for air passenger
short durations ( < 5.9 days), short distances (< 1163 km), service: Some additional evidence. Transportation Journal 20
solo travelers, booking before the day of travel, not staying (4): 93–96.
Saturday and flying domestically. Goodfellow, Ian, Yoshua Bengio, and Aaron Courville. 2016. Deep
learning. Cambridge: MIT.
Granados, Nelson, Alok Gupta, and Robert J. Kauffman. 2012.
Online and offline demand and price elasticities: Evidence from
References the air travel industry. Information Systems Research 23 (1):
164–181.
Abadi, Martín, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Granados, Nelson, Robert J. Kauffman, Hsiangchu Lai, and Huang-chi
Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Geof- Lin. 2012. À la carte pricing and price elasticity of demand in air
frey Irving, and Michael Isard, et al. 2016. Tensorflow: A sys- travel. Decision Support Systems 53 (2): 381–394.
tem for large-scale machine learning. In 12th {USENIX} Sympo- Hayden Boyd, J., and Robert E. Mellman. 1980. The effect of fuel
sium on operating systems design and implementation ({OSDI} economy standards on the us automotive market: An hedonic
16), 265–283. demand analysis. Transportation Research Part A: General 14
Abdullah, Sabah, Anil Markandya, and Paulo A.L.D. Nunes. 2011. (5–6): 367–378.
Introduction to economic valuation methods. In Research tools Heijman, Wim, and Pierre Mouche. 2011. New insights into the theory
in natural resource and environmental economics, 143–187. of Giffen goods, vol. 655. Berlin: Springer.
Singapore: World Scientific. Hruschka, Harald Werner Fettes, and Markus Probst. 2001. Analyzing
Baydin, Atılım Günes, Barak A. Pearlmutter, Alexey Andreyevich purchase data by a neural net extension of the multinomial logit
Radul, and Jeffrey Mark Siskind. 2017. Automatic differentia- model. In International conference on artificial neural networks,
tion in machine learning: A survey. The Journal of Machine 790–795. Cham: Springer.
Learning Research 18 (1): 5595–5637. Jin-Long, Lu. 2017. Segmentation of passengers using full-service and
Block, H.D., and Jacob Marschak. 1960. Random orderings and sto- low-cost carriers - evidence from taiwan. Journal of Air Transport
chastic theories of response. Contributions to Probability and Management 62: 204–216.
Statistics 2: 97–132. Jung, J.M. and E.T. Fujii. 1976. The price elasticity of demand for air
Bondoux, Nicolas, Anh Quan Nguyen, Thomas Fiig, and Rodrigo travel: some new evidence. Journal of Transport Economics and
Acuna-Agost. 2020. Reinforcement learning applied to airline Policy 10: 257–262.
revenue management. Journal of Revenue and Pricing Manage- Kolmar, Martin. 2017. Principles of microeconomics. Cham: Springer.
ment 19: 332–348. Lhéritier, Alix. 2020. PCMC-Net: Feature-based pairwise choice
Brons, Martijn, Eric Pels, Peter Nijkamp, and Piet Rietveld. 2002. markov chains. In International conference on learning represen-
Price elasticities of demand for passenger air travel: A meta- tations (ICLR 2020), Addis Ababa, Ethiopia, 26–30 April 2020.
analysis. Journal of Air Transport Management 8 (3): 165–175. Lhéritier, Alix, Michael Bocamazo, Thierry Delahaye, and Rodrigo
Candel, A., and M. Malohlava. 2020. Gradient boosted models. R Acuna-Agost. 2019. Airline itinerary choice modeling using
package version 3 (0.4). machine learning. Journal of Choice Modelling 31: 198–209.
Carlos Martín, Juan, Concepción Román, and Raquel Espino. 2008. Luce, R. Duncan. 1959. Individual choice behavior: A theoretical
Willingness to pay for airline service quality. Transport Reviews analysis. New York: Wiley.
28 (2): 199–217. Lundberg, Scott M., and Su-In Lee. 2017. A unified approach to inter-
Carlsson, Fredrik. 1999. Private vs. business and rail vs. air passen- preting model predictions. In Advances in neural information
gers: willingness to pay for transport attributes. Working Papers processing systems, 4765–4774.
in Economics No. 14. Lu, Jin-Long., and Zhang Yi Shon. 2012. Exploring airline passengers’
Castelli, Lorenzo Walter Ukovich, and Raffaele Pesenti. 2003. An willingness to pay for carbon offsets. Transportation Research
airline-based multilevel analysis of airfare elasticity for passen- Part D: Transport and Environment 17 (2): 124–128.
ger demand. In Air Transport Research Society (ATRS) world Martinez-Garcia, Esther, and Marcelo Royo-Vela. 2010. Segmenta-
conference tion of low-cost flights users at secondary airports. Journal of Air
Transport Management 16 (4): 234–237.
McFadden, Daniel, et al. 1973. Conditional logit analysis of qualita- Spiegel, Uriel. 1994. The case of a “giffen good”. The Journal of Eco-
tive choice behavior. Berkeley: Institute of Urban and Regional nomic Education 25 (2): 137–147.
