Enhancing Individual Mobility: A Multistage Personalization Approach

for Itinerary Planning in Multimodal Networks

Alexandra Wins

, Christoph Becker

, Sascha Alpers

, Lukas Kneis

and Andreas Oberweis

FZI Research Center for Information Technology, Haid- und Neu-Str. 10-14, 76131 Karlsruhe, Germany

Hochschule Heilbronn, Max-Planck-Str. 39, 74081 Heilbronn, Germany

Keywords:

Mobility, Itinerary Planning, Personalisation, Routing, Multimodal Networks.

Abstract:

Individual mobility is an essential element of a prosperous society. Multimodal transportation can offer greater

time, cost, and environmental efﬁciency than relying on a single mode of transport. Personalized itinerary

planning is crucial to enhance the appeal of multimodal transport. Our proposed approach for recommending

personalized itineraries tailors them by integrating diverse mobility preferences, routing services, and calibrat-

ing parameters of these services to provide individualized options. We optimize itineraries within the existing

routing services and available data. The aim of this approach is to enhance travel experiences, making them

more efﬁcient, cost-effective, and aligned with each traveler’s unique needs and preferences. The approach

was evaluated in a mid-sized German city by analyzing real-world mobility preferences, available routing

services, and mobility providers. Personalization criteria relevant to the evaluation area were selected. A sim-

ulation was conducted, which demonstrated a 10.48% increase in travel utility when compared to the shortest

path itinerary recommendation.

1 INTRODUCTION

Individual mobility is essential in advancing soci-

etal well-being, particularly in promoting work-life

balance. Multimodal transportation can be a cost-

effective and time-efﬁcient alternative to a reliance on

a single transport mode in urban areas, reducing traf-

ﬁc congestion and enhancing overall travel satisfac-

tion. Personalized travel suggestions can improve the

appeal of multimodal transportation by incorporating

the diverse preferences individuals consider when se-

lecting a route. These preferences, which may change

based on situational context or travel purpose, can

also vary throughout the day or within a route.

When choosing a route, individuals take into ac-

count a variety of mobility preferences, whether con-

sciously or unconsciously. However, the routing rec-

ommendation systems can only integrate preferences

that it can evaluate and interpret. For example, pref-

erences such as the “behavior of cyclists” may be im-

portant for car drivers, but their integration depends

on available data, which may not be uniformly col-

lected in all regions of the world. Incorporating such

preferences will not enhance personalization if the

system cannot assess them. In addition, each pref-

erence introduces a new optimization criterion, which

requires an estimation of its importance to the indi-

vidual users.

This paper addresses these challenges by conduct-

ing an analysis of mobility preferences and propos-

ing a uniﬁed itinerary recommendation system with

a multi-stage personalization strategy that integrates

diverse preferences, transport modes, and routing ser-

vices. To achieve this, we propose a strategic se-

lection process for integrating multiple routing ser-

vices into a uniﬁed system. The proposed approach

aims to incorporate a speciﬁc subset of routing ser-

vices that maximizes the overall number of supported

personalization options and travel modes. Further-

more, the paper presents an algorithm for identifying

utility-maximizing parameterization of the selected

routing services. Finally, our proposed approach ad-

dresses the issue of overchoice, where individuals are

faced with too many options, making the decision-

making process overwhelming. This is achieved by

dynamically estimating the optimal routes for indi-

vidual users based on their mobility preferences and

presenting the user with only the top three routes.

We evaluate our approach in a medium-sized

German city, where we analyze travelers’ mobility

preferences, available routing services, and mobility

providers. As an example, we identify a speciﬁc sub-

Wins, A., Becker, C., Alpers, S., Kneis, L. and Oberweis, A.

Enhancing Individual Mobility: A Multistage Personalization Approach for Itinerary Planning in Multimodal Networks.

DOI: 10.5220/0012637500003702

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 10th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2024), pages 319-326

ISBN: 978-989-758-703-0; ISSN: 2184-495X

319

set of mobility preferences and routing services that

can be integrated into the travel planner in the evalua-

tion region. Finally, we integrate the selected services

into a uniﬁed itinerary recommendation platform and

perform simulations on a real-world transport net-

work using real-time travel information to evaluate

the effectiveness of our proposed approach.

