When Should We Report the Trafﬁc Jams of Today? A Case Study on a

Swiss Highway Using Graph Neural Networks and Expert Knowledge

Jhonny Pincay

1,2 a

, Ana O

3,4 b

, and Damian Nomura

Viasuisse AG, Zentralstrasse 115, Biel, Switzerland

Pontiﬁcia Universidad Cat

olica del Ecuador, Avenida 12 de Octubre 1076 y Roca, Quito, Ecuador

Swiss Paraplegic Research, Guido A. Z

ach, Nottwil, Switzerland

Department of Health Sciences and Medicine, University of Lucerne, Lucerne, Switzerland

Keywords:

Speed Prediction, Graph Neural Networks, Transportation, Smart Logistics, PyTorch Geometric.

Abstract:

This case study manuscript details the conception and implementation of an artifact that uses ﬂoating car

data to forecast average speeds on a segment of a Swiss national road. To consider the spatial and temporal

dependencies when performing the predictions, the studied segment was modeled as a graph and as a time

series problem. Subsequently, to obtain a prediction model, the data collected over a month and augmented

to simulate the behavior during summer were used as the input to train a Graph Neural Network. After the

evaluation of the results it was concluded that despite the considerable differences between the forecasted

values and the reality, it was possible to perform such an implementation with limited data and resources.

Moreover, a handful of trafﬁc reporters still considered the results appropriate, and suitable.

1 INTRODUCTION

The deﬁnition of trafﬁc congestion has different

shades. It can be explained in terms of demand capac-

ity, travel time delay, and cost (Aftabuzzaman, 2007).

With demand capacity it is understood that travel de-

mand exceeds the capacity of a road; from the per-

spective of travel time delay, trafﬁc congestion means

that the time needed to reach a destination is higher

than the incurred under free-ﬂow travel conditions;

lastly, from the cost-related perspective, it is implied

that the actual need of resources to go from one place

to another is incremented. Regardless of the deﬁni-

tion that one adopts, one thing they have in common is

their consequences: a considerable increase in travel

time and required resources.

Nowadays, trafﬁc analysis and forecasting are a

very relevant topic since it impacts not only citizens

commuting or going on holiday but also the logistic

planning of services. For companies, estimating the

effects of trafﬁc congestion on their supply chain has

become critical to keep their operational costs low and

to ensure the ever-growing demand for just-in-time

https://orcid.org/0000-0003-2045-8820

https://orcid.org/0000-0002-7428-4574

delivery Pincay et al. (2020). Such facts have driven

a growing interest in the ﬁeld of trafﬁc modeling not

only from researchers but also from practitioners.

Furthermore, the availability of more data as well

as the development of methods to process larger

amounts of information have eased the development

of systems able to provide travel time information to

commuters and carriers. Usual data sources used for

trafﬁc analysis include sensors (e.g., trafﬁc counters

and loop detectors), on-site collected data, and ﬂoat-

ing car data (FCD) recorded from global positioning

systems (GPS) devices (Mori et al., 2015; Zhou et al.,

2012).

The usage of FCD for trafﬁc and travel time mod-

eling has been steadily increasing since it is a large

and affordable source of information. The methods

used to analyze and process FCD found in the liter-

ature include machine learning, fuzzy-based, proba-

bilistic, and deep learning-based methods. Despite

the satisfactory results that they offer, they require

vast amounts of data and the availability of enough

computational resources (Pincay et al., 2020; Zhu

et al., 2020).

This manuscript presents the method and out-

comes of a case study whose goal is to obtain a daily

prediction of speed on a road segment of one of the

150

Pincay, J., Oña, A. and Nomura, D.

When Should We Report the Trafﬁc Jams of Today? A Case Study on a Swiss Highway Using Graph Neural Networks and Expert Knowledge.

DOI: 10.5220/0011846900003473

In Proceedings of the 9th International Conference on Geographical Information Systems Theory, Applications and Management (GISTAM 2023), pages 150-158

ISBN: 978-989-758-649-1; ISSN: 2184-500X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

main highways of Switzerland. To that end, an artifact

following the principles of design science research

in conjunction with a transdisciplinary approach was

conceptualized. For the implementation of the arti-

fact, FCD provided by the Here Developer platform

was processed and methods based on a Graph Neural

Network (GNNs) were applied. Furthermore, trafﬁc

experts were included in all the stages of the develop-

ment of the project.

