Sustainable mobility behavior remains a challenging objective, as individuals rarely alter their transportation habits solely for environmental benefits. Creating efective incentives can prompt change, but first, it requires a comprehensive understanding of mobility patterns. This paper examines the prediction of personal travel activities (including next-trip forecasting and activity classification) utilizing recent journey data collected through a mobility application. A graph-based fusion ensemble, comprising a graph convolutional network and a statistical user pattern, is designed for structured prediction with multiple outputs, including origin, destination, transportation mode, and time. An explainable prediction pipeline is built on top of this ensemble. The results are then converted into a knowledge graph, enabling us to conduct sophisticated analyses and improve our weekly workflows.
Achieving sustainable transportation patterns is challenging since individuals rarely modify established routines solely for environmental benefits. Key considerations encompass personal convenience (overcrowded transit deterring usage), network development (separate cycling corridors), capability requirements, motivation systems (e.g., notifications or rewards, and inclusive design features. If individuals are to change their behavior, the proposed benefits of the motivation systems must exceed the perceived advantages of continuing current practices, including financial savings, comfort zones, and ingrained patterns.
The AI-CENTIVE project’s primary goal is to use neural network architectures to understand transportation behaviors in Austrian urban transit zones and identify the most efective reward structures to encourage a shift towards eco-friendly travel options, such as using bicycles or public transit instead of private vehicles for urban commuting. The research involves developing ensembles that combine statistical patterns and graph convolutional networks (GCNs). Our experiments quantify environmental impacts by combining baseline predictions with simulated incentives across mobility networks.
The main contribution presented in this paper is the semantic conversion workflow developed to support and analyze explainable AI predictions. The output of the ensembles is converted into a mobility knowledge graph, and the best predictions are analyzed using SPARQL queries.
The paper is organized as follows: Section 2 ofers a brief overview of related work, Section 3 describes the general method, while Section 4 presents the explainability pipeline and the associated analysis that can be performed with it. The paper concludes with a brief overview and future work.
This brief section covers only graph mobility network surveys and related work on explainability.
Examining graph-based computational methods, Jiang’s survey [
Beyond academic surveys identifying research directions, practical implementation insights from the
Google Maps arrival prediction system [
Understanding model decision processes remains crucial for discovering inherent biases often
stemming from unbalanced training samples. Schwalbe and Finzel [
A core challenge the AI-CENTIVE project addresses is the observation that individuals are often unwilling to change their mobility habits solely for environmental reasons. To overcome this, the project’s overarching mission is to develop AI-based incentivization techniques to influence citizens’ mobility choices, utilizing multimodal models of mobility activity and data analytics. The ultimate vision is to enable and incentivize Austrian citizens to adopt more sustainable mobility choices.
The work presented here builds upon the Ummadum platform which allows users to log their sustainable mobility activities via a mobile app. This process reports CO2 emissions reductions to raise users’ sustainability awareness and rewards sustainable behavior to incentivize environmentally friendly choices. The AI-CENTIVE community within Ummadum serves to engage participants, assess the efectiveness of incentives, conduct pilot tests for integrating AI-driven solutions, and plan rollout strategies across Austria. All trip data is collected with user consent, strictly adhering to GDPR best practices by anonymizing user identities and locations for analysis.
The project’s dataset for examining Vienna mobility patterns was provided by Ummadum. This commuter dataset comprises approximately 450,000 user trips collected between January 2024 and April 2025 from users of the Ummadum mobile app. The dataset focuses on activities such as biking, walking, public transport, and car sharing (as a rider or driver), also including data about activity status (e.g., cancelled, finished). To maintain GDPR compliance and route anonymity, user data and trip data are anonymized, with segments added or removed from trips. Locations for origins and destinations are expressed through zip codes, and information about location type (e.g., ofice, public venue) is also collected. The dataset also includes details on the incentives provided, such as points, rewards, carsharing, community types administering rewards, or activity challenges where users earn additional rewards based on their activity levels.
The project’s approach towards incentivization primarily involves rewards and notifications. Rewards include classic elements like: i) challenges; ii) lotteries, or iii) collection of points for sustainable activities. Notifications are displayed directly to the users based on their past history. Three types of notifications were implemented: i) success notifications to motivate participants based on their previous sustainable mobility actions; ii) AI recommendation notifications to suggest future sustainable mobility options using contextual information such as behavior, mobility preferences, location, and time; and iii) weather notifications for notable weather events. Explanations for the AI notifications are included, based on past user history. These notifications are part of a simple workflow that runs automatically each Monday morning, scheduling notifications after checking eligibility criteria (e.g., high confidence, no more than one notification per user per day).
Initial insights from the first 2025 pilot program indicated a total CO2 savings of 331 tons. These insights were used to refine a second pilot in 2025. This second pilot incorporates improvements such as explainable AI notifications (justifying notifications based on the user’s history, including similar trips or partial routes), weather alerts, and enhanced challenges within the rewards system. The overall goal of the second pilot was to test and refine AI-driven incentivization techniques to encourage sustainable mobility choices among Austrian citizens using the Ummadum platform.
We have examined several models designed for a structured prediction task that simultaneously
forecasts multiple outputs, including trip origin and destination locations, hour and day of the trip,
activity type, distance, and duration. These models, including Transformer and GCN ensembles that
include statistical models (e.g., ARIMA, XGBoost) are presented in a previous publication [
To enhance transparency and regulatory compliance of our Graph Neural Network mobility predictions, we developed a semantic conversion pipeline that transforms prediction outputs into structured RDF knowledge graphs. This approach addresses the requirements of the European AI Act for explainable AI systems in transportation applications.