2 RELATED WORK

Recently, multiple frameworks for personalized rout-

ing have been developed. In (Nuzzolo and Comi,

2016), researchers proposed a utility-based approach

for path suggestions that considers individual mobil-

ity preferences, improving path advice performance

over using average preferences. The FAVOUR frame-

work, which offers personalized route recommenda-

tions based on situational awareness, has been intro-

duced in (Campigotto et al., 2016). This method con-

sists of three stages: the initialization stage, which

chooses one of a few initial proﬁles based on gen-

eral user information; the second stage, which per-

sonalizes the proﬁle further through a stated pref-

erence survey; and the third stage, where the pro-

ﬁle is continuously updated through analysis of ac-

tual choices the users make. The method proposed in

(Lathia et al., 2012) focuses on personalized user ex-

periences by utilizing a hypernetwork that intercon-

nects private and public transport networks. It opti-

mizes advice based on individual preferences and for

different modes of transport. The system generates

customized routes for users based on factors such as

speed, cost-effectiveness, convenience, and environ-

mental impact. Personalized parameters are loaded

to the network to generate routes for each individual

case and user.

While the approaches mentioned above focus

on developing a personalized itinerary recommenda-

tion framework, our approach suggests an alternative

method for offering extensive personalization options

within a single platform without requiring the devel-

opment of a new routing service. This can be achieved

by integrating multiple services into a uniﬁed plat-

form. A similar system has been introduced in (Spi-

tadakis and Fostieri, 2012). A WISETRIP planner is

an innovative multimodal journey planner that inte-

grates multiple journey planners by providing a com-

munication interface to connect heterogeneous plan-

ners. The approach introduced in (Eszterg

ar-Kiss

et al., 2022) aims to create mobility solutions and

establish multimodal transport networks that connect

different systems. This is achieved by identifying ap-

propriate exchange points between separate networks

to implement the routing algorithm using various lo-

cal journey planners. In contrast to the approach pre-

sented in this paper, the aforementioned approaches

do not address the issue of the parameters calibration

of the integrated services.

3 PERSONALIZED ITINERARY

RECOMMENDATION SYSTEM

In order to integrate a wide range of mobility pref-

erences, transport modes, and routing services in a

uniﬁed platform, we propose a personalized itinerary

recommendation system with a multi-stage personal-

ization approach, as depicted in Figure 1. In the ini-

tial step, user preferences are obtained through sim-

ple surveys and choice experiments. Subsequently,

in the preprocessing phase, relevant data is gathered

from external sources. Personalization rules, derived

from user preferences and gathered data (e.g., avoid-

ing cycling if it is raining), are applied to prune pos-

sible travel options. In the following phase, external

routing services are requested using the start and des-

tination addresses speciﬁed by the user and a utility-

maximizing parameterization that is tailored speciﬁ-

cally to individual user preferences. In the ﬁnal post-

processing phase, routes are prioritized based on their

utility for a speciﬁc user and preference proﬁle. The

best three options are then selected and visualized to

the user. Subsequent sections will provide a detailed

description of each stage, including the necessary ini-

tialization and implementation steps, as well as a de-

scription of the functionality and workﬂow for each

stage.

3.1 Creation of Preference Proﬁle

The preference proﬁle enables the deﬁnition and esti-

mation of individual mobility preferences of the users.

These preferences can vary depending on the region

of operation.

3.1.1 Initialization and Conﬁguration

The prerequisite for implementing the preference pro-

ﬁle is the analysis of the accessibility of single pref-

erences in the region of operation. To achieve effec-

tive personalization, it is crucial for the system to ﬁl-

ter out inaccessible preferences, such as the compliant

behavior of other road users, and focus solely on pref-

erences with relevant available data. To identify such

preferences, one must ﬁrst establish their interpreta-

tion and then the source of the data for that prefer-

ence. Attributes like safety need interpretation spe-

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems

320

Start

Preference profile

creation

Initialize

preference

weights

General

preferences

Gather relevant

data

1. Creation of Preference Profile

Apply rule-based

constraints and

filter travel options

2. Preprocessing

Mapping of

preferences on

routing parameters

Request routing

services

OTP, Brouter,

Naviki etc.

