A Hybrid Knowledge-based Recommender for Mobility-as-a-Service

Konstantina Arnaoutaki, Babis Magoutas, Efthimios Bothos and Gregoris Mentzas

ICCS, National Technical University of Athens, Athens, Greece

Keywords: Recommender Systems, Mobility-as-a-Service, Knowledge-based, CSP.

Abstract: Mobility as a Service (MaaS) is the integration of various forms of transport services into a single “mobility

plan”, that can be considered as a bundled set of distinct services/products, bought and used as a single product.

The concept of MaaS is gaining an increasingly high interest however, there are still many challenges that

have to be dealt with when designing and offering viable MaaS products, including the suggestion of the

optimal MaaS plan that matches a user’s personal needs. In this paper, we propose a knowledge based

recommender system that builds upon constraint programming mechanisms and provides the necessary

functions to capture user preferences, exclude MaaS plans which do not match those preferences and infer the

similarity of the remaining plans to the user’s profile. The final outcome is a filtered and ranked list of MaaS

plans which allows the user to select the one that better matches her/his preferences.

1 INTRODUCTION

Mobility as a Service (MaaS) is a new mobility

paradigm that aims to provide integrated and

seamless access to transport services through one

single digital platform. The key concept behind MaaS

is to place the user at the core of transport services by

offering tailor made mobility solutions according to

users’ individual needs. In this respect, MaaS users

receive customised door-to-door transport services as

well as personalised trip planning and integrated

payment options (Durand et al., 2018).

MaaS is offered by a new type of mobility

operators the “MaaS Operators”. These are

intermediary companies that make agreements with

public and private transport operators on a city,

intercity or national level and offer subscriptions to

bundles of transport services, termed as “MaaS plans”

or mobility products (Kamargianni and Matyas,

2017). Access to the transport services in achieved

through mobility apps and related back-end platforms

that are maintained by MaaS operators and integrate

all the available transport services while providing a

single point for MaaS plans selection, route planning

and payment.

In a MaaS environment there can be a multitude

of MaaS plans with varying characteristics, in order

to meet the specific needs of different types of

travellers. These plans are derived from combinations

of available transport services. For example, MaaS

plans can combine and include public transport, taxi,

car sharing, bike sharing, car rental and/or other

related services such as parking or e-vehicle charging

stations.

It is evident that the selection space of MaaS plans

for end users increases according to the available

transport services, the combinations of which can

generate large choice sets with complex structures.

Moreover, despite the fact that travellers make use of

individual mobility services and are familiar with

them, they are not that familiar with the MaaS

concept where mobility services are bundled.

Consequently, finding a MaaS plan that is aligned to

the individual traveller’s needs and preferences

quickly and accurately is a cognitive task that

travellers will not be able to manage easily.

In this paper, we describe a hybrid knowledge-

based recommender system that supports travellers’

decisions related to the selection of MaaS plans, out

of a plethora of available plans that match their

preferences and needs. The recommender provides

the necessary functions to capture user preferences,

exclude plans that do not match those preferences and

infer the similarity of the remaining plans to the user’s

preferences. The final outcome is a filtered and

ranked list of MaaS plans. Users are presented with a

short list of plans that better match their preferences

and select the one they want to use. The proposed

approach is hybrid in the sense that it combines two

techniques. Firstly it incorporates constraint

Arnaoutaki, K., Magoutas, B., Bothos, E. and Mentzas, G.

A Hybrid Knowledge-based Recommender for Mobility-as-a-Service.

DOI: 10.5220/0007921400950103

In Proceedings of the 16th International Joint Conference on e-Business and Telecommunications (ICETE 2019), pages 95-103

ISBN: 978-989-758-378-0

modelling, where the MaaS package selection

problem is represented using constraint programming

formalisms. in order to infer a subset of potential

MaaS plans from a wider set of available plans.

Secondly, a weighted similarity calculation function

ranks the remaining MaaS plans, based on their

similarity to user preferences. A main advantage of

our approach is its ability to adapt to the needs of

different MaaS settings by integrating the knowledge

and requirements of domain experts as rules of a

constraint satisfaction problem.

