Presenters

Prof. Massimo Stella. Prof. Massimo Stella is Associate Professor and PI of CogNosco Lab, Department of Psychology and Cognitive Science, University of Trento. He is the PI of the FIS2 PENSO project on the cognitive repercussions of LLMs. His research develops cognitive network science, computational social science,and AI psychology methods to study language, emotions, perceptions, and human-AI interaction. In this tutorial he will lead the conceptual framing, methodological positioning, and discussion of responsible interpretation.

Dr. Ali Aghazadeh Ardebili. Dr. Ali Aghazadeh Ardebili is Senior Researcher at CogNosco Lab, Department of Psychology and Cognitive Science, University of Trento. His work focuses on LLM-generated cognitive digital shadows, persona-based simulation, data engineering pipelines, and pooling tools for synthetic corpora. In this tutorial he will lead the technical demonstration on prompt-to-data pipelines, wrangling, validation, and dashboard-based exploration.

Aims and Learning Objectives

The tutorial aims to provide a compact, practical introduction to mining and pooling synthetic cognitive data generated by large language models under controlled human-shadow and AI-assistant prompting conditions. By the end of the session, participants will be able to:

• define cognitive digital shadows as structured, observational traces of LLM behaviour, distinct from digital twins and from claims about real human populations;
• design persona-aware and AI-self prompting schemas that preserve metadata, task context, and machine-readable outputs;
• wrangle raw JSON artifacts into analysis-ready tables, including schema checks, persona consistency, categorical encoding, and task-level validation;
• pool synthetic cognitive records across model families, topics, persona attributes, psychometric scores, confidence, accuracy, and semantic features;
• interpret pooled language through psychometric summaries and textual or behavioural formamentis networks while avoiding over-interpretation;
• identify ethical, methodological, and data-governance limits in synthetic cognitive datasets.

More information here.

Abstract

With the rapid development of smart city infrastructures and the growing tendency of municipalities to install sensors across the urban environment, Machine Learning (ML) models are rapidly becoming a central tool for understanding and predicting human mobility and transportation dynamics. In this tutorial, we walk through state-of-the-art ML models to address three mobility-related core tasks, highlighting how the recent paradigm of physics-informed learning can add value in terms of both predictive accuracy and interpretability. We complement the lecture with a live coding session built on a newly released city-scale dataset derived from traffic sensors in Padua, Italy, demonstrating how to build static and dynamic spatio-temporal graphs and how to correctly train and evaluate a baseline forecaster.

Presenters

Riccardo Cappi (University of Padua). I am a second-year PhD student at the University of Padua, working at the intersection between machine learning, symbolic regression, and physics-informed learning. My research aims to model mobility and transportation patterns using interpretable, physically-inspired machine learning approaches. I earned my master's degree in Computer Science with honors from the University of Padua in December 2025, with a thesis on equation discovery for graph dynamical systems using interpretable neural networks. Previously, I completed my bachelor's degree in Computer Science at the University of Insubria in February 2022. I am currently collaborating with the Municipality of Padua to collect and organize data from traffic sensors deployed across the city, and to develop machine learning models for traffic forecasting.

Aims and Learning Objectives

This tutorial offers participants an overview of how Machine Learning approaches can be used to model mobility and transportation dynamics. We focus on three key tasks: traffic forecasting, mobility flow generation, and traffic state estimation. For each, we walk through state-of-the-art ML methods,emphasizing a recurring theme throughout the tutorial: how data-driven methods can be combined with physical and domain knowledge to improve ML models in terms of both predictive accuracy and interpretability. By the end of the tutorial, participants will be able to understand the main paradigms in Machine Learning for mobility and transportation, including best practices, common pitfalls, and where physics-informed and hybrid models add value. They will also be able to locate and make use of relevant open-source datasets, benchmarks, and evaluation metrics.

Scope and Length

This is a 1.5-hour tutorial that combines a lecture-style presentation with a live coding demonstration. The coding part centers on a dataset derived from traffic sensors in Padua, and shows how to load the data, build static and dynamic spatio-temporal graphs, and train and evaluate a baseline traffic forecaster. Table 1 reports the indicative time breakdown of each segment.

Detailed Description

Predicting human mobility spans several distinct tasks [1], each with its own modeling challenges and applications. This tutorial focuses on three of the most relevant ones: traffic forecasting, mobility flow generation, and traffic state estimation. For each task, we introduce the Machine Learning (ML)and Deep Learning (DL) approaches proposed in the literature, highlighting their advantages over traditional models. We then discuss how the emerging paradigm of Physics-Informed Machine Learning[2] can bridge data-driven and physics-based methods, combining their complementary strengths to improve both predictive performance and interpretability. For each task, we also point participants to a curated set of open-source datasets and benchmarks, and present the most used performance metrics.