3. Calibration of routing services

Priorisation and

visualisation of

routing results

Analysis of user

selection

4. Postprocessing

Route selection

Sociodemographic

factors

Transfer learning

External services

(e.g. Breezometer)

Choice

Experiments

Figure 1: Multi-stage Personalisation Process.

ciﬁc to each mode. For instance, safety while cycling

can be interpreted as the number of left turns, dis-

tance traveled along motorized roads, or the number

of accidents on a speciﬁc route. Preferences should

be considered in the personalization process if one of

the following conditions is met:

• The preference can be used as a means of pre-

selecting a transportation option, such as avoiding

biking in the rain.

• At least one routing service provides the neces-

sary personalization option, such as wheelchair

accessibility.

• Preference can be evaluated by utilizing available

data sources, such as user input (e.g., the presence

of a driver’s license) or open data (e.g., air qual-

ity).

3.1.2 Workﬂow and Functionality

In our itinerary recommendation system, users ﬁrst

mation and static preferences that remain constant

across different trip contexts, such as mobility im-

pairment. Users can then create separate preference

proﬁles for different situational contexts, incorporat-

ing dynamic preferences that adjust based on these

contexts. For example, a proﬁle labeled as “travel”

may require the accommodation of luggage, while a

proﬁle labeled as “night travel” may prioritize illumi-

nated paths. Static preferences are set during regis-

tration but can be modiﬁed within situational prefer-

ence proﬁles to reﬂect changes, such as traveling with

someone who has a mobility impairment. The system

proposed in (Nuzzolo and Comi, 2016) distinguishes

between rule-based preferences, which can be easily

deﬁned by users through simple surveys (such as a

maximum number of transfers), and more complex

preferences (such as “security” or “reliability”) that

lack a direct rule-based application and require pri-

oritization or comparison for optimization. Further-

more, these preferences may conﬂict with each other.

Utility theory can be used to represent conﬂicting ob-

jectives by quantifying the importance of each pref-

erence (Campigotto et al., 2016; Nuzzolo and Comi,

2016). The utility of a route is calculated as the sum

of weighted utilities for each route attribute, address-

ing conﬂicting criteria. The utility function of a route

r is illustrated in Equation 1, where β is a baseline

utility, α

is a binary value, depicting the relevance

of the attribute a

to the mode of the route (e.g., trafﬁc

is irrelevant for the train routes; therefore α will take

the value of 0), β

is the weight of the corresponding

attribute a

. The utility of a route U(r) is the sum of

weighted attributes a

U(r) = β +

∑

i=1

∗ β

∗ α

(1)

Quantifying the importance of a wide variety of pref-

erences can be complicated and time-consuming for

a user. Therefore, we propose the iterative approach

for learning individual user preferences based on the

approach outlined in (Campigotto et al., 2016). The

initial sociodemographic data is employed to estab-

lish default weights of the preferences. For example,

the speed parameter could be determined based on the

average speed for a speciﬁc gender and age. A more

sophisticated strategy involving transfer learning, as

suggested in (Campigotto et al., 2016), can be em-

ployed to initialize user preferences. An alternative

method for initializing the default weights is to uti-

lize the weights obtained from surveys conducted in

the operation region. These default weights can be

further adjusted for each individual user using choice

experiments (Campigotto et al., 2016). The method

used in this study for generating choice experiments

and estimating individual utility functions based on

the results of these experiments is described in (Wins

et al., 2024) and follows the method outlined in (Lou-

viere et al., 2008).

Enhancing Individual Mobility: A Multistage Personalization Approach for Itinerary Planning in Multimodal Networks

321

3.2 Preprocessing

The preprocessing phase is responsible for ﬁltering

travel options before the actual optimization process.

The ﬁlter criteria are derived from the available data

and rule-based constraints deﬁned in the preference

proﬁle.

3.2.1 Initialization and Conﬁguration

In order to customize the route, data needs to be gath-

ered from different services, such as Breezometer for

weather and air quality. The choice of these services

depends on the travel region. To enable preprocess-

ing, the analysis and integration of the available ser-

vices in the operational region must be conducted dur-

ing implementation.