The remainder of this paper is organized as

follows. In Section 2 we discuss the related work. In

Section 3 we present our hybrid knowledge based

recommender, while in Section 4 we elaborate on the

implementation details. Section 5 presents an

indicative usage scenario of the proposed approach

and finally, Section 6 concludes the paper and

provides directions for future work.

2 BACKGROUND AND RELATED

WORK

2.1 Background

The problem of suggesting personalized MaaS plans

resembles that of generating bundle

recommendations which has been mainly addressed

by data-driven approaches that rely on the analysis of

past user choices (see Section 2.2 for an overview of

the related work). However, a data-driven approach

in our case would require significant amounts of

historical data concerning user’s past selections of

MaaS plans, which are not available in any newly

deployed MaaS solution.

Knowledge-based recommender systems (RS)

help to tackle the absence of data and user feedback,

i.e. the so-called cold-start challenge, by combining

explicit requirements, stated by the users within a

recommendation session, and deep knowledge about

the underlying domain for the computation of

recommendations (Felfernig et al., 2015).

Our approach relies on the use of constraint

programming theory embedded in knowledge-based

recommenders, which fits well to the problem of

identifying and recommending personalized MaaS

plans. More specifically, we consider a Constraint

Satisfaction Problem (CSP) that involves finding a

value for each one of a set of problem variables where

constraints specify that some subsets of values cannot

be used together (Freuder and Mackworth, 2006).

Following this idea, we considered the task of “MaaS

plan selection”, where each transport service included

in a MaaS Plan can be represented as an option in a

constraint satisfaction problem. Under the CSP

principles, two discrete phases of the problem solving

process are defined: i) the problem is modelled as a

set of decision and parameter variables, and ii) a set

of constraints are applied on these variables which

must satisfy a solution. Decision variables represent

the available choices and their potential values

coincide with the available decision options. In our

case, decision variables are derived from the

characteristics of the mobility services which are part

of the MaaS plans (such as the available quota of

public transport, bike sharing or taxi). The second

phase of the process refers to applying a set of

constaints in order to find solutions to the problem, so

that the values of the decision variables satisfy all the

applied constraints. In our case, by applying the

constraints, we filter out MaaS plans that do not

satisfy the defined constraints.

2.2 Related Work

Generating recommendations and providing

personalized suggestions for bundles of products is a

problem that has been investigated in domains, such

as tourism, telecommunications and e-commerce. An

analysis of the types of recommender systems (RS)

that can be used for dynamic bundles

recommendation of touristic services (e.g. activities,

places to stay) is provided by Schumacher and Rey

(2011). Zhang et al., (2013) present a hybrid

recommendation approach which combines user-

based and item-based collaborative filtering

techniques with fuzzy set techniques and knowledge-

based methods (business rules) and apply it for

telecom products and services recommendations.

Beheshtian-Ardakani et al., (2018) approach the

problem of suggesting product bundles in e-

commerce websites from a marketing perspective.

They propose a novel model for bundles

recommendations by using market segmentation

variables and customer loyalty analysis. Customer

loyalty is calculated by employing the so-called

recency, frequency, and monetary value (RFM)

model that considers the recency of the last purchase,

the frequency of purchases, and their monetary value

(Linoff and Berry, 2011).

Constraint-based recommender systems have

been successfully applied in various domains.

Felfernig et al. (2006) present CWAdvisor, a domain-

independent knowledge-based recommender that

assists customers in the product selection process via

a personalized conversation. The aforementioned

ICE-B 2019 - 16th International Conference on e-Business

recommender has been successfully applied to

support decisions for selecting financial services and

electrical equipment. Jannach et al. (2009) introduce

a virtual advisor for tourists called “VIBE”. The

proposed advisor uses knowledge-based

conversational RS technology to provide a

personalized way of choosing plans offered by a spa

resort from a predefined catalogue, that meet users’

individual requirements. Reiterer et al. (2015)

describe a constraint-based recommender that

supports households to select the optimal waste

disposal strategy that corresponds to their needs.

while Murphy et al. (2015) design a constraint-based

energy saving recommender system. The proposed

system exploits real-world energy use data of

appliances, and suggests behaviour changes and

optimized appliance usage schedules so that users can

reach domestic energy saving goals. Zanker et al.