Our system employs a multilayered semantic framework combining established ontologies with
custom ontologies 1. The core structure integrates PROV-O (W3C Provenance Ontology) [
For mobility-specific concepts, we defined custom RDF vocabularies in the modultech.eu/aicentive namespace, establishing classes such as TripPrediction, PredictionExplanation, and PatternEvidence. These vocabularies maintain compatibility with existing transport standards like GTFS (General Transit Feed Specification) while extending semantic representation for explainable AI applications. The explainability vocabulary captures structured reasoning through properties that include routeFrequency, timeFrequency, and activityFrequency, allowing systematic analysis of prediction confidence factors. Each prediction entity links to explanation instances that contain machine-readable pattern evidence and human-readable justification text.
The conversion pipeline automatically processes GNN outputs into RDF using rdflib, preserving all prediction metadata including confidence scores, historical pattern evidence, and explanatory reasoning. The resulting knowledge graph enables direct SPARQL querying without database infrastructure, 1The ontologies were designed for ofline use and are publicly available through https://github.com/modultechnology/aicentiveontologies GitHub supporting complex analytical queries for model evaluation and regulatory auditing. Semantic queries can systematically identify low-confidence predictions, analyze temporal patterns in model performance, and extract compliance reports for regulatory review. This queryable format facilitates automated detection of prediction biases and supports continuous model improvement through pattern analysis.
This semantic conversion enables sophisticated analytical capabilities through SPARQL querying that would be dificult or impossible with traditional CSV formats. Researchers can execute complex queries to identify high-confidence predictions for regulatory audit trails, systematically analyze low-confidence routes that require model improvement, and investigate user behavior patterns by correlating confidence scores with historical pattern evidence (see Table 1).
The semantic structure supports automated quality assurance through queries that filter predictions by confidence thresholds, group routes by performance metrics, and extract explanation metadata for transparency reporting. Beyond basic filtering, the pipeline enables advanced analytical patterns such as identifying users with consistently high prediction confidence, discovering temporal patterns in model performance, and analyzing the relationship between route familiarity and prediction accuracy. The explainability metadata becomes queryable, allowing researchers to systematically study which types of historical evidence correlate with prediction success and identify biases in model performance across diferent user groups or activity types. This semantic approach transforms static prediction output into a dynamic knowledge graph that supports continuous model evaluation, regulatory compliance reporting, and data-driven insights for model improvement. The standardized format also enables integration with other transportation data sets and supports collaborative research through interoperable semantic web technologies. Most importantly, queryable explanations provide the foundation for automated compliance checking required by emerging AI regulations, while simultaneously supporting research into the relationship between explainability quality and prediction accuracy.
One particular advantage of this approach is speed. It is generally 3x-5x times faster than the pure CSV approach. This is mainly due to the SPARQL engines being optimized for triple pattern matching than traditional SQL or pandas operations on CSV data. The processing eficiency is higher also due to SPARQL’s lazy evaluation mechanism (e.g., only process the data needed to answer specific queries) and selective loading (e.g., retrieving only data needed for each analysis). RDFlib itself is also optimized to handle larger datasets more eficiently through streaming.
Table 1 showcases several SPARQL queries that are routinely used for explainable mobility analysis. Table 2 presents the kind of automated analysis that can be done automatically on top of the evaluation results. All the results presented in this table refer to the last week of evaluation from our second pilot.
The semantic approach directly addresses EU AI Act mandates for transparency and auditability in AI applications. The structured explainability format enables automated compliance monitoring and systematic verification of AI decision-making processes. Beyond regulatory compliance, the standardized RDF format enhances research reproducibility and enables integration with broader smart city transportation planning systems, positioning our work at the intersection of responsible AI development and practical mobility applications.
The semantic conversation pipeline and the rapid analysis it enables represents one of the highlights of our projects. Due to examining the results through such tools, we were able to quickly improve our best model, and also deliver high quality notifications to our community members.
Future investigations will incorporate emissions calculations from the carbon assessment tool developed by project partner BOKU to evaluate environmental benefits across diferent reward mechanisms. This enables concrete quantification regarding pollution reduction achieved through altered journey selections. Additionally, data about weather conditions will enhance the ensemble accuracy.
Building widespread participation in eco-friendly journey choices remains equally crucial. Monitoring increased adoption rates of sustainable options helps evaluate progress in behavioral transformation. Furthermore, we intend to target problematic urban locations lacking vehicle storage facilities or convenient transit access within a reasonable walking range. Quantifying improvements within these zones demonstrates enhanced accessibility plus livability improvements. Such enhanced measurements strengthen our predictive capabilities while supporting inclusive, evidence-based approaches toward sustainable transportation development. Through our project initiatives, we aim to empower community members and organizational stakeholders in creating more responsive, equitable, and environmentally responsible urban transit zones.
The AI-CENTIVE project (FFG G.A. 892238) was funded within the RTI Initiative of the ICT of the Future “IKT der Zukunft” – an initiative of the Federal Ministry for Climate Protection, Environment, Energy, Mobility, Innovation and Technology (BMK). 6.1. Declaration on Generative AI During the preparation of this work, the authors have used Grammarly for grammar and spelling checks.