Traffic Forecasting with Spatio-Temporal Models. This section focuses on traffic forecasting, i.e., predicting how a given traffic variable (such as traffic volume or mean vehicle speed) evolves overtime. These variables are typically measured by sensors deployed across the city, such as loop detectors[3, 4, 5], or Automatic Vehicle Identification (AVI) systems [6, 7, 8]. We show how traffic dynamics can be modeled through different graph-based representations that connect sensors over the urban network, including distance-based, similarity-based, and dynamic graphs [9]. Additionally, we review state-of-the-art spatio-temporal ML models, clarifying the distinctions between Space-Then-Time (STT),Time-Then-Space (TTS), and Time-And-Space (TAS) architectures [10].

Mobility Flow Generation. This section addresses mobility flow generation, where the goal is to estimate the flows between a set of geographic locations from their characteristics, such as Points of Interest (POIs), land use, demographic features, and distance. We first present traditional transportation approaches, such as the gravity model, which assumes that flows follow a gravity-like law inspired by Newton's law of gravitation [11]. We then walk through recent ML models for this task, including the Deep Gravity model [12], spatial-aware models [13], and diffusion-based approaches for cross-city generation [14].

Physics-Informed Neural Networks for Traffic State Estimation. This section presents recent approaches that apply Physics-Informed Neural Networks (PINNs) to Traffic State Estimation (TSE), i.e.,reconstructing traffic state variables, such as vehicle density and mean velocity, along road segments over time from partially observed sensor data. We show how PINNs encode the principle of vehicle conservation by adding a penalty term to the loss function, derived from first- and second-order conservation laws [15, 16, 17, 18, 19, 20], drawing an analogy between vehicular traffic and fluid flow.

Live coding: a new city-scale dataset. This section centers on a multi-source dataset we recently released [8], derived from AVI recordings in Padua, Italy. It combines 10-minute traffic-volume timeseries with trajectory-based flow statistics (transition-probability matrices, average travel times, and flow residuals) and rich urban context (points of interest, demographics, weather, and the road network). We show how to build static and dynamic spatio-temporal graphs (with road-proximity or flow-based adjacency) using the released Python loader, and how to correctly train and evaluate a baseline forecaster, with attention to fair evaluation and avoiding data leakage.

Future Directions The tutorial concludes with an overview of open challenges and promising future directions in Machine Learning for mobility and transportation. We discuss recent efforts to extend the PINN framework to graph-based models [21], geographic transferability [14], the discovery of interpretable mobility equations from data [22, 23], and the emerging role of foundation models for mobility [24].

Target Audience

The tutorial targets researchers, graduate students, and industry or government practitioners in datascience with basic familiarity with machine learning. No prior background in the transportation domain is assumed. It is relevant to anyone working with spatio-temporal data, urban analytics, or mobility, and to public-sector practitioners interested in data-driven traffic management.