(2010) have approached the composite task of

configuring product bundles, namely travel packages

combining accommodation and activities services,

within the constraint-based framework. Their work

concluded in a generic Web configurator, that

combines recommendation functionality together

with constraint solver principles and results in a range

of personalized product bundles, tailored to tourists

needs while respecting e-tourism domain restrictions.

Their work can be considered as a hybrid paradigm

strategy that mixes knowledge-based techniques with

collaborative filtering recommendation methods.

3 OUR APPROACH

The proposed recommender system for MaaS plans,

relies on state-of-the-art techniques and follows a

novel hybrid knowledge based approach that i)

encodes the MaaS plans filtering problem as a

constraint satisfaction problem (CSP) by leveraging

knowledge from domain experts, and ii) uses explicit

feedback from users to derive a personalized ranked

list of MaaS plans through a similarity function. The

proposed approach addresses the cold start problem

(Lam et.al, 2008), and can be used to derive

recommendations even for newly registered users, for

who the system does not have any information

regarding their past preferences on the available

items.

Figure 1 provides an overview of the proposed

approach. Since different combinations of offerings

and MaaS plans may be available in a city depending

on the available transport services and business

environment, we have designed a MaaS plan

configurator tool that allows MaaS operators to

define and configure the MaaS plans to be offered.

Our knowledge-based approach exploits a

recommender knowledge base that contains explicit

rules (MaaS constraints) about how to relate user

requirements (customer variables) with MaaS

product features (product variables). Such rules are

defined by knowledge engineers with knowledge of

the field, while user requirements are acquired

through questions incorporated into a graphical

knowledge acquisition user interface (see Figure 4 for

an indicative example).

Figure 1: Overview of our knowledge-based CSP and

similarity-based MaaS plans recommender.

The MaaS product selection problem is

formulated as a Constraint Satisfaction Problem

(CSP), with the goal of limiting the size of the space

that must be searched in order to identify the plans to

present to the user among those available. The CSP is

integrated into a recommender engine, where the

solution objective is to derive a list of preferred MaaS

plans by filtering out plans not satisfying the

constraints. The list of remaining MaaS plans is

further processed using a weighted similarity function

which sorts the results in a ranked list of plans which

are aligned with user preferences. In the case that no

matching product is found as a solution to the CSP,

the similarity-based approach is applied to all

available products to rank them based on their

similarity to the user profile, under user-provided

budget constraints.

3.1 MaaS Knowledge Base and

CSP-based MaaS Plans Filtering

The knowledge base of a constraint-based

recommender system can be described through two

A Hybrid Knowledge-based Recommender for Mobility-as-a-Service

sets of variables (V

, V

PROD

) and two different sets of

constraints (C

, C

PROD

) (Felfernig et al., 2015). These

variables and constraints are the vital elements of a

constraint satisfaction problem (Tsang, 1993). A

solution for a constraint satisfaction problem consists

of concrete instantiations of the variables such that all

the specified constraints are fulfilled. In

correspondence to the Recommender Knowledge

Base, under the CSP formalisms stated above, we

define the various components of the knowledge base

that was developed as the basis for CSP-based MaaS

plans filtering as follows:

MaaS Customer Variables V

, refer to each user’s

individual properties. In the domain of MaaS, the

frequency of public transport usage is an example for

a customer variable and public transport

usage=Every day represents a concrete customer

requirement, indicating a daily use of public

transportation services. The most important customer

variables along with the questions used to derive them

are the following:

 Driving license; derived through the question

“Do you hold a full driving license?”

 Public Transport usage; derived through the

question “how often do you use public

transport?”

 Fare reductions; derived through the question

“Are you eligible for any public transport travel

fare reductions?”