Development, University of California. Strauss, Arne K., Robert Klein, and Claudius Steinhardt. 2018. A
McFadden, Daniel. 1974. The measurement of urban travel demand. review of choice-based revenue management: Theory and meth-
Journal of public economics 3 (4): 303–328. ods. European Journal of Operational Research 271 (2): 375–387.
McFadden, Daniel. 1980. Econometric models for probabilistic choice Tahanisaz, Sahar, and Sajjad shokuhyar. 2020. Evaluation of passenger
among products. Journal of Business 53 (3): S13–S29. satisfaction with service quality: A consecutive method applied
Merkert, Rico, and Matthew Beck. 2017. Value of travel time sav- to the airline industry. Journal of Air Transport Management 83:
ings and willingness to pay for regional aviation. Transportation 101764.
Research Part A: Policy and Practice 96: 29–42. Teichert, Thorsten, Edlira Shehu, and Iwan von Wartburg. 2008. Cus-
Molnar, Christoph. 2019. Interpretable machine learning. https://chris tomer segmentation revisited: The case of the airline industry.
tophm.github.io/interpretable-ml-book/. Transportation Research Part A: Policy and Practice 42 (1):
Morlotti, Chiara, Mattia Cattaneo, Paolo Malighetti, and Renato 227–242.
Redondi. 2017. Multi-dimensional price elasticity for leisure and Tsamboulas, Dimitrios A., and Anastasios Nikoleris. 2008. Passen-
business destinations in the low-cost air transport market: Evi- gers’ willingness to pay for airport ground access time savings.
dence from easyjet. Tourism Management 61: 23–34. Transportation Research Part A: Policy and Practice 42 (10):
Mottini, Alejandro, and Rodrigo Acuna-Agost. 2017. Deep choice 1274–1282.
model using pointer networks for airline itinerary prediction. In Veblen, Thorstein. 1899. The Theory of the Leisure Class. New York:
Proceedings of the 23rd ACM SIGKDD international conference The Macmillan Company.
on knowledge discovery and data mining, 1575–1583. Vinod, B. 2008. The continuing evolution: Customer-centric revenue
Njegovan, Nenad. 2006. Elasticities of demand for leisure air travel: management. Journal of Revenue Pricing Management 7: 27–39.
A system modelling approach. Journal of Air Transport Manage- Vivion, Nick. 2016. Bleisure travel: The benefit of mixing business
ment 12 (1): 33–39. travel with leisure. Online (visited on March 2, 2020). https://
Oum, Tae Hoon, William G. Waters, and Jong-Say. Yong. 1992. Con- www.sabre.com/insights/ble-travel-the-benefi ts-of-mixing-busin
cepts of price elasticities of transport demand and recent empirical ess-travel-with-leisure.
estimates: An interpretative survey. Journal of Transport Econom-
ics and Policy 26 (2): 139–154. Publisher’s Note Springer Nature remains neutral with regard to
Paszke, Adam, Sam Gross, Francisco Massa, Adam Lerer, James jurisdictional claims in published maps and institutional affiliations.
Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Nata-
lia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative
style, high-performance deep learning library. In Advances in
neural information processing systems, 8024–8035. Red Hook: Rodrigo Acuna‑Agost serves as the Head of Artificial Intelligence,
Curran Associates. Research and Technology team at Amadeus. Before joining Amadeus
Pedregosa, Fabian, Gaël. Varoquaux, Alexandre Gramfort, Vincent in 2010, he worked on applications of Operations Research and AI for
Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter different industries, including manufacturing and travel.
Prettenhofer, Ron Weiss, Vincent Dubourg, et al. 2011. Scikit-
learn: Machine learning in python. The Journal of Machine Eoin Thomas is with Amadeus since 2014, and currently part of the
Learning Research 12: 2825–2830. AI, Research and Technology team, exploring how machine learning
Ragain, Stephen, and Johan Ugander. 2016. Pairwise choice markov can help the travel industry. Prior to this, he completed his Ph.D. in
chains. In Advances in neural information processing systems, University College Cork, Ireland and a postdoctoral contract with Inria,
3198–3206. France in the domain of biomedical signal processing.
Ribeiro, Marco Tulio, Sameer Singh, and Carlos Guestrin. 2016. Why
should I trust you? Explaining the predictions of any classifier. In Alix Lhéritier is with Amadeus since 2015, working on machine learn-
Proceedings of the 22nd ACM SIGKDD international conference ing research and applications for the travel industry. He was previously
on knowledge discovery and data mining, 1135–1144. with Inria, France as a doctoral student and received his doctorate in
Richard, David B. 2009. The changing price elasticity of demand for computer science from the Université Nice Sophia Antipolis.
domestic airline travel. In 50th Annual transportation research
Forum, Portland, Oregon.
Schiff, Aaron, and Susanne Becken. 2011. Demand elasticity estimates
for new zealand tourism. Tourism Management 32 (3): 564–575.