This article is structured as follows: Section 2 in-

troduces the concepts and related works on which this

initiative is grounded. Then, the methods followed in

the design of the artifact are described in Section 3.

Section 4 presents the results of the implementation

of the project. Lastly, Section 5 closes the curtains of

this research effort with a summary and concluding

remarks.

2 THEORETICAL BACKGROUND

This section introduces the theories and concepts used

for the foundation of this research effort. Related

work is also examined.

2.1 Trafﬁc and Speed Prediction

According to Lin et al. (2005), the main components

of a road trafﬁc environment are humans, vehicles,

and facilities (e.g., signaling, roads and streets). In

this context, the facilities constitute the supply, and

humans and vehicles are the trafﬁc demand.

In regards to road trafﬁc, it can be classiﬁed into

two states: congested/jammed and uncongested/free

ﬂow (Treiber and Kesting, 2013). There are several

trafﬁc characteristics or variables that enable identi-

fying the trafﬁc in any of these two states. These

variables are known as trafﬁc state variables, being

the most relevant the ﬂow, vehicle density, and speed

(Pincay et al., 2020).

Trafﬁc and speed forecasting has been studied

mainly through knowledge-driven and data-driven

approaches. Knowledge-driven methods implement

queueing theory and perform simulations about the

user behavior in trafﬁc; data-driven initiatives focus

on the study of time series to enable the implementa-

tion of Auto-Regressive Integrated Moving Average

(ARIMA) models for instance (Li et al., 2017).

Although ARIMA models remain popular they

work only under stationary situations, which in traf-

ﬁc, is not always true. Thus, most recent efforts are

directed towards designing and implementing neural

https://developer.here.com/

network and deep learning-based methods, as a way

of considering temporal and spatial structures in the

deﬁnition of more reliable and accurate trafﬁc predic-

tion models.

2.2 Artiﬁcial Neural Networks and

Deep Learning

Artiﬁcial Neural Networks (ANN) or simply neural

networks were conceived as a way of taking to the

computational world how the human brain works: nu-

merous neurons are interconnected and together they

process information (Wang, 2003). They had proven

to be useful and efﬁcient to solve tasks where per-

forming inferences from previous data are required.

Current applications of ANNs include image segmen-

tation, pattern recognition, face recognition, and pre-

diction tasks, among others (Abiodun et al., 2018).

A special type of ANN is the deep neural network.

The main difference with traditional ANNs is that

deep neural networks are composed of a large num-

ber of layers. Current research is directed towards

extending deep learning methods with approaches

based on graph data aiming to consider the interac-

tion between agents in the deﬁnition of prediction

models. Such networks are known as Graph Neural

Networks (GNNs) and their study is called geometric

deep learning (Wu et al., 2020). There is a plethora

of architecture deﬁnitions of GNNs; authors Wu et al.

(2020) proposed a taxonomy to classify them: Recur-

rent Graph Neural Networks (RecGNNs), Convolu-

tional Graph Neural Networks (ConvGNNs), Graph

Autoencoders (GAEs), and Spatial-Temporal Graph

Neural Networks (STGNNs).

Circulation on roads can be represented as a

graph. For instance, locations along the way consti-

tute the nodes and the weight of the edges correspond

to the distance between places. The assumption that

the circulation in a certain location at a certain time

inﬂuences its neighboring places can also be made.

Under this deﬁnition, the problem of forecasting the

average speed over time based on GNNs is possi-

ble. The speciﬁc type of neural network that enables

such reasoning is named Temporal Graph Convolu-

tional Network (T-GCN). They capture both spatial

and temporal dependencies among time series. This

characteristic makes this type of ANN suitable for a

broad range of spatiotemporal forecasting tasks (Bai

et al., 2021).

2.3 Related Works

Previous publications dealing with the task of predict-

ing vehicle speeds on the roads through GNNs are

When Should We Report the Trafﬁc Jams of Today? A Case Study on a Swiss Highway Using Graph Neural Networks and Expert

Knowledge

151

presented in this section.

Guo et al. (2019) proposed an attention-based

spatial-temporal GNN to capture the dynamic spatio

and temporal properties of trafﬁc data simultaneously.

According to the reported results, the method showed

a superior accuracy than others used for trafﬁc-related

problems such as the ARIMA and Long short-term

memory (LSTM). However, their solution was yet to

be proved in larger-scale experiments.