 CarSharing usage; derived through the

question “How often do you use car sharing?”

 Taxi usage; derived through the question “How

often do you use Taxi services?”

 BikeSharing usage; derived through the

question “How often do you cycle?”

MaaS Product Variables V

PROD

, refer to the

various attributes of a MaaS plan, including its id,

price and the quota per transport mode that is

available within the period the plan is valid for (e.g. a

month). Examples include the number of taxi, bike

sharing and/or car sharing trips included in the plan,

as well as number of days a Public Transport service

can be used.

MaaS Products C

PROD

, refer to the allowed

instantiations of product properties, which define the

set of available MaaS plans. Indicative examples of

MaaS plans are presented in Table 1, illustrating

product properties’ values (e.g.PT,Taxi etc) within

monthly scale.

Unlimited corresponds to the Large quantity

Table 1: Indicative examples of MaaS Plans.

PROD

Value

Public_transport

30 days

Taxi

4 trips

Bike_Sharing

Unlimited

Car_Sharing

Unlimited

Price

90 euros

Public_transport

15 days

Taxi

2 trips

Price

60 euros

MaaS Constraints C

, refer to the relationship

between Customer and product variables, with the

former constraining the values of the latter. Indicative

examples of MaaS constraints are provided in Table

2, following an object-oriented annotation language.

For example, C

denotes that MaaS plans which

include car sharing are filtered out for users that do

not possess a driving license.

Table 2: Indicative MaaS constraints.

If user.driving license=’No’ then MaaS product.

CarSharing=’0’

If user.Public Transport usage =’Every day’

then MaaS product. Public Transport=’30’ days

If user.Fare Reductions = ‘Yes’ then MaaS

product. Id=’50’or ‘51’ or ‘52’ (special

discounted MaaSPlans)

If user.CarSharing usage =’ Every day’ MaaS

product. CarSharing =’Unlimited’ trips

Given the user preferences (V

) provided through

the aforementioned questions which are embedded in

a knowledge acquisition interface, the MaaS product

definitions (C

PROD

) and the MaaS constraints (C

one or more solutions for the constraint satisfaction

problem are provided by a CSP solver. The solutions

consist of concrete instantiations of the product

variables such that all the specified constraints are

fulfilled, and correspond to specific MaaS plans that

are tailored to the user preferences.

3.2 Similarity-based Plans Ranking

As already mentioned, our approach includes the

calculation of a weighted similarity between a user

and MaaS plans, in the direction of ranking the MaaS

products that satisfy the constraints (i.e. the output of

CSP-based MaaS plans filtering process), on the basis

of user preferences for the various modes of transport

included in the MaaS plan. Many similarity

mechanisms have emerged in Case Based Reasoning

ICE-B 2019 - 16th International Conference on e-Business

(CBR) and data mining research as well as other areas

of data analysis. Most of them assess similarity based

on feature-value descriptions of cases (e.g. items,

users etc.) using similarity metrics that use these

feature values. We adopt such an approach that

follows the so-called intentional concept description

strategy, according to which a concept is defined in

terms of its attributes (e.g. a monthly MaaS plan has

public transportation, taxi, bike sharing and car

sharing usage quotas). This notion of a feature-value

representation is underpinned by the idea of a space

with cases (e.g. MaaS plans) located relative to each

other in this space (Tummas and Ricci, 2009).

Similarly, users are represented as a set of feature-

value pairs with features representing their

preferences for the different modes of transport

included in the MaaS plans, in order to allow the

calculation of similarity between a user and an item,

i.e. a MaaS plan.

Each feature in the representation space is

considered to have a different contribution to

measuring similarity, i.e. each feature is given a

different weight in the user-item similarity

calculation. This is because there may be a variance

in the importance of each feature for similarity

computations, depending on the willingness of each

user to include the respective mode in his/her MaaS

plan. The higher the willingness to include a mode,

the bigger the weight of the respective feature will be.

For example, in case a user is more willing to include

taxi than bike sharing in a MaaS plan, the taxi feature

will be given a bigger weight than the bike sharing

one.

