Providing personalized o ers, and services in general, for the users of a system requires perceiving the context in which the users' preferences are rooted. In this work, we introduce the use of an already known model and methodology { based on the so-called Context Dimension Tree { along with a conceptual architecture to build a recommender system that o ers personalized services for travelers. The research is performed in the frame of the Shift2Rail initiative as part of the Innovation Programme 4 of EU Horizon 2020.
The demand for systems that provide personalized services increases the need to
extract knowledge from di erent sources and appropriately reshape it. Besides,
services cannot be properly adapted just by considering the static information
obtained from the users' pro les: using instead a combination of such pro les
with the context in which the user is going to be served is de nitely more realistic.
Generally speaking, Context can be recognized as a set of features (i.e. values
for variables) contributing to the decisions of a user in a system [
This work explores and presents the essential elements vital to design a usercentered Recommender System (RS) for the Travel Companion (TC) module, currently being developed within the Shift2Rail (S2R) initiative as part of the Innovation Programme 4 (IP4). TC acts as an interface between users (typically travelers) and the other modules of the S2R IP4 ecosystem, supporting the
Copyright c 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). This volume is published and copyrighted by its editors. SEBD 2020, June 21-24, 2020, Villasimius, Italy.
This work was supported by Shift2Rail and the EU Horizon 2020 research and
innovation programme under grant agreement No: 881825 (RIDE2RAIL)
users in all steps of their travel. Precisely, since supporting context-dependent
data and service tailoring is paramount to ensure personalized services, we aim
at extending the work carried out in [
Recommender Systems: RS are designed to recommend items to users considering
their needs. The user pro ling approaches, that emerged in the literature with
the aim to determine the users' requirements and behavioral patterns [
This section explores the travelers' contexts and preferences through the CDT methodology, and introduces a conceptual system architecture that includes the main components for the ranking of the trips according to the travelers' context.
To enable context-aware recommendations for travel purposes, we identi ed the aspects characterizing contexts which correspond to the TC users' choice criteria that are potentially useful to score the available trips (see Figure 1).
Note that, in the application design phase, designing a CDT is performed
independently of, yet in parallel with, the other routine activities involved in this
phase [
Practically, the CDT design is done iteratively, and depends on the nal requirements of the application. The TCDT we propose de nes the fundamental dimension and concept nodes in such a way that any further interesting features can be added by increasing or decreasing the level of granularity of the model. 3.1
In the TCDT of Figure 1, the Personal Data ND captures the socio-economic characteristics of the users through the Profile NC . Di erent groups can be extracted according to the values of the socio-economic factor concept node, such as geographical origin, profession, and so on. Each group carries a set of preferences that are rather stable, thus can be associated with the notion of user pro le. The main motivations for this dimension are to enable a warm start for the system and also to provide the chance of detecting and possibly supporting some group behaviors. As an example, consider two regions X and Y in the category of geographical groups, and suppose that, for some reason, users from X tend to choose eco-friendly travels much more frequently than users from Y. Investigating the reasons behind this tendency enables the authorities and the system to take the required actions (if applicable) for increasing the popularity of eco-friendly travels for the users living in region Y.
Moreover, a person can be member of some communities. The Memberships NC captures the memberships of the user along with their payment methods. The Loyalty Cards ND captures membership of the user in a community that potentially provides speci c discounts. Moreover, we introduced ND|Other Memberships|as a placeholder to capture other, less structured, communities that may follow di erent patterns compared to those based on Loyalty Cards. Tailoring the Memberships NC with more levels of granularity through a combination of domain experts' knowledge and machine learning approaches like clustering is one of our future works.
We put a particular emphasis on supporting the needs of people with disabilities and health-related issues (HI), dedicating the Health Issues ND to this purpose. The HI Category ND distinguishes the case that the issue belongs to the PRM category, which stands for Person with Reduced Mobility. The descendants of HI Type ND list some of the most critical issues. Moreover the TCDT
A. Javadian et al. mutually exclusive concepts. increases the level of granularity for the Walking NC through the Aid Tool ND to exemplify the concepts which should be considered during the future expansion. Knowing this concept is essential because of the particular space each of the Aid Tools require while recommending trips. The same importance is also applied to the Severity ND, which can potentially limit the travel choices. Lastly, HI Ends in ND, enables the TC to determine if the issue is permanent or temporary.
Among the other top dimensions, Behavioral Status captures the current situation of the user as follows: the Traveling concept captures the state in which the user is traveling, or has purchased a travel o er and is waiting for the upcoming trip. Obviously, its two sibling NC , drawn with dashed lines, are mutually exclusive: the Inactive NC is true if the user is not interacting with the TC, while the Surfing NC incorporates both implicit and explicit momentary user behaviors while interacting with the TC. More precisely, the Interface and Gestures ND capture implicit behaviours: while users are interacting with the TC through a Computer, since there is extra visual space, and presumably the user has no urgent travel request, the TC proposes information regarding \eco-friendly" o ers, which have lower CO2 emissions. The users may decide to click, scroll, or ignore this information, which in turn can provide useful insights about their preferences regarding eco-friendly o ers.
Explicit behaviours, instead, are captured via the Request ND. Eco-friendly traveling behaviors can be promoted through so-called ride-sharing. For this reason, we foresee that when the user requests a travel o er through the TC and driving a car is a possibility, they can specify whether their Role is that of Driver or of Passenger, as well as, their Purpose and preferred Segments.