Researchers Yu et al. (2017) proposed a deep

learning framework implemented through the training

of Spatio-Temporal Graph Convolutional Networks

(STGCN). After conducting simulations on two real-

life datasets they found that formulating the problem

of trafﬁc and average speed forecasting as a graph

and convolutional structures enabled faster training

with fewer parameters than other state-of-the-art ap-

proaches. Nevertheless, their framework needed fur-

ther optimizations in its network structure to enable

its application in large-scale industries.

Another initiative is by Li et al. (2017). They at-

tempted to incorporate the spatial and temporal de-

pendencies in trafﬁc ﬂow using Diffusion Convo-

lutional Recurrent Neural Network (DCRNN). The

studied segments were modeled as a graph, where the

nodes corresponded to the speed measuring stations

and the weight of the edges was the distance between

the nodes. The diffusional nature architecture of the

neural network enabled us to consider the effect of

trafﬁc over space and time to improve the results of

the predicted speeds.

Lastly, Bai et al. (2021) proposed to use an Atten-

tion Temporal Graph Convolutional Network (A3T-

GCN) as a means to model the short-time trend in

time series by using Gated Recurrent Units (GRU)

and a graph convolutional network to consider the

spatial dependencies according to the topology of the

road network. An attention mechanism was intro-

duced to ﬁne-tune the importance of different time

points to improve prediction accuracy. The experi-

mental results in real-world datasets demonstrated the

effectiveness and robustness of the proposed method.

In contrast to the precedent work, this research

effort presents the results of a transdisciplinary case

study of speed predictions for trafﬁc using ﬂoating

car data collected over a month from one segment

of one of the most important highways of Switzer-

land, augmented through statistical methods to sim-

ulate the behavior during summer, and used to train

a GNN network implemented in Python. The goal

is to demonstrate how adequate results can be ob-

tained with limited computational resources, with rel-

atively low amounts of data, and by incorporating ex-

pert knowledge in the design of the solution.

3 METHODOLOGY AND USE

CASE

The development of this case study were guided by

the principles of the design science research for infor-

mation systems methodology with a transdisciplinary

approach (i.e., incorporating practical experiences

into the solution process (Hadorn et al., 2008)). This

research methodology was selected because its ap-

plication entails the development of an artifact while

extending existing knowledge (Hevner and Chatter-

jee, 2010) and since this project was realized in a

conjoined effort between academia and an industrial

practitioner namely Viasuisse AG

Five main phases encompassed the execution of

this study: i) preliminary analysis, ii) data collection

and augmentation; iii) training; iv) evaluation; and,

v) visualization. Figure 1 depicts these phases and

intermediate steps.

Figure 1: Methodology followed in the research project.

3.1 Preliminary Analysis

Selecting a suitable data source for the prediction

tasks is crucial. Before starting the development of

this project, different data sources that record infor-

mation about trafﬁc ﬂow in Switzerland were evalu-

ated:

• Real-Time Data from Road Trafﬁc Counters: The

open data platform for the mobility of Switzer-

land

offers real-time data from trafﬁc sensors de-

ployed through the main roads and highways of

the country. The recorded data corresponds to the

https://viasuisse.ch/

https://opentransportdata.swiss/en/

GISTAM 2023 - 9th International Conference on Geographical Information Systems Theory, Applications and Management

152

average speed of vehicles that passed near the lo-

cation of the sensor in the previous minute.

Aiming to evaluate the quality of the retrieved

data, points of interest from the entire network

were selected based on the location of trafﬁc cam-

eras. Access to those cameras was granted by our

partner. After performing several observations of

what was happening and reality and the data that

was being retrieved, the researchers came to the

conclusion that although this source offers data

with ﬁne granularity and is well documented, the

results were not reliable enough to deﬁne a train-

ing set that meets the goals of this case study.

• Trafﬁc Message Channel-Based Records: Trafﬁc

messages delivered through the Trafﬁc Message

Channel (TMC) technology (Gao and Wen, 2007)

and processed by our partner company were also

studied. Such messages record a variety of inci-

dents that may cause trafﬁc anomalies (e.g., traf-

ﬁc congestion, accidents, and road works) and

they are reported by trafﬁc monitoring responsi-

ble (e.g., road police and municipalities).