3.2.2 Workﬂow and Functionality

Based on the selected preference proﬁle and travel de-

mand, the travel options and requested routing ser-

vices are ﬁltered in the preprocessing phase. In the

ﬁrst step, data required for the personalization (e.g.,

weather) is requested from various services available

in the region of operation. In the subsequent step,

appropriate transport modes and routing services are

determined based on the user preferences and avail-

able data (including real-time data). For instance, if

a user prefers not to cycle during pollination, modes

involving cycling (bike, bike-sharing, bike and ride)

will be eliminated from further consideration. Fur-

thermore, routing services are selected based on the

required personalization options and available data.

For instance, if a user prefers public transport and

wheelchair accessibility, only services with these op-

tions are chosen. However, preferences relying on

real-time data, such as weather and pollination, may

not be used if information is unavailable at the re-

quest time. Therefore, service selection should con-

sider not only feature support, like avoiding speciﬁc

streets when computing the route but also data avail-

ability.

3.3 Calibration of Routing Services

The third phase of our proposed approach to comput-

ing personalized itineraries involves requesting rout-

ing services that support the speciﬁed personalization

options and desired travel modes and cover the re-

quired areas. To achieve route personalization, map-

ping of the mobility preferences from the selected

proﬁle to the routing request parameters of each re-

quested routing service is essential.

3.3.1 Initialization and Conﬁguration

The ﬁrst prerequisite for the implementation of the

routing calibration stage is the selection of the rout-

ing services to integrate into the itinerary recommen-

dation platform. Routing services such as Brouter and

OpenTripPlanner(OTP) offer extensive customization

of algorithm parameters to cater to diverse user needs.

However, due to their complexity, integrating these

services requires time to understand their features

through documentation and source code. In contrast,

services like Naviki offer predeﬁned proﬁles, such as

leisure and mountain biking, without the option to ad-

just internal parameters. This simpliﬁes use but limits

personalization.

To identify the most effective combination of rout-

ing services for a speciﬁc region, an analysis should

be conducted to determine the intersection of routing

services that maximize the support for preferences,

modes, and region coverage without causing a neg-

ative impact on performance. Although routing ser-

vices can be requested concurrently, prolonged pro-

cessing time for a single service may adversely affect

the overall system performance. Thus, routing ser-

vices should be selected under consideration of the

trade-off between supporting additional features and

maintaining optimal performance.

The second prerequisite for the implementation of

the routing calibration stage is the integration of var-

ious routing services in the system. We propose us-

ing the Adapter Pattern (Gamma, 1995) for this pur-

pose. This pattern assists in adapting the interface of

the itinerary planner to various routing service inter-

faces. Each newly added routing service requires an

adapter implementation that supports routing requests

and provides information on the supported prefer-

ences, transport modes, and regions. This data is uti-

lized during the preprocessing stage to determine the

suitable routing services. Additionally, the adapter

consolidates routing responses and returns them to the

itinerary planner in a uniﬁed data format.

3.3.2 Workﬂow and Functionality

The personalization of routing can be achieved

through the utility-maximizing calibration of the pa-

rameters of routing services. In this process, users’

mobility preferences are mapped onto the technical

parameters of the respective routing services. These

mappings can be complex due to variations in pa-

rameter value ranges, data types and sensitivities. To

overcome these issues, we suggest a parameter cal-

ibration approach based on simulated annealing and

utility theory.

The proposed approach, which is based on the

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems

322

method described in (Teodoro et al., 2017), involves

identifying and pruning non-inﬂuential parameters

while simultaneously auto-tuning inﬂuential param-

eters. To enhance calibration performance, a sensi-

tivity analysis is conducted to prune non-inﬂuential

parameters - those with minimal or no impact on

the routes - and exclude them from further analysis.