The vector representing a user in the X-

dimensional feature space (with X denoting the

number of distinct modes included in MaaS plans), is

instantiated based on user responses to the questions

about the frequency of public transport, taxi, bike

sharing, and car sharing usage, as described in section

3.1. For example, a value of 0 is given to the taxi

feature of the user vector, in case the user replies in

the relative question, that s/he never uses a taxi

service, while a value of 30 taxi rides is given if the

user replies in the same question that h/she is using a

taxi service “Every day”. The values for other

possible responses will vary between these two

extremes.. The values for the other features of the user

vector are calculated in a similar manner.

The item vectors are instantiated for each MaaS

plan based on the values of the features of MaaS

Product Variables, i.e. the quota per transport mode

that is available within the period the plan is valid for

(e.g. a month), as described in section 3.1. After the

user and MaaS plans vectors have been instantiated

for a specific user and a specific list of MaaS plans

(the output of the CSP-based filtering), all vectors are

normalised and the weighted similarity formula given

below is applied to calculate the similarity between

the user preferences and all MaaS plans of the list.

























where F is the number of attributes (i.e. features) in

each vector (in our case equals the number of distinct

modes included in MaaS plans; four indicatively for

Public Transport, Taxi, BikeSharing and CarSharing

modes), i is an individual feature from 1 to F, w

is the

weight of feature i (derived from a likert scale

question that follows below) andT and S are the two

input vectors for which similarity should be

calculated (i.e. a user and a specific MaaS plan

vectors), Typically, the weights sum to 1 and are non-

negative. The weights are derived from user’s

response to the following question:

“Please define your willingness to include the

following modes of transport in your new MaaS

Plan:”

 Public Transport

 Taxi

 Bike Sharing

 Car Sharing

given within a likert-scale 1-5, with 1 indicating

“Very much” and 5 “Totally not” option. This

question is also embedded in the knowledge

acquisition graphical interface depicted in Figure 1.

The calculated similarities between the user and the

MaaS plans of the list are used to rank the latter and

present them to the user in a tabular form in

descending order, i.e. the first plan is the most similar

to the user preferences and the last one the least

similar.

4 IMPLEMENTATION

For the implementation of the MaaS plans

Recommender we followed a three-tier architecture

as illustrated in Figure 2. The data tier consists of

three data sources already discussed in previous

sections, namely the user profile data, the MaaS plans

data and the list of domain constraints. The user

profile data contain users’ individual preferences.

These are acquired when a user interacts with the

MaaS app, and the corresponding MaaS plan

selection screen, through a set of questions as

A Hybrid Knowledge-based Recommender for Mobility-as-a-Service

described in Section 3.1. The MaaS plans data refer

to a list of available MaaS products, which are

configured by the MaaS operator and form the search

space of the recommendation engine. Both the user

profile data and the MaaS plans data are stored in a

no-SQL MongoDB database whereas the list of

domain constraints is generated and stored in the

filesystem as a data file in .mzn extension, that

corresponds to the MiniZinc model files. The

business logic layer integrates a CSP library based on

the MiniZinc

open-source constraint problem

solving software and a modular similarity calculation

component which provides the means to infer the

similarity of each plan to the user’s profile and

preferences. For the implementation of the RS we

used the Meteor web application framework which is

built on top of NodeJS and the Javascript

programming language. Node.js packages and

modules were used in order to deploy the MiniZinc

CSP solver within the Meteor JavaScript platform.

Figure 2: MaaS plans recommender system architecture.

5 USAGE SCENARIO

Figure 3 depicts the MaaS plans recommendation

process, including all the steps from setting the user

requirements to the recommendation of the final list

of MaaS plans, while highlighting the user-

recommendation engine interactions. Notable here is

the fact that the User Profile record is editable,

meaning that knowledge acquisition from the user

side is performed once and stored in system’s db with

a unique user id, while it is updated every time the

user states different preferences within other MaaS

plans selection efforts.