RESEARCH ARTICLE
Received: 10 April 2020 / Accepted: 19 September 2020 / Published online: 15 March 2021
© The Author(s), under exclusive licence to Springer Nature Limited 2021
Abstract
This paper presents a novel methodology to develop itinerary choice models (ICM) for air travelers that addresses the limita-
tions of the traditional utility-maximization approach. The methodology integrates a reinforcement learning algorithm and
an airline network competition analysis model. The reinforcement learning algorithm searches for the values of parameters
of the itinerary choice model while considering maximizing a reward function. The reward function is measured as the nega-
tive of the difference between the estimated and observed system metrics. The airline network competition analysis model
is used to calculate the estimated system metrics. It is a simulation model that represents passenger-itinerary assignment.
It captures the demand–supply interactions at the network level while considering the competition among all airlines. An
ICM system is calibrated using the developed framework considering the global airline network, which includes more than
500,000 airport pairs. Validating the model against ground truth data shows that the developed model adequately captures
the travelers’ itinerary choice behavior and replicates the competition pattern among airlines.
Keywords Airlines · Competition · Itinerary choice · Reinforced learning · Simulation · Market shares
(1992), Proussaloglou and Koppelman (1995), Yoo and Ash- air-travel itinerary choice behavior. The methodology inte-
ford (1996), Algers and Beser (1997), Corsi et al. (1997), grates a reinforcement learning (RL) algorithm and an air-
Proussaloglou and Koppelman (1999), Suzuki et al. (2001), line network competition analysis model. The RL algorithm
Coldren et al. (2003), Coldren and Koppelman (2005), War- searches for the values of parameters of the itinerary choice
burg et al. (2006), Carrier (2008), Hess et al. (2013), and model while considering maximizing a reward function. The
Delahaye et al. (2017). reward function is measured as the negative of the difference
While ICM that is based on the theory of utility-maximi- between the estimated and observed system metrics. The
zation is widely adopted (Ben-Akiva and Lerman (1985)), airline network competition analysis (NCA) model is used
the theory has several basic assumptions that are evidently to calculate the estimated system metrics. It is a simula-
violated by the itinerary choice problem, which could sig- tion model that represents passenger-itinerary assignment.
nificantly impact the interpretation and accuracy of these It captures the demand–supply interactions at the network
models. First, it is assumed that travelers of an airport pair level, while considering the competition among all airlines.
make itinerary choice considering the same set of itinerar- This paper is organized as follows. “Overall framework”
ies (i.e., fixed choice set). In reality, the itinerary choice set section presents the overall framework. The application of
is dynamic and requires continuous update. If all seats of the framework considering the worldwide airline services is
a flight are sold out, all itineraries that include this sold- presented in “Model application” section. “Results” section
out flight should be eliminated from the choice set of all gives the model results as well as the results of several case
travelers making subsequent bookings. Thus, future travelers studies to illustrate its accuracy and interpretation. Finally,
will have a smaller set of itineraries to choose from, result- “Summary and conclusions” section provides concluding
ing in different itinerary choice sets across travelers of the comments and suggestions for research extensions.
same airport pair. Second, the utility-maximization choice
theory requires that options in the choice set of travelers
to be fully independent (i.e., mutually exclusive). However, Overall framework
this requirement could be violated as connecting itineraries
between most airport pairs in a typical airline network could This section describes the overall framework used for devel-
share one or more flights. Having a common fight among oping an ICM that addresses the limitations of adopting the
itineraries in a traveler’s choice set makes them not fully utility-maximization theory for modeling itinerary choice
independent. Third, most developed itinerary choice models behavior. The proposed ICM model overcomes these limi-
use number of passenger bookings of the itineraries as an tations by explicitly capturing the impact of the airlines’
indication of their level of attractiveness. However, number seat capacity constraints on the travelers’ itinerary choice
of bookings of any itinerary is censored by the available seat behavior. Figure 1 illustrates the different components of the
capacity making it a biased measure for its level of attrac- framework, which consists of a RL algorithm that integrates
tion (Zeni 2001; Vulcano et al. 2012; Nikseresht and Ziarati, a full-scale airline network competition analysis model. The
2017). For instance, an attractive itinerary with a limited seat RL search algorithm is developed to estimate the param-
capacity will always have less bookings. Thus, using the eters of the itinerary choice models. The algorithm seeks to
capacity-constrained number of bookings could be mislead- maximize a reward function, which represents the negative
ing for estimating the relative attractiveness of itineraries in of the deviation between the actual and estimated airline
a traveler’s choice set. Finally, a utility-maximization ICM network metrics. The purpose of the full-scale airline net-
only estimates choice probabilities at the itinerary level. The work competition analysis model is to estimate the network
statistical significance of the model is only considered at metrics at each iteration, which is passed to the calculations
this level. Predicting the demand at the flight level, which of the reward function that facilitates the learning mecha-
is needed for most studies, requires aggregating the choice nism in the RL algorithm. The following subsections explain
probabilities of all itineraries that include each flight. The these components, which include input data, the structure
accuracy of a flight’s predicted demand could be question- of itinerary choice models, the NCA model, and the param-
able considering that the flights’ seat capacities are ignored eter estimation of the ICM using the reinforced learning
while estimating the itinerary-level choice probabilities. algorithm.