This pruning step can substantially reduce the search

space. Parameter calibration is performed individ-

ually for each preference proﬁle and transportation

mode. This is necessary because certain parameters

may only be relevant for speciﬁc modes and prefer-

ence proﬁles. For example, preferences such as bi-

cycle speed are only relevant to modes such as bik-

ing, bike-sharing, and bike-and-ride. It is also cru-

cial to consider the interactions between parameters

to ensure accurate calibration. For instance, giving

equal weight to conﬂicting parameters such as “scenic

route” and “fastest route” may lead to unsatisfactory

outcomes. Resolving conﬂicting parameterization of-

ten relies on the internal algorithms of the routing ser-

vice being used. To address this problem, it is crucial

to identify an appropriate parameter mapping.

For sensitivity analysis, we propose using the

Morris Method (Morris, 1991) to evaluate the impact

of changes in individual parameters on routing. As-

sessing the impact based on a single route can be mis-

leading due to route variability, including the avail-

ability of direct public transport connections. There-

fore, we propose a systematic selection of multiple

routes across the evaluation region, aiming for com-

prehensive coverage that considers the trade-off be-

tween high-quality results and the computational efﬁ-

ciency of the analysis. These routes form a test set T .

We consider parameters that affect at least one route

as inﬂuential, recognizing that changes to parameters

can affect routes differently.

To assess the elementary effect of a parameter

, routes are generated for start and destination

using the default parametrization of a routing ser-

vice P

= {p

,...p

...p

}. Subsequently, routes are

computed with an alternative parametrization P

,...p

...p

}, where p

̸= p

, generated based on

the Morris Method, while keeping all other parame-

ter values unchanged. The resulting routes R

and R

under parametrizations P

and P

are then compared.

Initially, the comparison is based on spatial factors,

particularly the area enclosed by the two routes R

and R

(see Figure 2). Subsequently, the compari-

son focuses on temporal factors. The temporal factors

considered are:

• Difference in the start times of routes R

and R

• Difference in the end times of routes R

and R

• Duration of routes R

and R

Route A

Route B

Area between routes

Figure 2: Area between two routes.

Input: P

de f ault

= default parametrization

T = set of start/end coordinates

U(x) = utility function

N = maximum iterations

Result: utility-maximizing parametrization

max

← null;

current

← P

de f ault

;

while i ← 0 to N do

if (P

max

̸= null) then

← random({p

,..p

} ∈

current

)

val ←

random(set o f possible p

values)

current

← val

end

sum ← 0

forall t in T do

r ← computeRoute(t, P

current

)

u ← U(r)

sum ← sum + u

end

f (P

current

) ← sum/T.size()

if P

max

= null or f (P

current

) > f (P

max

)

then

max

← P

current

;

end

Algorithm 1: Simulated Annealing.

After completing the sensitivity analysis and prun-

ing non-inﬂuential parameters, the subsequent step is

auto-tuning the remaining parameters. Similar to the

sensitivity analysis, the auto-tuning process is carried

out for a selected mode, preferences proﬁle, and test

set T . We suggest employing Simulated Annealing

for the auto-tuning process. The pseudocode of the

used algorithm is outlined in Algorithm 1. Neigh-

bours are generated by randomly choosing a param-

eter p

and then randomly selecting a value from the

value range associated with that parameter p

. The

ﬁtness function is determined by the utility function

of a chosen preference proﬁle, as deﬁned in Equation

1. To compute the ﬁtness of a new parametrization

current

, routes are requested for every combination

of start and destination coordinates from the test set

Enhancing Individual Mobility: A Multistage Personalization Approach for Itinerary Planning in Multimodal Networks

323

T using the parametrization P

current

. The utility of

each calculated route is then evaluated using the util-

ity function U(x) corresponding to the selected pref-

erence proﬁle. The ﬁtness of the new parametrization

f (P

current

) is calculated as the mean of the utilities

across all generated routes from test set T .

Once the utility-based parametrization is com-

puted, it can be used to request personalized routes

for individual users and preference proﬁles.

3.4 Postprocessing

In the postprocessing phase routes are consolidated

and rated based on user preferences. Postprocessing

begins with the consolidation of responses: these are

transformed into a single format whereby duplicates

are eliminated. The remaining routes are rated and

prioritised based on the selected preference proﬁle.

To avoid overchoice, the user is presented with only

three routes with the highest utilities. The actual route

choice of a user can be used to update the utility func-

tion of the preferences proﬁle using the method de-

scribed in (Campigotto et al., 2016).