This data source however recorded the duration

of trafﬁc anomalies and not the speed of the cir-

culating vehicles, and thus, it was not suitable to

reach the objectives of this work. Nevertheless,

the TMC technology encodes points of interest

over the highways in the form of a location code

list expressed in the standard ISO 14819-3 (Arco

et al., 2017). For instance, the A1 highway is com-

posed of 151 locations (as per version 7 of the

Swiss location code table) and each of them has

a unique identiﬁer (e.g., the location Niederbipp

is identiﬁed by the location code 10256). The ex-

istence of such location codes helped in the def-

inition of points of interest and to understand the

trafﬁc ﬂow over the Swiss highways.

• Floating Car Data from Here WeGo: Here

WeGo

is a navigation service operated by HERE

Technologies that offers among others trafﬁc and

location services. Through an API for develop-

ers, HERE offers trafﬁc ﬂow data aggregated by

location; this data contains information about the

average speed and the tendency of trafﬁc jams.

Moreover, some tests were conducted to estimate

the quality of this data source. The tests consisted

of traversing a segment of interest and comparing

what was happening while in a car and what the

data being retrieved was showing. It was deter-

mined then that the average speed provided by the

API was close to reality. Considering this fact and

https://wego.here.com/

also that the speed can be aggregated by TMC lo-

cation codes, the authors decided to use this data

source to build our speed prediction artifact.

3.2 Data Collection and Augmentation

With the selection of the data source, it was possible

to proceed with the data collection.

Firstly, a segment of interest was deﬁned. To that

effect, discussions with trafﬁc reporters (part of our

partner company) were held. The goal of deﬁning a

segment of interest in this way is to later use the ex-

pertise of the reporters to corroborate the results of the

speed predictions. The selected segment was part of

the A1 highway, which traditionally is the one with

the highest amount of trafﬁc hours every year; the

segment had a length of 50Km and encompassed 24

locations from Verzweigung H

arkingen (TMC Loca-

tion code 10041) until W

urenlos (TMC Location code

11211).

A script written in the programming language

Python was deployed on a server to retrieve the trafﬁc

ﬂow data. The data was retrieved every 5 minutes dur-

ing May of 2022 and for 30 days. The extracted data

contained among other ﬁelds the timestamp, TMC lo-

cation code, location name, and average speed.

Furthermore, since one month of data was not

enough, the collected data was augmented to simu-

late a database of readings for the whole summer. To

that end, the average of the speeds grouped by times-

tamp, location, and direction was computed and used

to simulate the values for June, July, and August. The

values were selected from a uniform distribution on

the interval from a minimum value of 93% and a max-

imum 103% of the observed average speed.

3.3 Training

Once the data was available, the training stage took

place. The goal of the training is to predict the future

trafﬁc average speed considering previously observed

trafﬁc ﬂow from N correlated observed locations on

the selected highway segment. In a similar manner,

as proposed by Li et al. (2017), the monitored loca-

tions can be represented as a weighted directed graph

G = (V , E, W )) where V corresponds to the set of

nodes (the observed locations), E denotes the set of

edges, and W ∈ R

N×N

corresponds to a weighted ad-

jacency matrix denoting the proximity of the nodes.

Furthermore, the training stage consisted of two

main steps: Neural network setup & Implementation,

and Training & ﬁne-tuning of parameters. Details are

presented below:

When Should We Report the Trafﬁc Jams of Today? A Case Study on a Swiss Highway Using Graph Neural Networks and Expert

Knowledge

153

Figure 2: Architecture of the A3T-GCN Network. Adapted from Bai et al. (2021).

3.3.1 Neural Network Setup and

Implementation

When modeling the trafﬁc ﬂow with GNNs, the prob-

lem takes the shape of a diffusion process over a di-

rected graph. This abstraction enables the capturing

of the stochastic nature of trafﬁc dynamics (Li et al.,

2017; Cowan and Jonard, 2004). Thus, the locations

of the road segment are modeled into a graph network,

where the trafﬁc state on the different locations is de-

picted as the node attributes (i.e., the average speed at

a certain time and day).

The A3T-GCN architecture proposed by Bai et al.

(2021) was adapted to deﬁne the structure of the

model to perform the speed forecasting task. The rea-

son behind this choice was the good results they pre-

sented in their study and related literature and since

our dataset had a similar structure. The characteris-

tics of the A3T-CGN model are described as follows:

• The A3T-GCN is based on the T-GCN but uses

attention.