First, the user opens the MaaS Plans selection screen

and a dialog box-wizard appears asking him/her to

https://www.minizinc.org/

answer a set of questions used for eliciting user

requirements. An indicative view of the interface

showing the questions asked to the user is provided in

Figure 4. This set of questions is used to build the

user’s profile database. The answers are linked to the

MaaS Customer variables defined in Section 3.1. For

instance, the answer to the question “Do you hold a

driving license?”, is used to set the customer variable

“driving_license” as either yes or no (1 or 0). In a

similar manner, all the customer variables are set in

line with the answers to the questions.

Figure 3: Overview of plans recommendation process.

Moreover, the preconfigured list of MaaS Plans is

also stored in the database. Both user profile data and

MaaS plans are fetched by the Mobility Plans

Recommender where the recommendations are

calculated using the CSP and similarity-based

mechanism. In the following, we provide an example

of MaaS plans recommended by our approach for a

particular user and list of available MaaS plans.

A MaaS operator has configured a number of

MaaS plans by following the “McDonald’s self-

customization strategy”, which allows the provision

of small, medium and large quantities of offerings per

mobility service, with each one of the aforementioned

levels corresponding to specific quotas of e.g. bike

sharing rides or days that Public Transport (PT) can

be used. In our example, we consider that the

configured plans contain all the combinations of four

mobility services as follows: a PT service that can be

used for 5 (small), 15 (medium), or 30 (large) days

per month, a Taxi service with values of 3 (small), 7

(medium) and 12 (large) rides per month, a

BikeSharing service with values of 3(small), 6

(medium), or Unlimited (large) hours per month

and a CarSharing service with values of 3

(small), 6 (medium) and Unlimited (large)

hours per month. Note that the aforementioned MaaS

plans configuration is based on a real case follow-

ed by the MaaS service operated by Hannoversche

ICE-B 2019 - 16th International Conference on e-Business

100

Figure 4: The MaaS plans recommender knowledge acquisition Graphical User Interface (GUI).

Verkehrsbetriebe Aktiengesellschaft, the public

transportation company in Hannover, that is

considered a pioneer in MaaS (Röhrleef, 2018). The

above mobility services combinations result in 120

different MaaS Plans.

In our example scenario, a MaaS traveller with no

driving license, who uses frequently public transport

and has a friendly attitude towards Bike Sharing

schemes, is interested in purchasing one of the above

mentioned pre-configured MaaS plans. An instance

of user preferences for MaaS as captured by his/her

responses to the corresponding list of questions is

depicted in Figure 4. The relevant Customer

Variables’ instances are the following:

 User. Driving license= “No”

 User. Public Transport usage= ”Every Day”

 User. Fare reductions= ”No”

 User. CarSharing usage= ”Never”

 User. Taxi usage= ”Once/few times per week”

 User. BikeSharing usage= ”Once/few times per

month”

The aforementioned customer variables

instantiations, along with the MaaS constraints and

the pre-configured MaaS plans are used by our CSP

mechanism to identify the MaaS plans matching the

user preferences. In our example scenario, a list of

MaaS plans satisfying the constraints were given by

the CSP, but for reasons of simplicity only four are

depicted in Table 3, and will further be processed by

the similarity mechanism ( Table 4).

Table 3: Maas Plans Derived by the CSP Mechanism.

Plan

Public

Tran.

Taxi

Bike

Sharing

Car

Sharing

Price

Note that in the example scenario, the CSP-based

filtering resulted in MaaS plans with no car sharing

(i.e. the CarSharing product variable values have been

set to zero) since the user has no driving license and

therefore is not allowed to drive. Moreover, the plans

derived from CSP-based filtering have large

quantities of public transport offerings, since the

particular user stated his preference for a daily use of

that mode of transport. Similarly, the user stated

preference about a low frequency of bike sharing use,

resulted in MaaS plans with small and medium

quantities of bike sharing offerings. In the opposite

direction, the user’s frequent needs for taxi are

covered through medium and large quantities of taxi

offerings. Note that the price values per product have

been calculated based on basic assumptions regarding

the pricing policy of each Mobility provider included

in the MaaS schema.