4 EVALUATION

The evaluation of the proposed approach was carried

out in a mid-sized city in Germany. Initially, we cre-

ated training and validation sets T and T

′

by dividing

the evaluation region into 30 equal grids. We have de-

ﬁned the geofence of the evaluation region as a quad-

rangle with the following coordinates: (49.08144;

8.34344), (49.08189; 8.55081), (48.95559; 8.54806),

(48.95491; 8.25967). The overall area of the evalua-

tion region is 254 km

, which we divided into 30 grid

areas of approximately 8.5 k m

each. From each grid,

two coordinates t

and t

are selected and added to co-

ordinate lists T

and T

, respectively. The training set

T is subsequently formed by considering all potential

combinations of coordinates from the list T

, exclud-

ing those where the start and end destinations are the

same. The divergent validation set T

′

is formed anal-

ogously from the coordinates from the list T

Regional routing services and mobility and data

providers were analyzed to select the mobility pref-

erences that can be integrated into itinerary planning

based on relevant data availability. Our proposed

itinerary planner considers several static preferences,

such as mobility impairment, subscriptions, payment

methods, favorite places and routes, and type of pri-

vate vehicle (e.g., electric or diesel car). This infor-

mation is gathered through a general survey, along

with basic sociodemographic data. The itinerary plan-

ner allows for the deﬁnition of multiple situational

preference proﬁles, which incorporate further prefer-

ences of two types: those with and without rule-based

application, as suggested in (Nuzzolo and Comi,

2016). The following preferences with rule-based ap-

plications have been integrated: luggage, maximum

number of transfers, minimum transfer time, car park

time, car pickup time, speed, maximum distance for

walking and cycling, and environmental friendliness.

The user can directly adjust the default values of these

preferences, which are based on the default values of

OTP (OpenTripPlanner, 2024).

The user’s utility function (see Equation 1) in-

corporates more complex preferences without rule-

based application. In our itinerary planner, we have

incorporated the following preferences into the util-

ity function: travel mode preference, travel time re-

luctance, travel cost reluctance, waiting time reluc-

tance, access and egress walk time reluctance, access

and egress mode preference, and elevation reluctance.

These preferences can be accessed through choice ex-

periments, as described in section 3.1, or learned by

the system based on the user’s previous choices (Ar-

entze, 2013) or transfer learning (Campigotto et al.,

2016). The preferences are then incorporated through

parameterization and calibration of routing services,

followed by prioritization of routes during the post-

processing phase.

It is important to note that the preferences in-

volved in routing parametrization and post-processing

may be different. For example, cycling speed is only

relevant for routing parametrization, not for evaluat-

ing route utility in the postprocessing phase. When

selecting preferences for routing parametrization, it is

important to examine the available routing services

in the evaluation region thoroughly. In our analy-

sis, we have considered the following routing ser-

vices: BlaBlaCar, Brouter, Cycle.travel, Graphhop-

per, HERE, Mapbox, MAPQUEST, Naviki, Open-

RouteService, OpenTripPlanner, Trassenﬁnder, Trias,

TripGo, TomTom. An analysis of the personalization

options and supported modes provided by each rout-

ing service has been conducted. This analysis identi-

ﬁes the preferences applicable during the calibration

phase of routing services, which are detailed in Table

1. The selection of OTP, Valhalla, and TomTom con-

stitutes a minimal set of routing services that collec-

tively maximize the range of personalization options

in the evaluation region.

The route planning in this study focuses on inte-

grating parameters aligned with individual mobility

preferences, neglecting those that could inﬂuence the

performance of routing algorithms. For instance, such

parameters as “search window” have not been incor-

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems

324

Table 1: Routing parametrization.

