LEARNING SUSTAINABLE MOBILITY BEHAVIOUR IN POST-PANDEMIC VIENNA

Sustainable mobility behaviour is a difficult goal to reach, as people are not willing to change their habits simply because it might help the environment.
One method to change their mind is to create various incentives. Before doing this,
however, it is important to understand their behaviour. This paper is focused on
understanding people’s activity (e.g., next trip prediction, classification of their activity)
based on their recent trip data collected through a mobility app. The early experiments show that a classical method based on gradient boosting leads to better results that more state if the art deep learning methods for these tasks.

1. LEARNING SUSTAINABLE
MOBILITY BEHAVIOUR IN
POST-PANDEMIC VIENNA
Adrian M.P. Brașoveanu
Lyndon J.B. Nixon
Arno Scharl
ICITT2024

2. Introduction
ICITT2024
Project partners
1) Modul Technology
2) Ummadum
3) webLyzard
Technology
4) GeoSphere Austria
5) BOKU
6) DIO
Main idea:
Anticipate and change
mobility behaviour
while promoting
sustainability
First tasks:
Activity and Distance
prediction inside and
outside districts

3. Ummadum Dataset
ICITT2024
Ummadum
October 23 data
83k data points
All trips logged with
user consent,
GDPR (e.g., all trips
anonymzed for
analysis)
Important features
Activity (type,status)
Date (startat)
Distance
Duration
Rewards
(not displayed)

4. Data Split by District
Activities
BIKE
WALK
CARRIDER
CARDRIVER
PT – Public
Transport
Models
Transformer
XGBoost
Tasks: Activity prediction and Distance prediction
Early results were only marginally better than random
guesses. We decided to experiment with different splits.
Split by District (Bezirk)
23 in Vienna and 94 in Austria
Inside – Inner
Outside – Outer
Some districts bigger than others
ICITT2024

5. Adding Geographical Information
OSM
Various APIs:
Overpass
Nominatim
Nominatim
Docker Image
Handles Geo
names
Granular
information
Country
Region
District
City
ICITT2024

6. Activity Prediction - Inner/Outer
Task
Predict activity
(e.g., WALK,
BIKE)
Activity
BIKE
WALK
CARRIDER
CARDRIVER
PT
Models
Transformer
XGBoost
The differences between Accuracy and F1 for Transformers
- class imbalance
- high false positive or false negative rates
- misclassification costs
Model Dataset AccuracyF1
TransformerInner 0.46 0.40
Outer 0.76 0.68
XGBoost Inner 0.97 0.97
Outer 0.87 0.88
ICITT2024

7. Activity Prediction – Inner/Outer
Explaining Transformer Results
Split by District
Inside – Inner
Outside - Outer
Activity
BIKE
WALK
CARRIDER
CARDRIVER
PT – Public
Transport
Models
Transformer
XGBoost
Interpreting Transformer Results:
- Inside Districts (left) – attention is scattered through the
whole set of features; results are not great
- Outside Districts (right) – attention focused on a smaller
set of features; results are also better
ICITT2024

8. Activity Prediction – Inner
Explaining XGBoost Results
Split by District
Inside – Inner
Outside - Outer
Activity
BIKE
WALK
CARRIDER
CARDRIVER
PT
Models
Transformer
XGBoost
Inside districts the picture is more nuanced.
Distance (and duration) matters!
Rewards (upforsharing, upcommunitytype) matters!
Rest of the features not so much.
ICITT2024

9. Activity Prediction – Outer
Explaining XGBoost Results
Split by District
Inside – Inner
Outside - Outer
Activity
BIKE
WALK
CARRIDER
CARDRIVER
PT
Models
Transformer
XGBoost
When traveling outside districts things are more
straightforward!
reward_upcommunitytype seems to be the most
important feature when going outside the city.
ICITT2024

10. Distance Prediction
Task
Predict distance
Split by District
Inside – Inner
Outside - Outer
Models
Transformer
XGBoost
RMSE: lower is better.
Inside districts seems to be almost
solved!
Outside districts can still be improved.
Model Dataset RMSE
XGBoost Inner 56.96
Outer 2566.56
ICITT2024

11. Distance Prediction
Feature Importance
Split by District
Inside – Inner
Outside - Outer
Models
XGBoost
For the entire dataset (both inside and outside districts):
- reward_upcommunitytype is the best feature
- rest of the features have somewhat similar
contributions!
ICITT2024

12. Distance Prediction
Feature Importance - Inner/Outer
Split by District
Inside – Inner
Outside - Outer
Models
XGBoost
ICITT2024

13. Conclusion and Future Work
Tasks
Activity and
Distance
Prediction
Split by District
Inside – Inner
Outside - Outer
Models
XGBoost
Transformer
XGBoost is better
now, but
Transformers can
be seriously
improved
Adding Various Features

webLyzard topics

Weather features from GeoSphere

Date features (e.g., day of week)

Events (e.g., concerts)
Structured Prediction (multiple targets at once)

Activity

Distance

Duration
Using Different Models

Hybrid models (XGBoost Transformer)

Graph Neural Networks

Time Series LLMs
ICITT2024

14. Acknowledgments
THANK YOU!
https://aicentive.eu
/
District images from slide 3 from English and German Wikipedia published under
CC-BY-SA 3.0 license by Rosso Robot and Alexxw Ailura