Thereafter, the weighted similarity function

described in Section 3.2 is applied on the plans

derived by the CSP mechanism. The weights’ values

for each attribute of the example will be set to

 w

=1/(1+2+1+5)=0.11

 w

=2/(1+2+1+5)=0.22

 w

=1/(1+2+1+5)=0.11

 w

=5/(1+2+1+5)=0.56

The calculated similarities are used to rank the four

plans as depicted in Table 4. The plans are presented

to the user in a tabular form in descending order, i.e.

A Hybrid Knowledge-based Recommender for Mobility-as-a-Service

101

the first plan is the most similar to the user

preferences and the last one the least similar.

Table 4: The ranked list of MaaS plans based on the

weighted similarity to the user’s preferences.

Plan id

Weighted similarity to the user’s

preferences

0.9969

0.9967

0.9897

0.9896

It should be noted that the user can set the maximum

price s/he is willing to pay for a MaaS plan through a

slider widget embedded in the plans selection screen.

MaaS plans with a higher price than the user-provided

maximum are filtered out, while the rest are passed to

the MaaS recommender system for CSP and

similarity-based filtering. The user choice about the

maximum MaaS price can be changed at any time in

the context of a single session. Each time the user

choice changes, the recommender is triggered to

recommend a subset of plans out of those that have a

lower price than the maximum. Finally, the user

chooses one of the suggested plans for the given

budgetary constraints and preferences.

6 CONCLUSIONS

In this paper, we presented a hybrid knowledge-based

recommender system for users of the Mobility as a

Service (MaaS) mobility paradigm. MaaS aims to

provide integrated and seamless access to transport

services through one single digital platform. In a

MaaS environment there can be a multitude of MaaS

plans, that include combinations of transport services,

in order to meet the specific needs of different types

of travellers. Our recommender supports travellers’

decisions related to the selection of MaaS plans by

combining Constraint Satisfaction Problem solving

and weighted similarity mechanisms in order to

compute a personalized ranked list of MaaS plans

aligned to the preferences of travellers who are about

to use them.

To the best of our knowledge the proposed hybrid

recommender constitutes the first attempt for

personalizing the MaaS plans selection process while

the knowledge-based approach tackles the cold start

problem which refers to the lack of data for deriving

user needs and preferences. However, the approach

relies on knowledge engineers who need to define the

set of rules for filtering the MaaS plans. Such

engineers may not always be available whereas the

knowledge acquisition process can become

complicated when many rules need to be defined. In

order to mitigate the above limitations, we plan to

explore combinations of our approach with data-

driven ones. More specifically, by analysing user

mobility data, such as GPS tracks, we could

automatically infer user needs and preferences for

specific transport services, as well as understand how

these change in time. Such information can be used to

automatically modify the suggestions when user

needs and preferences change, and overcome the

knowledge acquisition challenge.

As part of our next steps, we are in the process of

evaluating our proposed approach and system in real

life conditions where travellers from the cities of

Manchester, Budapest and Luxemburg will be using

a MaaS app integrating our knowledge-based MaaS

plans recommender. Our aim is to test our approach

and measure the effectiveness and benefits of MaaS

plans suggestions to travellers.

ACKNOWLEDGEMENTS

Research reported in this paper has received funding

by the H2020 EC project MaaS4EU (GA no.

723176).

REFERENCES

Beladev, M., Rokach, L., and Shapira, B. (2016).

Recommender systems for product bundling.

Knowledge-Based Systems, 111, 193-206.

Brailsford, S. C., Potts, C. N., and Smith, B. M. (1999).

Constraint satisfaction problems: Algorithms and

applications. European Journal of Operational

Research, 119(3), 557-581.

Christakopoulou, K., Radlinski, F., and Hofmann, K. (2016,

August). Towards conversational recommender

systems. In Proceedings of the 22nd ACM SIGKDD

International Conference on Knowledge Discovery and

Data Mining (pp. 815-824). ACM.