• For the spatial aggregation a GCN structure is

used. For the temporal aggregation a GRU.

• Historical time series are inputted to the T-GCN

model, then n hidden states h are obtained. These

hidden states cover spatiotemporal characteristics.

• The hidden states are fed into the attention model

to determine a context vector that covers the trafﬁc

variation information: a

t−n

, ..., a

t−1

, a

• Final results are computed using a fully connected

layer.

Figure 2 illustrates the architecture of the A3T-

GCN model applied.

3.3.2 Training and Fine-Tuning

The original dataset was split into 80% for the train-

ing process and 20% for testing purposes. The hyper-

parameters to deﬁne included the learning rate (lr),

epochs, and the number of hidden units.

The Python implemented neural network was

trained using the Adam optimizer with learning rate

annealing. The hyper-parameters were chosen using

the Tree-structured Parzen Estimator (TPE) on the test

set (Bergstra et al., 2011).

The data sample for the training consisted of 24

nodes; each node containing 2 features (speed and

time). Each bucket of data contained 12 timesteps

(12 × 5 min = 60 min). The edges attributes were

deﬁned on the distances between the locations and a

threshold. This threshold was deﬁned with the value

of 4000 (k = 4000), meaning that the effects of low

speeds on a location will have an effect up to the loca-

tions in the following 4000 meters. This distance was

deﬁned after holding meetings and discussions with

trafﬁc experts. Lastly, the neural network was trained

in 30 epochs.

3.4 Evaluation

The evaluation was conducted using 20% of the avail-

able data. The loss and Mean Squared Error (MSE)

were used to deﬁne how accurate the trained model

was. Moreover, the prediction for a whole day was

obtained and the predicted average speed of several

locations was compared to what really happened, this

helped us to have a closer-to-reality appreciation of

the results.

The results were then shared with the trafﬁc ex-

perts and contrasted with their day-to-day insights.

3.5 Visualization

The results of the daily predictions were shown

through a web heatmap. Moreover, some insights

about the daily predictions are also depicted.

4 RESULTS

This section presents the artifact’s implementation re-

sults built upon the methods explained in Section 3.

A script written in Python was used to perform

the data collection and another one was written in R

GISTAM 2023 - 9th International Conference on Geographical Information Systems Theory, Applications and Management

154

Figure 3: Predicted and real speeds for the locations of K

olliken S

ud (top) and Verzweigung H

arkingen (bottom) for the 1st.

of September 2022 between 10H00 and 18H00.

to augment and simulate the data. The neural net-

work model was implemented with the deep learning

framework PyTorch Geometric Temporal. Intro-

duced by Rozemberczki et al. (2021), this framework

was implemented with the existing libraries of the

PyTorch ecosystem, temporal snapshot generators for

batching, and streamlined neural network layer deﬁ-

nitions. The library Folium was used to create the

heatmap presenting the results.

4.0.1 Dataset Conformation

The data source used to conform the dataset was the

FCD obtained from HERE; moreover, the data were

collected during May 2022 and augmented to simu-

late readings for the whole summer of 2022 (until Au-

gust 2022). Furthermore, only the average speeds in

the direction west to east were taken into considera-

tion.

The ﬁnal dataset used for the training and the test-

ing was composed of approximately 211 800 records,

which registered the timestamp, TMC location name,

code, and average speed at the moment of the read-

ing. To construct the sensor graph, the distance be-

tween the 24 studied locations was computed and

used to build an adjacency matrix applying a thresh-

olded Gaussian kernel (Shuman et al., 2013).

4.1 Training and Validation

The training was conducted on a Windows computer

with 32 GB of RAM memory and with an Intel (R)

Core i7 processor.

A learning rate lr = 0.001 and 30 training epochs

were set for the training process. During the ﬁrst

training epoch, an MSE of 1.0497 was obtained; after

all the training epochs were executed, the MSE was

0.327 and 0.2947 on the testing set.

The speed forecast for the 1st. of September

2022 between 10H00 and 18H00 was later obtained.

These values were used to compare with the actual

records of the day. Two locations were selected to

that end: K

olliken S

ud und Verzweigung H

arkingen.

These locations were selected after consulting with

the trafﬁc professionals; the location K

olliken S

ud is a

rather straight sub-segment whereas the Verzweigung

arkingen is rather a complex road junction where

multiple roads meet.