The user provides the locations that they are going to visit (at least source and destination). In the TCDT this value is transformed into appropriate concepts such as Country, City and Zone. For example, the Zone ND enables the TCDT to capture the factors contributing to the user's decision through its children nodes i.e. Distance from Public Transportation services, Hotels and Landmarks (PT). Moreover, the Weather Forecast NC is used to capture weather information according to the Time when the user will be in that Zone. The same strategy is applied to transform the actual value of the requested departure and arrival times to the Time NC and its descendant nodes.
It may happen that the user has some Accompanying Items (e.g., a bike) and Pets, whose characteristics|such as their Type, Species, Size and Number |should be considered when recommending trips. Also, accompanying Persons not only a ect travel choices from the logistic aspect but, if the Person is also a user of the TC, their preferences should be considered for recommending trips.
The Services ND encompasses the variety of Travel- and Meal-related preferences according to the optional user-provided scores. Precisely, the TC, through scores, enables its users to provide scores between 0 and 5 representing excluded and requested respectively for the available services. Beside the preferences obtained by analyzing the history of the user's choices, the services scored as 0 allow the system to lter out such travel o ers. 3.2
Figure 2 shows the conceptual architecture envisaging three main actors, namely End Users, Travel Companion, and Third Parties, along with their main elements to learn the users' preferences and recommending the best travel options accordingly. It provides an overall view, hiding the details of all the TC's blocks and functions; also, it does not specify the modules' physical locations (cloud, etc.).
Vitally, since the system deals with sensitive data, best practices concerning
issues related to security and individual privacy through technical solutions [
Experts Weather Service Organizations Travel Shoppers Travel Sevice
Providers Social Media
Third Parties
End Users
Travel Companion
User Interface
Recommender Core
Preferences Learner Evaluation Metrics Community Personal
Learner Learner
Data User
Centered Centered Ranker
Filter
Data
User-Data Knowledge Models
Service-Related Weather Forecats
Travel Offers Services & Products
Features Interoperability Framework Trip Tracker
Business Analytics Dashboard
Social Media Core
SM Miner Pipeline SM Publisher
Naturally, the Travel Companion is, for our purposes, the most important
actor, for which we provide a further breakdown into modules. The Knowledge
Models block is useful to have a warm start for the newly-registered users, and for
whose behavioral activities and speci c preferences the system does not have any
prior data. It also provides the opportunity to investigate the behavioral drifts
that happen for the person, groups, and communities, and to acquire initial
information for logistic purposes. The Social Media (SM) Core is composed of the
two main blocks SM Miner Pipeline and SM Publisher. The SM Miner Pipeline
employs Natural Language Processing techniques to harvest knowledge from
SM platforms, seeking explicit travel-related mentions and keywords. The SM
Publisher enables the TC to publish tailored news, promotions, and responses
to speci c users on SM, and allows users to share their trip information and
other socializing functionalities. Recommender Core elements take as input user
data, knowledge models, and service-related information and accordingly
provide a ranked list of the trips for the user. The TC uses the S2R Interoperability
Framework of the Sprint Project [
The main Third parties, playing di erent roles about the provision of information and services to the TC, are the following: (i) Data Providers serve various information about the user, service, etc., e.g. data about the safety of zones; (ii) Experts from di erent domains like sociology, transportation, etc. provide and modify the knowledge models of the TC; (iii) Weather Service Organizations provide weather forecasts associated with the places; (iv) Travel Shoppers (TS) are the organizations and services that arrange the journeys; (v) SM can play two primary roles: they can collect data regarding the users' attitudes and preferences, and the TC can use them to publish tailored news and promotions. 3.3
The travel o ers received by the TC from TS contain variables describing their characteristics (e.g. duration, CO2 emissions, etc.). As a rst step in the recommendation of travel o ers to the user, the Filter block (see Figure 2) hides some of them according to the knowledge provided by the values associated with speci c TCDT dimensions that are stronger preferences and act as a kind of personal constraints|e.g. o ers that include TSPs with Services which users already excluded by scoring them 0. Then, the Ranker receives the remaining travel offers, plus a vector of preferences containing the weights capturing the importance of the TCDT values to the user and to divers communities and groups. For each received travel o er, the Ranker computes a numerical score in the interval [0; 1] according to suitable evaluation metrics and uses this score to rank o ers.
For example, consider a pregnant traveler, traveling with her partner for leisure; for the trip, she has excluded a speci c type of meal by giving 0 score. Besides, through knowledge models, TC knows that leisure travels, have higher score for economy products (assuming that neither she has provided the optional scores for any of the products nor the TC has any history regarding her preference in leisure travels). Among the travel o ers provided by the TS, TC lters out those that include the excluded meal. Considering her current context, the weight of her pregnancy condition potentially might be higher than that of her possible preference for economical o ers, thus a travel o er that includes a direct ight will take over indirect ights (even though more economical), and, due to the accompanying partner, two seats next to each other might be more scored than seats with more space but in di erent places (even though more comfort). 4
The design of an advanced learning system for travelers' preferences should not only provide the best possible rankings (and, consequently, suggestions) for travel o ers, but should also adapt to changes in the behaviors and preferences of users.
The latter requirement is of great importance because preferences are highly dynamic and prone to changes according to di erent contexts.
In this work, we proposed a methodology to describe, at the conceptual level, the di erent contexts in which travelers can nd themselves, with the advantage of being able to specify, for each traveler, how their preferences are a ected by context changes. The methodology consists, on one side, in representing the characteristics of users, services and speci c circumstances by means of a TCDT, and, on the other side, in designing a system architecture that identi es the potential sources of data and the interactions among the various system elements.
We are currently working on enriching the proposed TCDT by increasing the dimensions' level of granularity to explore the other contexts whose characteristics can contribute to the users' preferences and to their traveling decisions.