Preference Routing Services Type

Mobility

impairment

OTP, Valhalla,

Trias, TripGo,

Graphhopper

Travel time all F

Number of

transfers

OTP RB

Transfer time OTP RB

Transfer OTP F

Waiting time OTP F

Board with bike OTP F

Mode reluctance OTP F

Car park time, cost OTP RB

Car pickup time OTP RB

Speed OTP, Valhalla,

Brouter,

Graphhopper,

TomTom

Distance

(walking/cycling)

all RB,

Elevation Brouter, Valhalla,

OTP

Road type (e.g.

highway)

Brouter, Valhalla F

Road condition

(e.g. paved)

Valhalla F

Existence of a

cycling path

Brouter, Valhalla F

Existence of a

sideway

Brouter, Valhalla F

Safety OpenTripPlanner F

Environmental

friendliness

TomTom RB

Windigness TomTom F

Hilligness TomTom F

Barriers Brouter, OTP F

porated. Routing services commonly employ both

rule-based preferences (e.g., speed) with direct as-

signments and preferences that can be calibrated to

match a user’s speciﬁc preferences or situational con-

text (e.g., walk reluctance). The latter type of parame-

ters is ﬂexible and lacks universal values across differ-

ent routing services, requiring identiﬁcation for each

speciﬁc service. The “Type” column in Table 1 spec-

iﬁes whether a preference is rule-based (RB) or ﬂexi-

ble (F).

After selecting the routing services and deﬁning

the utility function, the routing services calibration,

detailed in section 3.3, has been evaluated for OTP

(version 2.2). Considering space constraints, we will

solely present the results acquired for OTP for the bi-

cycle mode. The parameters “Car reluctance”, “Car

pickup cost” were not considered during the calibra-

tion for the bicycle mode. The sensitivity analysis re-

sults, both spacial and temporal, showed that the pa-

rameters “Bike switch cost” and “Optimize type” ap-

peared to have the most inﬂuence on the course of the

route. To account for this, the simulated annealing al-

gorithm was adjusted to modify this parameter more

frequently than the others. None of the parameters

was be excluded from further consideration.

To evaluate the effectiveness of the utility-based

auto-tuning approach in the bicycle scenario, we have

generated 100 utility functions u ∈ U . The values of

the weights of route attributes have been randomly se-

lected from the range [-1; 0.1]. This range was se-

lected based on the results of the travel preferences

studies, such as (Arentze and Molin, 2013). As we are

conducting a simulation, we have set the base utility

β to zero for all utility functions. Subsequently, we

have performed calibration of OTP with each utility

function u ∈ U

bike

, for bicycle mode and a previously

deﬁned training set T to obtain a utility-maximizing

parametrization for each particular utility function u ∈

bike

−60 −50

−40 −30 −20 −10 0

default

with utility

Figure 3: Average utility distributions of bicycle itineraries

with and without personalization.

To comprehensively evaluate the calibration re-

sults, we use a validation set T

′

. Routes for the

mode bicycle are computed with OTP for each com-

bination of coordinates from the validation set T

′

us-

ing the default parametrization P

de f ault

and the per-

sonalized parametrization P

obtained from the cal-

ibration with the utility function u. For each route

computed using parametrization P

and a route

de f ault

computed using parametrization P

de f ault

, we

employ the utility functions u to estimate the utilities

of these routes, denoted as U (r

) and U (r

de f ault

) re-

spectively. The difference in utilities U(r

de f ault

) and

U(r

) is then compared to assess the enhancement

in satisfaction with the proposed route when utiliz-

ing the calibrated parametrization. On average, there

is a 10.48% improvement in utility. Figure 3 visually

presents the utility distributions for routes computed

with and without parameter calibration for OTP bi-

Enhancing Individual Mobility: A Multistage Personalization Approach for Itinerary Planning in Multimodal Networks

325

cycle mode. Finally, a carried-out paired t-test with a

resulting p-value of 2.2e-16 ensured the statistical sig-

niﬁcance of the observed differences, afﬁrming that

it cannot be attributed to random ﬂuctuations. How-

ever, it is essential to evaluate each routing service

separately to ensure that routing results based on cal-

ibrated parametrizations improve travel satisfaction.