Durand, A., Harms, L., Hoogendoorn-Lanser, S., and

Zijlstra, T. (2018). Mobility-as-a-Service and changes

in travel preferences and travel behaviour: a literature

review. KiM| Netherlands Institute for Transport Policy

Analysis.

Felfernig, A., and Burke, R. (2008, August). Constraint-

based recommender systems: technologies and research

issues. In Proceedings of the 10th International

Conference on Electronic Commerce (p. 3). ACM.

Felfernig, A., Friedrich, G., Jannach, D., and Zanker, M.

(2006). An integrated environment for the development

of knowledge-based recommender applications.

ICE-B 2019 - 16th International Conference on e-Business

102

International Journal of Electronic Commerce, 11(2),

11-34.

Felfernig, A., Friedrich, G., Jannach, D., and Zanker, M.

(2015). Constraint-based recommender systems.

In Recommender systems handbook (pp. 161-190).

Springer, Boston, MA.

Freuder, E. C., and Mackworth, A. K. (2006). Constraint

satisfaction: An emerging paradigm. In Foundations of

Artificial Intelligence (Vol. 2, pp. 13-27). Elsevier.

Guo-rong, L., and Xi-zheng, Z. (2006, October).

Collaborative filtering based recommendation system

for product bundling. In 2006 International Conference

on Management Science and Engineering (pp. 251-

254). IEEE.

Jannach, D., Zanker, M., and Fuchs, M. (2009). Constraint-

based recommendation in tourism: A multiperspective

case study. Information Technology & Tourism, 11(2),

139-155.

Junker, U., and Mailharro, D. (2003). Preference

programming: Advanced problem solving for

configuration. AI EDAM, 17(1), 13-29.

Kamargianni, M., and Matyas, M. (2017). The business

ecosystem of mobility-as-a-service. In transportation

research board (Vol. 96). Transportation Research

Board.

Lam, X. N., Vu, T., Le, T. D., and Duong, A. D. (2008,

January). Addressing cold-start problem in

recommendation systems. In Proceedings of the 2nd

international conference on Ubiquitous information

management and communication (pp. 208-211). ACM.

Linoff, G. S., and Berry, M. J. (2011). Data mining

techniques: for marketing, sales, and customer

relationship management. John Wiley & Sons.

Mailharro, D. (1998). A classification and constraint-based

framework for configuration. Ai Edam, 12(4), 383-397.

Murphy, S. Ó., Manzano, Ó., and Brown, K. N. (2015,

August). Design and evaluation of a constraint-based

energy saving and scheduling recommender system.

In International Conference on Principles and Practice

of Constraint Programming (pp. 687-703). Springer,

Cham.

Reiterer, S., Felfernig, A., Jeran, M., Stettinger, M.,

Wundara, M., and Eixelsberger, W. (2015). A Wiki-

based Environment for Constraint-based Recommender

Systems Applied in the E-Government Domain.

In UMAP Workshops.

Rohrleef, (2018), Mobility as a Service (MaaS) – One stop

Mobility Shop for Hannover (Germany). Available at

https://slideplayer.com/slide/13282823/

Schumacher, M., and Rey, J. P. (2011, January).

Recommender systems for dynamic packaging of

tourism services. In ENTER(pp. 13-23).

Tsang, E. Foundations of constraint satisfaction.

1993. Cited on, 14.

Tumas, G., and Ricci, F. (2009). Personalized mobile city

transport advisory system. Information and

communication technologies in tourism 2009, 173-183.

Zanker, M., Aschinger, M., and Jessenitschnig, M. (2010).

Constraint-based personalised configuring of product

and service bundles. International Journal of Mass

Customisation, 3(4), 407-425.

Zhang, Z., Lin, H., Liu, K., Wu, D., Zhang, G., and Lu, J.

(2013). A hybrid fuzzy-based personalized

recommender system for telecom products/services.

Information Sciences, 235, 117-129.

A Hybrid Knowledge-based Recommender for Mobility-as-a-Service

103