When Should We Report the Trafﬁc Jams of Today? A Case Study on a Swiss Highway Using Graph Neural Networks and Expert

Knowledge

155

Figure 4: Example of the visualization results for the trafﬁc prediction on the studied segment; besides a heatmap depicting

the low speeds, some stats are also provided.

These results are depicted in Figure 3; for

olliken S

ud (see Fig. 3 top), the predictions follow

very well the tendency of the real values, which can

be interpreted as encouraging results. On the other

hand, in the outcome for the Verzweigung H

arkingen

(see Fig. 3 bottom), the differences are more evident.

To get an expert opinion about the obtained pre-

dictions, discussions with trafﬁc experts working at

our partner company were held. From their point of

view, the results are promising and they reﬂect what

they observe on a regular Thursday. Moreover, for

their daily activities, when they have to report trafﬁc

incidents they always have to verify on-road cameras

and decide whether an incident is going to be solved

soon or if it is going to last for a longer period. “Hav-

ing such daily predictions will help us to make our job

more effectively” manifested one of the participants

in the discussion; “Even if the predicted speeds differ

from the real values, having an indicator of an ap-

proximate time when a sudden reduction could hap-

pen is also useful” manifested another.

In light of these results, one can say that despite

the 0.2947% MSE on the validation set seeming high

at ﬁrst glance, the predicted speed values can still be

practical when assisting trafﬁc reporters to report in-

cidents in a faster and more agile manner.

4.1.1 Visualization

A dashboard-like web interface was implemented to

present the results of the speed predictions. The in-

terface had two main parts: a heatmap and some

indicators. The heatmap had the goal of depicting

the locations with the lower speeds during the day

whereas the indicators were designed to provide a bet-

ter overview of the locations where the low speeds

were predicted and the times when this was happen-

ing. Figure 4 presents an example of the described.

Moreover, the prototype was intended to help trafﬁc

reporters in their duties and not drivers on the road.

Therefore, a richer interface was preferred instead of

a simpliﬁed one.

5 SUMMARY AND

CONCLUSIONS

Nowadays, convenience is the currency to buy users’

attention and thus decides whether a service remains

relevant or not. As Kano Noriaki stated, over time de-

lightful innovation becomes another basic need. As a

result, the expectations of users grow accordingly to

the technical standards set by the industry. Providers

of trafﬁc information itself are facing demands for

more and more precise data to route from A to B.

It has not been long ago that general information

about major trafﬁc jams on national highways broad-

casted in 30-minute intervals over radio channels was

enough.

This transdisciplinary research/applied project

presents the results of a case study that attempts

to demonstrate how speed predictions on highways

can be obtained without a large amount of data and

GISTAM 2023 - 9th International Conference on Geographical Information Systems Theory, Applications and Management

156

with limited resources but by exploiting deep learning

methods and including experts’ knowledge in the de-

velopment. This project was completed in ﬁve stages:

i) preliminary analysis; ii) data collection and aug-

mentation; iii) training; iv) evaluation; and iv) visu-

alization of results. The preliminary analysis allowed

us to identify the most reliable data source to predict

speed on the roads; a collection and cleansing process

was conducted and an augmentation of the data took

place afterward. The overall dataset was conformed

of average speed readings every 5 minutes from May

to August of 2022 of a segment on the A1 motorway

in Switzerland. A training process of a neural network

architecture based on GNNs took place then followed

by an evaluation process based on different metrics

and experts’ opinions. Lastly, the speed predictions

were depicted on a heatmap through a web interface.

The methodology applied and the usage of GNN

models proved to be suitable for the task of forecast-

ing speeds. Although there is still plenty of room

for improvement, it was shown that with a relatively

small amount of data and limited computer resources,

it is still possible to obtain predictions. As expressed

by the trafﬁc experts that analyzed the results pro-

vided, even when these are not as accurate as one

could wish they are still useful to ease the task of

trafﬁc reporting. Aspects such as a gradual or sud-

den change of speed might be an indicator that a

trafﬁc perturbation will occur. Regardless of the re-

lated work, this research project is characterized by its

practical nature, transdisciplinarity, and the real data

used for the implementation. The methods presented

could guide similar studies in which the involvement

of people is key to solving a rather complex problem.