Considering these violations and their possible adverse
impact on the fidelity of PFM, there are increasing calls in Input data
the airline network planning community to revisit method-
ologies adopted for developing ICM. This paper contrib- The framework requires four main input data elements for
utes to the literature by introducing a novel methodology all airlines competing in the markets under study. These data
to develop ICM that addresses the limitation associated elements include (1) the OD passenger demand matrix, (2)
with adopting the utility-maximization theory to model the the fare data, (3) the flight schedule data (i.e., timetable),
and (4) the available actual system metrics. The OD demand distance. Historical flight schedule data are available for
matrix data give the number of passengers traveling by all each month from several vendors (Sabre 2020). These ven-
airlines for each airport pair. These data are usually reported dors also provide the latest flight schedule data for near
monthly for the entire month and also for the average day future months as published by the airlines. The last data
of the month, which is known as passengers per day each element includes the actual system metrics. These data are
way (PPDEW). The second input data element is the fare used primarily to calculate the reward function of the RL
data which are widely available in the form of the average algorithm. The reward function is to benchmark the model
fare for each airport pair calculated based on all itinerar- estimation results against the ground truth in the successive
ies (e.g., non-stop and connecting) serving that airport pair. iterations of the RL-based calibration process. It should be
Both these data elements are available for each past month mentioned that the framework is flexible to adopt any type
of the year through several commercial vendors (e.g., Diio of metrics for benchmarking and calibration including disag-
by Cirium 2020; Sabre 2020), which compile and validate gregate, aggregate, or both.
the data from different data sources. For the purpose of this study and because they are
The flight schedule data give the list of scheduled flights widely available, two aggregate metrics are used in the cur-
(timetable) for all airlines. Each flight is reported in terms rent implementation of the framework, which are (a) the
of its main attributes including origin airport, destination airline leg/route data and (b) the market share data at the
airport, departure time, arrival time, fleet type, seat capacity, airline-stop level. These data are also available as part of
code-share information, day of week, and flight aeronautical the Sabre Data & Analytics Market Intelligence 6.3 platform
(Sabre 2020). The airline leg data include the average load form, the probability that individual i chooses itinerary k is
factor and the total PPDEW for each operating airline for expressed as given in (1).
each airport pair. The load factor is a measure of the aver-
(1)
( )
age flight occupancy measured as the percentage of sold Pik = S xik , xil≠k , Si , 𝛽 ∀i ∀k
capacity. For example, in August 2018, airport pair John F.
where Pik is the probability that traveler i choses itinerary
Kennedy International–London Heathrow (JFK-LHR) was
k , xik a vector of attributes of itinerary k in the choice set of
served by four operating airlines, which are British Airways
traveler i , xil≠k a vector of attributes of the other itineraries
(BA), Virgin Atlantic Airways (VS), American Airlines
( l other than k ) in the choice set of traveler i , Si a vector
(AA), and Delta Airlines (DL). These four airlines had an
describing the characteristics/preferences of traveler i , 𝛽 a
average monthly load factor of 82.1%, 78.0%, 74.1%, and
set of parameters that describes the effects of the variables
76.9%, respectively (Sabre 2020). The PPDEW for these
on probabilities, which are estimated statistically, and S(.) is
four airlines were 1812, 1268, 836, and 359, respectively.
a score (utility) function, which measures the attractiveness
The market share data at the airline-stop level include the
of traveler i to itinerary k.
PPDEW for each service. For example, in August 2018, for
airport pair Daytona Beach International–John F. Kennedy
International (DAB-JFK), there were 122 passengers flew Classification of airport pairs
using JetBlue Airways (B6), four passengers flew by Delta
Air Lines (DL) via Hartsfield–Jackson International airport The travelers’ itinerary choice behavior could significantly
(ATL), and one passenger flew by American Airlines (AA) vary across airport pairs (Coldren and Koppelman 2005).