5 CONCLUSIONS

Individual mobility is pivotal for societal well-being,

and multimodal transportation offers an efﬁcient al-

ternative to exclusive car use in urban areas. The

process of personalizing travel suggestions based on

diverse preferences can enhance the attractiveness of

multimodal transportation. The proposed multi-stage

personalization approach exhibits the capability to ef-

ﬁciently integrate a broad range of mobility prefer-

ences and routing services. Leveraging the adapter

pattern makes it highly adaptable for different regions

worldwide. New operational regions can be inte-

grated by including additional routing services and

data sources in the system. The proposed routing cal-

ibration approach helps establish utility-maximizing

mapping rules for each preference proﬁle and routing

service, which is particularly important for efﬁciently

exploiting the personalization capabilities of highly

customizable routing services. Additionally, a utility-

based comparison of routing options tailored to indi-

vidual users promises an enhanced user experience.

However, challenges persist, including managing ex-

tensive preferences, estimating their signiﬁcance, and

ensuring data availability. An additional critical issue

is the quality of data, which can be addressed through

crowdsensing. The platform’s route recommendation

quality will improve with more users, mitigating the

need for a physical route assessment. Despite these

complexities, our proposed approach has the poten-

tial to enable the transition towards more personalized

and efﬁcient itinerary recommendations in the realm

of mobility services.

ACKNOWLEDGEMENTS

The content of this paper is the result of the project

“MobAPlan - Mobility and Activity-based Planning

Assistant”. This research and development project is

funded by the Vector Stiftung (Vector Foundation).

REFERENCES

Arentze, T. A. (2013). Adaptive personalized travel infor-

mation systems: A bayesian method to learn users’

personal preferences in multimodal transport net-

works. IEEE Transactions on intelligent transporta-

tion systems, 14(4):1957–1966.

Arentze, T. A. and Molin, E. J. E. (2013). Travelers’

preferences in multimodal networks: Design and re-

sults of a comprehensive series of choice experiments.

Transportation Research Part A: Policy and Practice,

58:15–28.

Campigotto, P., Rudloff, C., Leodolter, M., and Bauer, D.

(2016). Personalized and situation-aware multimodal

route recommendations: the favour algorithm. IEEE

Transactions on Intelligent Transportation Systems,

18(1):92–102.

Eszterg

ar-Kiss, D., Ansariyar, A., and Katona, G. (2022).

Interconnecting separate transportation systems by in-

troducing exchange points. In 2022 IEEE Interna-

tional Smart Cities Conference (ISC2), pages 1–6.

IEEE.

Gamma, E. (1995). Design patterns. Pearson Education

India.

Lathia, N., Capra, L., Magliocchetti, D., De Vigili, F.,

Conti, G., De Amicis, R., Arentze, T., Zhang, J., Cal

ı,

D., and Alexa, V. (2012). Personalizing mobile travel

information services. Procedia-Social and Behavioral

Sciences, 48:1195–1204.

Louviere, J. J., Street, D., Burgess, L., Wasi, N., Islam, T.,

and Marley, A. A. J. (2008). Modeling the choices

of individual decision-makers by combining efﬁcient

choice experiment designs with extra preference infor-

mation. Journal of Choice Modelling, 1(1):128–164.

Morris, M. D. (1991). Factorial sampling plans for pre-

liminary computational experiments. Technometrics,

33(2):161–174.

Nuzzolo, A. and Comi, A. (2016). Individual utility-based

path suggestions in transit trip planners. IET Intelli-

gent Transport Systems, 10(4):219–226.

OpenTripPlanner (2024). Multimodal trip planning. https:

//www.opentripplanner.org/. Accessed: Jan. 3, 2024.

Spitadakis, V. and Fostieri, M. (2012). Wisetrip-

international multimodal journey planning and deliv-

ery of personalized trip information. Procedia-Social

and Behavioral Sciences, 48:1294–1303.

Teodoro, G., Kurc¸, T. M., Taveira, L. F., Melo, A. C., Gao,

Y., Kong, J., and Saltz, J. H. (2017). Algorithm sen-

sitivity analysis and parameter tuning for tissue image

segmentation pipelines. Bioinformatics, 33(7):1064–

1072.

Wins, A., Barthelmes, L., Alpers, S., Becker, C., Kager-

bauer, M., and Oberweis, A. (2024). Personalized

day-trip planning: A TSP-TW-based multimodal mul-

ticriteria optimisation approach. Procedia Computer

Science (In Press).

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems

326