The pressure of meeting expectations multiply in

the logistics branch by the factor of ever-growing

pressure on the prices and thus on the cost a product

is allowed to cause. The supply chain of commodi-

ties has been optimized to a very high extent. Goods

nowadays usually do not rest for a very long time at

a storage facility since storing creates cost. Capac-

ities of trucks and containers are being maxed out

in order to waste as little as possible. At the same

time, the goods need to be delivered just in time, in

order to not cause delays, which are again cost rele-

vant. The demands from users directly, as from in-

dustrial stakeholders, i.e. logistic suppliers, exceed

the capabilities pure and accessible data can provide.

In a hypothetical situation: Even if all the data avail-

able would be accessible for one solution that could

perform the analysis in order to provide such infor-

mation, it would consume way too much energy and

storage space. Ecological ethics would heavily be vi-

olated by predictions for mainly fuel-based vehicles.

From a sustainable business perspective, we are

urged to perform with the same efﬁciency as the logis-

tic suppliers with their trucks and containers, in terms

of using data. We must ﬁnd ways to make the most

sense out of data in order not to overproduce sense-

lessly but still provide precise outcomes. The results

of this study might still not have used the most eco-

logically sustainable of methods but it is a ﬁrst step

taken in the right direction. At the same time, it is

a clear indicator, if not proof, that transdisciplinary

research allows for ﬁnding viable solutions more efﬁ-

ciently. We must be aware that business and science

are not two separate silos, but act as chain links. Sci-

ence has no value to humans if it is not made acces-

sible to the greater public. This happens by forming

products. At the very same time, business needs real

innovation in order to prosper. This study shows that

joint forces of both areas are the future of business as

of science.

Future work will focus on gathering more data to

capture seasonality and to improve the models with

the further tuning of parameters and training cycles.

other aspects to be addressed are the inclusion of un-

expected events (e.g., accidents and the presence of

objects or animals on the road) to adjust the predic-

tions. Lastly, the visualization dashboard must be also

improved and enriched with maybe the inclusion of

linguistic summaries (Pincay et al., 2021) as a means

of addressing the uncertainty and imprecision of the

predicted values.

ACKNOWLEDGEMENTS

The authors would like to thank the director of Via-

suisse AG Marc Br

onnimann for supporting this re-

search project as well as Marco De Carli, Michael

Krein, Barbara Roelli, Karin Sch

ar, and Chico Koch

for contributing with their trafﬁc expertise in the def-

inition of aspects of the methodology and evaluation

of the results.

This project has received funding from the Euro-

pean Union’s Horizon 2020 research and innovation

program under the Marie Sklodowska-Curie grant

agreement No 801076, through the SSPH+ Global

Ph.D. Fellowship Program in Public Health Sciences

(GlobalP3HS) of the Swiss School of Public Health.

REFERENCES

Abiodun, O. I., Jantan, A., Omolara, A. E., Dada, K. V.,

Mohamed, N. A., and Arshad, H. (2018). State-of-the-

When Should We Report the Trafﬁc Jams of Today? A Case Study on a Swiss Highway Using Graph Neural Networks and Expert

Knowledge

157

art in artiﬁcial neural network applications: A survey.

Heliyon, 4(11):e00938.

Aftabuzzaman, M. (2007). Measuring trafﬁc congestion-

a critical review. In 30th Australasian Transport Re-

search Forum.

Arco, E., Ajmar, A., Arneodo, F., and Boccardo, P. (2017).

An operational framework to integrate trafﬁc message

channel (tmc)in emergency mapping services (ems).

European Journal of Remote Sensing, 50(1):478–495.

Bai, J., Zhu, J., Song, Y., Zhao, L., Hou, Z., Du, R., and Li,

H. (2021). A3t-gcn: Attention temporal graph convo-

lutional network for trafﬁc forecasting. ISPRS Inter-

national Journal of Geo-Information, 10(7):485.

Bergstra, J., Bardenet, R., Bengio, Y., and K

egl, B. (2011).

Algorithms for hyper-parameter optimization. Ad-

vances in neural information processing systems, 24.

Cowan, R. and Jonard, N. (2004). Network structure and

the diffusion of knowledge. Journal of economic Dy-

namics and Control, 28(8):1557–1575.

Gao, Y. and Wen, H.-m. (2007). Technique and standard-

ization research of radio data system-trafﬁc message

channel (rds-tmc). Journal of Transportation Systems

Engineering and Information Technology, 3.