via Douglas International Airport (CLT). The market shares For example, the travelers’ itinerary choice behavior in
of the three services were 95.7%, 3.4%, and 0.9%, respec- short-haul markets could be different from medium or long-
tively (Sabre 2020). Unfortunately, these data are not avail- haul markets. This behavior is also expected to be different
able by the time of day. Of course, if time-dependent version in airport pairs with only connecting itineraries, where the
of these data becomes available, it can be incorporated to connection duration and itinerary circuity might be the main
better explain the travelers’ itinerary choice behavior with attributes governing the travelers’ choices. As such, in this
respect to their preferences to departure/arrival times. study, the different airport pairs are classified into homo-
geneous classes and an itinerary choice classifier model is
Itinerary choice models developed for each class. Three main criteria are used to
classify airport pairs, including (1) the number of stops in
The itinerary choice problem can be generally described the best itinerary scheduled by any of the airlines in the air-
as a statistical classification problem, which determines to port pair, (2) the distance (aeronautical miles) between the
which of a set of itineraries a new traveler will belong (i.e., airport pair, and (3) the number of daily flight departures
choose). To categorize this new traveler choice observation, in the airport pair. It worth mentioning that future work is
a training dataset is used which contains historical observa- underway to investigate other/additional classification crite-
tions. The individual traveler observations are analyzed into rion such as region, time zone, and type of market (domestic/
a set of quantifiable properties, also known as explanatory international). The main goal is to generate homogenous
variables or attributes. These explanatory variables, hope- classes with minimum variation among the members of each
fully, can explain the membership of each traveler choice class. Figure 2 illustrates the airport pairs’ classification con-
observation to an itinerary. Many classification algorithms sidering these three criteria, resulting in thirteen different
adopt a linear function to combine these explanatory vari- classes of airport pairs. Accordingly, there will be thirteen
ables. The linear function assigns a score to itinerary in the ICM to be calibrated. As shown in the figure, airport pairs
choice set by combining the attribute vector of a traveler are first classified into non-stop, single-stop, and two-stops
observation with a vector of parameters (weight) using a dot or more. Airport pairs served by non-stop and single-stop are
product. As such, the score can be viewed as the utility asso- further classified into three classes based on the aeronauti-
ciated with traveler i choosing itinerary k . Algorithms with cal miles between the airport pair. These classes include
this basic setup are known as linear classifiers. Classification distances less than or equal 2000 miles, 2000–4000 miles,
algorithms vary in terms of the procedure for determining and greater than 4000 miles. Finally, based on the number of
(training) the optimal parameters (coefficients) and the way daily flight departures, airport pairs are classified into three
that the score is interpreted. For the purpose of this study, classes including greater than or equal 15 daily flights, 5–15
a probabilistic function is used to represent the relationship daily flights, and less than 5 daily flights. This classification
between the choice of an itinerary and its attributes and the is only considered for non-stop airport pairs. The reason for
attributes of the other competing itineraries. In its general not further classifying the one-stop and the two-stop classes
is to ensure that a representative sample of airport pairs exist of stay, day of week, etc. It would be ideal if information
in all classes. on these variables is available for all demand in all airport
pairs, and for all bookings. Typically, variables that describe
Explanatory variables the itinerary characteristics are widely available compared
to travelers and trip attributes. Accordingly, this study is
Several explanatory variables are suggested in the litera- focusing on using these variables in model specification.
ture to explain the travelers’ itinerary choice behavior in Table 1 gives the list of explanatory variables considered
markets with different characteristics (Coldren et al. 2003; for the models developed in this study. The table also gives
Coldren and Koppelman 2005). These variables could be the definition of each variable. As shown in the table, five
categorized into three main classes: itinerary characteristics, main variables are considered which are the itinerary level
traveler attributes, and trip attributes. Examples of variables of service, type of the itinerary, connection quality, time of
that describe the itinerary characteristics include trip travel day, and airline presence and brand. No variables are con-
time, circuity, connection quality, airline presence and brand, sidered for travelers or trip attributes mainly because such
and departure/arrival times. Variables related to traveler data are not available for the purpose of this study. Includ-
attributes include, for example, income, gender, and age. ing these variables would be a recommendation for further
Variables related to trip attributes include trip purpose (busi- research. The itinerary level of service is a dummy variable
ness or leisure), trip type (domestic or international), length that indicates the number of connections/stops per itinerary.
Level of service
Less than non-stop A dummy variable, which is equal to one if the itinerary has at least one stop in non-stop airport pairs, and zero
otherwise
Less than one-stop A dummy variable, which is equal to one if the itinerary has at least two stops in one-stop airport pairs, and zero
otherwise
Itinerary type
Code share A dummy variable, which is equal to one if the itinerary is a code share or interline itinerary, and zero otherwise
Operated by regional A dummy variable, which is equal to one if any flights of the itinerary is operated by a regional airline, and zero
otherwise
Connection quality
Distance ratio A ratio that measures the itinerary aeronautical distance compared to the aeronautical distance between the airport
pair of the itinerary
Connection duration Total connection duration, if any (in hours)
Time of day
0:00–4:59 AM A dummy variable, which is equal to one if the itinerary’s departure time is between 0:00–4:59 AM, and zero
otherwise
5:00–5:59 AM A dummy variable, which is equal to one if the itinerary’s departure time is between 5:00–5:59 AM, and zero
otherwise
6:00–6:59 AM A dummy variable, which is equal to one if the itinerary’s departure time is between 6:00–6:59 AM, and zero
otherwise
7:00–7:59 AM A dummy variable, which is equal to one if the itinerary’s departure time is between 7:00–7:59 AM, and zero
otherwise
8:00–8:59 AM A dummy variable, which is equal to one if the itinerary’s departure time is between 8:00–8:59 AM, and zero
otherwise
9:00–9:59 AM A dummy variable, which is equal to one if the itinerary’s departure time is between 9:00–9:59 AM, and zero
otherwise
10:00–10:59 AM A dummy variable, which is equal to one if the itinerary’s departure time is between 10:00–10:59 AM, and zero
otherwise
11:00–11:59 AM A dummy variable, which is equal to one if the itinerary’s departure time is between 11:00–11:59 AM, and zero
otherwise
12:00–12:59 PM A dummy variable, which is equal to one if the itinerary’s departure time is between 12:00–12:59 PM, and zero
otherwise
1:00–1:59 PM A dummy variable, which is equal to one if the itinerary’s departure time is between 1:00–1:59 PM, and zero
otherwise
2:00–2:59 PM A dummy variable, which is equal to one if the itinerary’s departure time is between 2:00–2:59 PM, and zero
otherwise
3:00–3:59 PM A dummy variable, which is equal to one if the itinerary’s departure time is between 3:00–3:59 PM, and zero
otherwise
4:00–4:59 PM A dummy variable, which is equal to one if the itinerary’s departure time is between 4:00–4:59 PM, and zero
otherwise
5:00–5:59 PM A dummy variable, which is equal to one if the itinerary’s departure time is between 5:00–5:59 PM, and zero
otherwise
6:00–6:59 PM A dummy variable, which is equal to one if the itinerary’s departure time is between 6:00–6:59 PM, and zero
otherwise
7:00–7:59 PM A dummy variable, which is equal to one if the itinerary’s departure time is between 7:00–7:59 PM, and zero
otherwise
8:00–8:59 PM A dummy variable, which is equal to one if the itinerary’s departure time is between 8:00–8:59 PM, and zero
otherwise
9:00–9:59 PM A dummy variable, which is equal to one if the itinerary’s departure time is between 9:00–9:59 PM, and zero
otherwise
10:00–11:59 PM A dummy variable, which is equal to one if the itinerary’s departure time is between 10:00–11:59 PM, and zero
otherwise
Airline presence and brand
Fare ratio A ratio that measures the itinerary fare compared to the average fare in the airport pair
Table 1 (continued)
Variable Definition
Airline presence A ratio that measures the presence of the marketing airline of the itinerary in the airport pair. It is calculated by
dividing the non-stop seats of the airline by the total non-stop seats in the airport pair
Airline name A dummy variable, which is equal to one if the itinerary is marketed by this airline, and zero otherwise
For Non-Stop airport pairs, a dummy variable is considered her choice does not necessarily reflect his/her true prefer-
to indicate connecting itineraries. The Less than Non-Stop ence because this traveler is captive to choose only from the
variable is an indicator to represent an itinerary that has at available intervals. For example, assume there is an airport
least one connection. For the Single-Stop airport pairs, a pair that has only two departures in the morning (e.g., 7:00
variable is considered to indicate itineraries with two stops AM and 10:00 AM). In this case, all travelers are selecting
or more. The Less than One-Stop variable is when the itiner- either the 7:00 AM or 10:00 AM departure. When these data
ary has at least two connections. are used to calibrate a model, the parameters of the departure
The itinerary type is represented by two dummy vari- time intervals in the afternoon, for example, will indicate
ables. The first variable is the Code Share dummy variable, that the afternoon departures are not preferred. This is not
which represents the case when the itinerary is a code share correct. The reality is that these afternoon intervals are not
(or interline). The second variable is Operated by Regional, available and not represented in the choice set.
which represents itineraries that have at least one flight oper- Also, in the short-haul market, the departure and arrival
ated by a regional/affiliated airline. Regional airlines are times are correlated. For example, an itinerary that has a
airlines that operate regional aircraft to provide air service morning departure will most likely have a morning arrival
to small communities. They are usually contracting with a due to the short trip time. In other airport pairs in other
major airline, operating under their brand name, to deliver clusters (i.e., medium- and long-haul clusters), time zones
passengers to the airline’s hub from surrounding small might make the departure time variable to have a biased
airports. impact. For example, in the United States and for a cluster
The Connection Quality is measured by the Distance that includes the transcontinental routes, a midnight depar-
Ratio and the Connection Duration. The Distance Ratio ture from the west coast to the east coast (red eye flight)
measures the itinerary aeronautical distance compared to will arrive in the morning of the second day. However, a
the direct aeronautical distance between the airport pair of midnight departure from the east coast to the west coast will
the itinerary. The Connection Duration is the total connec- arrive in the middle of the night. This example shows that
tion duration in hours (i.e., ground time between flights) of the midnight departure interval has different impact depend-
the itinerary. ing on the location of the origin and destination and their
The Time of Day variable represents the departure time zones. Accordingly, the Time of Day variable is not
hour of the itinerary. This variable is included to control included in models of medium- and long-haul airport pairs.
for travelers’ assignment to the different departure times. Many of these airport pairs are expected to have significant
This variable is only included in Model 1 (i.e., airport pairs time-zone differences and in this case the departure time
that are short haul and have at least 15 daily departures). It preference will not be meaningful in the model. To correct
is not included when the number departures are less than for this factor, an additional clustering of the airport pairs
15 because in this case some departure times are not rep- based on their location (time zone) is needed. While this
resented. The departure time variables are exceptionally addition is expected to improve the model accuracy, it is
important, but they must be modeled correctly. They are expected to increase the computational burden of the model
included in the short-haul market with at least 15 daily because of the new parameters. We recommend considering
departures because this is the only case that they can be other specifications with additional clusters and variables in
correctly modeled. In this case, when there are at least 15 future research.