Guo, S., Lin, Y., Feng, N., Song, C., and Wan, H. (2019).

Attention based spatial-temporal graph convolutional

networks for trafﬁc ﬂow forecasting. In Proceedings

of the AAAI conference on artiﬁcial intelligence, vol-

ume 33, pages 922–929.

Hadorn, G. H., Biber-Klemm, S., Grossenbacher-Mansuy,

W., Hoffmann-Riem, H., Joye, D., Pohl, C., Wies-

mann, U., and Zemp, E. (2008). The emergence of

transdisciplinarity as a form of research. In Handbook

of transdisciplinary research, pages 19–39. Springer.

Hevner, A. and Chatterjee, S. (2010). Design research in

information systems: theory and practice, volume 22.

Springer Science & Business Media.

Li, Y., Yu, R., Shahabi, C., and Liu, Y. (2017). Diffusion

convolutional recurrent neural network: Data-driven

trafﬁc forecasting. arXiv preprint arXiv:1707.01926.

Lin, H.-E., Zito, R., Taylor, M., et al. (2005). A review

of travel-time prediction in transport and logistics. In

Proceedings of the Eastern Asia Society for trans-

portation studies, volume 5, pages 1433–1448.

Mori, U., Mendiburu, A.,

Alvarez, M., and Lozano, J. A.

(2015). A review of travel time estimation and fore-

casting for advanced traveller information systems.

Transportmetrica A: Transport Science, 11(2):119–

157.

Pincay, J., Mensah, A. O., Portmann, E., and Ter

an, L.

(2020). Forecasting travel times with space partition-

ing methods. In Grueau, C., Laurini, R., and Ragia,

L., editors, Proceedings of the 6th International Con-

ference on Geographical Information Systems The-

ory, Applications and Management, GISTAM 2020,

Prague, Czech Republic, May 7-9, 2020, pages 151–

159. SCITEPRESS.

Pincay, J., Portmann, E., and Ter

an, L. (2021). Mining

linguistic summaries in trafﬁc. In B

ack, T., Wag-

ner, C., Garibaldi, J. M., Lam, H. K., Cottrell, M.,

Merelo, J. J., and Warwick, K., editors, Proceedings

of the 13th International Joint Conference on Compu-

tational Intelligence, IJCCI 2021, Online Streaming,

October 25-27, 2021, pages 169–176. SCITEPRESS.

Rozemberczki, B., Scherer, P., He, Y., Panagopoulos, G.,

Riedel, A., Astefanoaei, M., Kiss, O., Beres, F.,

opez, G., Collignon, N., et al. (2021). Pytorch ge-

ometric temporal: Spatiotemporal signal processing

with neural machine learning models. In Proceedings

of the 30th ACM International Conference on Infor-

mation & Knowledge Management, pages 4564–4573.

Shuman, D. I., Narang, S. K., Frossard, P., Ortega, A., and

Vandergheynst, P. (2013). The emerging ﬁeld of signal

processing on graphs: Extending high-dimensional

data analysis to networks and other irregular domains.

IEEE signal processing magazine, 30(3):83–98.

Treiber, M. and Kesting, A. (2013). Trajectory and ﬂoating-

car data. In Trafﬁc Flow Dynamics, pages 7–12.

Springer.

Wang, S.-C. (2003). Artiﬁcial Neural Network, pages 81–

100. Springer US, Boston, MA.

Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., and Philip,

S. Y. (2020). A comprehensive survey on graph neural

networks. IEEE transactions on neural networks and

learning systems, 32(1):4–24.

Yu, B., Yin, H., and Zhu, Z. (2017). Spatio-temporal

graph convolutional networks: A deep learning

framework for trafﬁc forecasting. arXiv preprint

arXiv:1709.04875.

Zhou, P., Zheng, Y., and Li, M. (2012). How long to wait?:

predicting bus arrival time with mobile phone based

participatory sensing.

Zhu, H., Xie, Y., He, W., Sun, C., Zhu, K., Zhou, G.,

and Ma, N. (2020). A novel trafﬁc ﬂow forecasting

method based on rnn-gcn and brb. Journal of Ad-

vanced Transportation, 2020.

GISTAM 2023 - 9th International Conference on Geographical Information Systems Theory, Applications and Management

158