departures per day, it is expected that most departure time Finally, the Airline Presence and Brand are represented
intervals are represented in the choice set and travelers are in terms of three variables, which are the Fare Ratio, Airline
selecting among all options. In other words, the choice Presence, and Airline Name. The Fare Ratio measures the
reflects the actual preference of the traveler. On the con- itinerary fare compared to the average fare in the airport pair.
trary, for the other airport pairs in other clusters, which have For instance, low-cost airlines are expected to have low Fare
a few departure times, not all the departure time intervals Ratio compared to full-service airlines. The Airline Pres-
throughout the day will be represented in the choice set. ence measures the presence of the marketing airline of the
This implies that when a traveler chooses an itinerary, his/ itinerary in the airport pair. It is calculated by dividing the
“Kenelm,” said Lady Alden, raising her earnest eyes and clasping
her hands, “you cannot be cruel; you cannot forget every tie that
binds you to Ronald. Oh, why, my God! why did I bring that fatal box
here? You cannot forget all that Ronald was to you. He left his wife
and children in your charge, and you would rob them of all their
natural protectors—of husband and father! Oh, Kenelm! you must
not do it. No man could live and be so cruel.”
“I must do justice,” he said, firmly.
“What do you call justice?” she cried, wringing her hands.
“I shall deliver the man who did the deed to the laws of his
country, and they shall punish him for the murder done.”
Her face grew ghastly pale as she listened. It was terrible to gaze
upon, awful to see. Great drops of agony gathered on the white
brow. He turned his face away lest he should see the torture he was
bound to inflict.
She knelt at his feet, and raised her hands as though she were
supplicating the mightiest power.
“Kenelm, have pity on me if you will not on him! Have mercy on
me; if you injure him you kill me. You can only reach him through my
heart. See, dear,” she continued, with a low sob, “if you stood here,
and you took deadly aim at him I would fling myself before him and
die first. You should walk over me dead before you touched him! All
my life is bound up in his. I live in him. My soul is one with his. Oh,
Kenelm! for God’s dear sake, have pity and spare!”
But he never even turned his face toward her.
“I love him so dearly,” she moaned—“oh, so dearly, Kenelm!
Have pity on me. I have never wronged you. I have been a true,
good friend to you all my life. I have sorrowed for your sorrows.
Spare me now!”
“I would not injure you, Hermione,” he said, in a low, hoarse
voice. “I——”
But she interrupted him.
“You would not injure me, yet you would take him who is the life
of my life from me. Oh! Kenelm, see, I would not raise that finger to
save my own life, but I kneel to plead for his, and I shall kneel until
you grant my prayer!”
“It is useless. The most solemn oath that a man could take I have
taken; it was on the dead, white lips of my murdered love, and I
cannot break it!”
Her lips grew parched and dry with the terrible agony that
possessed her. Her voice was weak and faint as though she were
exhausted by long and wearisome pain.
“You must spare him, Kenelm. See, I pray you with tears—I pray
to you as woman never prayed before. For God’s dear sake, let him
go free!
“You are not listening!” she cried; “you are turning from me—you
who hold what is dearer than life in your hand. I will give you—— Oh!
my God! what can I give you? How will I bribe you? Would that my
lips were touched with fire! Would that my heart lay before you that
you might see its love and its despair!”
“Justice!” he said, slowly; “we must have justice. Remember, it is
an attribute of the most high God, just as mercy is. Remember who
said, ‘Blood for blood.’ Remember my oath.”
She fell forward then with her face on the ground, and such
passionate prayers went from her white lips he could with difficulty
withstand them.
“You will never be happy again if you do this ruthless deed! If
she, poor Clarice, could speak, she would plead for him! Oh, spare
him, Kenelm; spare him!”
She seized his hand, and the tears from her weeping eyes fell on
it.
“You will be kind to me. You are chivalrous and kind. You will not
let a woman kneel here at your feet, and refuse her prayer?”
“It is to avenge a woman’s cruel death that I act,” he said,
gloomily.
“‘Vengeance is mine,’ said the Lord; ‘and I will repay it.’”
“Hermione, this is not revenge, but justice. You know it; I know it.
If I could save your husband by laying down my own life, I would
gladly do it, but I cannot.”
“You will not hear my prayer, then?”
“I cannot. You should not kneel to me in vain, Hermione, if I
could.”
He turned away, leaving her kneeling there—white and cold, and
as one half dead—the blood in her veins frozen with fear. He walked
to the window. The golden sunlight still lay on flowers and trees, a
little bird was singing its sweet, melodious song. It seemed to him
that years had passed since he stood there before, and the crimson
shade of murder had come between him and the bright sunshine.
He stood still, his whole heart and soul given over to a mighty
tempest. He knew the secret at last—after years of patient waiting,
after spending a fortune in searching for the criminal—he was living
here, at his own doors—he was the man he had called brother and
friend.
He bit his lip to keep down the anger that was fast rising in his
heart. No pitying thoughts came to him of the man who had been his
friend. Hot, bitter, long-pent-up anger raged in his heart.
“For that which he has done he deserves to die,” he said, “and
die he shall.”
He was startled by the touch of a soft hand, and turning he saw a
sight that might have melted a heart less angry than his. Lady
Hermione had stolen gently from the room in search of her children.
She had brought them in with her, and they were kneeling there at
his feet, and she like a sheltering angel behind.
“Harry, baby Maude, pray to him, clasp your hands, my babies—
look in his face, and ask him to spare papa. Listen, Maudie: ‘Spare
papa!’”
The lovely baby face was raised to his, the pretty lips lisped the
words, “Spare papa!” and Harry, with great tears shining in his blue
eyes, said, “Spare papa!”
“My God,” cried the unhappy mother, who saw no signs of
relenting on that stern face, “soften his heart; take Thou pity on us,
since he will have none.”
“I have all pity, Hermione,” he said, “all true and tender pity for
you and yours, but justice must be done!”
Then she stood up before him, and raised the little ones.
“Look your last, my children,” she said, “on the face of the man
who is a traitor to your father’s trust—who can look at you and take
that father from you. You shall see him no more.”
He steeled his heart against them. In vain little Harry went back
and clasped his arms round him.
“You will not make my mamma cry—my own beautiful mamma?
Do what she wants you to do, Mr. Eyrle. I know she is not wrong.
You love us—you would never hurt papa.”
He had to recall the dead white face of his murdered love before
he could resist that prayer; then he kissed the child and led him
sadly to his mother. Lady Alden took them sadly from the room.
When she returned there was a look of determination on her gentle,
lovely face. She went up to him.
“Once more, Kenelm,” she said, “I ask you, for God’s sake, will
you give up your scheme of vengeance? It is years now since the
deed was done; it must have brought its own punishment. Will you
not let it die—pass into oblivion?”
“I will not,” he replied, sternly, “I will keep my oath.”
“Tell me, can no prayers, no pleading, move you from this
purpose?”
“None; that which I have sworn I will do at any cost.”
“Give me one hour’s grace, then come to me again. I have
something to tell you in one hour. See, I will not leave this spot
where I am standing—only one hour.”
“I will obey you,” he said, and, without looking at her face, he
quitted the room. What passed there only God and herself knew.
Two hours passed before he returned, and he found her where he
had left her, the sweet face white and exhausted, but with a look of
resignation upon it.
“You have returned,” she said, “and I ought to tremble, for in you I
see the messenger of doom and death. Kenelm, I have something to
tell you!”
He looked from the trembling hands to the pale face.
“I am ready to hear,” he replied. “Do not waste time in making
excuses for your husband, Hermione; it is labor in vain.”
A strange, wan smile came to her lips.
“I have nothing to excuse in him,” she said, gently, “for that which
I have to tell you does not touch him. When you hear it you may fling
me down and trample the life from me if you will.”
“Talk reasonably, Hermione; then I may understand you.”
She went to the table where the box containing that terrible
evidence lay. She opened it and took from it a long, slender dagger,
with its rusty stain.
“You are right,” she said, in a low, dreary voice. “That was the
instrument with which the deed was done, and this belonged to me.”
“To you!” he cried. “What do you mean, Hermione?”
“It was given to me years ago by my cousin, who had been
traveling in Greece. God grant that I may not go mad. He gave it to
me one summer evening like this. My mother said it was a foolish
present. My father bade me lock it away, but my cousin told me it
was a great curiosity; that in ancient times the Grecian ladies wore
such deadly toys fastened to their girdles.”
“Why do you tell me this now?” he asked.
She bent down and whispered something to him that made his
face grow pale with horror, while he sprang from her as though the
air she breathed and the words she spoke were poison.
CHAPTER XLVIII.
HOW WILL IT END?
“You are mad, Hermione,” cried Kenelm Eyrle. “You cannot mean
it; it is not true.”
“It is true,” she replied; “that dagger was mine, and I—hear me,
Kenelm Eyrle—I confess it, I did the deed. I, and I alone, am guilty!”
“My God!” he cried, “it is surely impossible. Those hands of yours
are surely not stained with crimes so abhorrent.”
“I am guilty,” she said, “and I alone. Do your worst to me now.”
“I refuse to believe you. I cannot credit it. You to do such a deed.”
“Yes,” she replied, and there was no hesitation, nor fear in her
voice, “I did it, Kenelm; I am guilty.”
He stood in silence, his emotion too great for words. This gentle,
gracious, lovely lady a murderess! Ronald innocent after all? She,
whom he had looked upon as pure and peerless, guilty of this
monstrous crime?
“I cannot believe it,” he repeated.
“Yes; it is true, I would not have told you had you promised to let
the dread memory of it die. I would not have mentioned it if you had
promised to—to spare my husband. Guilty as I am, I dare not double
my guilt by letting him die for what was my crime.”
He was still looking at her, as though he were in a dream.
“Do with me what you will,” she said. “I prayed, I pleaded for
Ronald’s life; I do not even ask for mine.”
“Why not?”