References

1. M. Luca, G. Barlacchi, B. Lepri, L. Pappalardo, A survey on deep learning for human mobility,ACM Computing Surveys (CSUR) 55 (2021) 1-44.
2. M. Raissi, P. Perdikaris, G. E. Karniadakis, Physics-informed neural networks: A deep learningframework for solving forward and inverse problems involving nonlinear partial differentialequations, Journal of Computational physics 378 (2019) 686-707.
3. Y. Li, R. Yu, C. Shahabi, Y. Liu, Diffusion convolutional recurrent neural network: Data-driventraffic forecasting, in: 6th International Conference on Learning Representations, ICLR 2018,Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings, OpenReview.net,2018. URL: https://openreview.net/forum?id=SJiHXGWAZ.
4. S. Guo, Y. Lin, N. Feng, C. Song, H. Wan, Attention based spatial-temporal graph convolutional net-works for traffic flow forecasting, in: Proceedings of the AAAI conference on artificial intelligence,volume 33, 2019, pp. 922-929.
5. C. Song, Y. Lin, S. Guo, H. Wan, Spatial-temporal synchronous graph convolutional networks:A new framework for spatial-temporal network data forecasting, in: Proceedings of the AAAIconference on artificial intelligence, volume 34, 2020, pp. 914-921.
6. Y. Wang, Y. Chen, G. Li, Y. Lu, Z. He, Z. Yu, W. Sun, City-scale holographic traffic flow data basedon vehicular trajectory resampling, Scientific Data 10 (2023) 57.
7. Q. Ma, X. Guo, W. Zhong, Z. He, Z. Su, W. Ma, R. Zhong, City-scale high-resolution traffic datasetswith refined networks for hierarchical traffic control, Scientific Data (2026).
8. R. Cappi, M. Luca, P. Fontolan, N. Navarin, B. Lepri, A. Sperduti, A city-scale dataset of trafficflows, travel times, and urban context, arXiv preprint arXiv:2605.18782 (2026).
9. G. Fan, A. Q. M. Sabri, S. S. A. Rahman, L. Pan, S. Rahardja, Emerging trends in graph neuralnetworks for traffic flow prediction: A survey: G. fan et al., Archives of Computational Methodsin Engineering 32 (2025) 4811-4855.
10. A. Cini, I. Marisca, D. Zambon, C. Alippi, Graph deep learning for time series forecasting, ACMComputing Surveys 57 (2025) 1-34.
11. G. K. Zipf, The p 1 p 2/d hypothesis: on the intercity movement of persons, American sociologicalreview 11 (1946) 677-686.
12. F. Simini, G. Barlacchi, M. Luca, L. Pappalardo, A deep gravity model for mobility flows generation,Nature communications 12 (2021) 6576.
13. X. Yao, Y. Gao, D. Zhu, E. Manley, J. Wang, Y. Liu, Spatial origin-destination flow imputation usinggraph convolutional networks, IEEE Transactions on Intelligent Transportation Systems 22 (2020)7474-7484.
14. Y. Li, J. Wang, X. Yu, P. Wang, Q. Huang, Cross city traffic flow generation via retrieval augmenteddiffusion model, Advances in Neural Information Processing Systems 38 (2026) 70025-70055.
15. A. J. Huang, S. Agarwal, Physics informed deep learning for traffic state estimation, in: 2020 IEEE23rd International Conference on Intelligent Transportation Systems (ITSC), IEEE, 2020, pp. 1-6.
16. A. J. Huang, S. Agarwal, Physics-informed deep learning for traffic state estimation: Illustrationswith lwr and ctm models, IEEE Open Journal of Intelligent Transportation Systems 3 (2022)503-518.
17. J. Zhang, S. Mao, L. Yang, W. Ma, S. Li, Z. Gao, Physics-informed deep learning for traffic stateestimation based on the traffic flow model and computational graph method, Information Fusion101 (2024) 101971.
18. R. Shi, Z. Mo, X. Di, Physics-informed deep learning for traffic state estimation: A hybrid paradigminformed by second-order traffic models, in: Proceedings of the AAAI conference on artificialintelligence, volume 35, 2021, pp. 540-547.
19. R. Shi, Z. Mo, K. Huang, X. Di, Q. Du, A physics-informed deep learning paradigm for traffic stateand fundamental diagram estimation, IEEE Transactions on Intelligent Transportation Systems 23(2021) 11688-11698.
20. X. Di, R. Shi, Z. Mo, Y. Fu, Physics-informed deep learning for traffic state estimation: A surveyand the outlook, Algorithms 16 (2023) 305.
21. J. Ji, J. Wang, Z. Jiang, J. Jiang, H. Zhang, Stden: Towards physics-guided neural networks fortraffic flow prediction, in: Proceedings of the AAAI conference on artificial intelligence, volume 36,2022, pp. 4048-4056.
22. O. Cabanas-Tirapu, L. Danús, E. Moro, M. Sales-Pardo, R. Guimerà, Human mobility is welldescribed by closed-form gravity-like models learned automatically from data, Nature Communi-cations 16 (2025) 1336.
23. H. Guo, W. Zhang, J. Yang, Y. Hou, L. Dong, Y. Liu, Distilling human mobility models with symbolicregression, Geographical Analysis 58 (2026) e70043.
24. M. Luca, C. Beneduce, B. Lepri, Time series foundation models are flow predictors, arXiv preprintarXiv:2507.00945 (2025).

Abstract

Large Language Models (LLMs) have recently demonstrated significant potential across several application domains. On the other hand, as with all AI-based models, LLMs are vulnerable to adversarial attacks, which involve deliberately manipulating input data to achieve malicious goals. The tutorial will provide an overview of how adversarial learning algorithms are categorized by goal, timing, and knowledge. The tutorial will also provide an overview of several state-of-the-art adversarial algorithms, ranging from pixel manipulations to prompt manipulations.

Presenters

Giuseppina Andresini. Giuseppina Andresini is a tenure-track Assistant Professor at the Department of Computer Science, the University of Bari Aldo Moro. She got a PhD in Machine Learning for Cybersecurity in March 2022. Her PhD thesis was awarded by Con.Scienze award in 2022. Her research activity mainly concerns deep learning, XAI and adversarial learning with applications in cybersecurity and remote sensing. On these topics, she has published several papers in international journals (e.g., Computer &Security, Journal of Intelligent Information Systems, Knowledge-Based Systems, Future GenerationComputer Systems, Information Sciences, Machine Learning Journal) and international conferences (e.g., Discovery Science and ISMIS). She was a co-chair of the Workshop on Machine Learning for Cybersecurity (MLCS) from 2021 to 2025, and she is a co-chair of the Workshop on Secure and Trustworthy AI (STAI@2026). She is Program Chair at ISMIS 2026. She presented a tutorial entitled "XAI in Cybersecurity" at ITADATA2023. She presented a talk entitled "Explainability to Foster Transparency and Trust in AI-powered Cybersecurity" in the workshop TAIC@ITASEC2025. She was a teacher of the course "Bridging Adversarial Learning and Data-Centric AI for Robust AI" at the Summer School ESSAI 2026. She was/is a member of the organizing committee of CyberChallenge.IT from 2021 to 2026. She served as a reviewer for several international conferences and journals and as a member of the program committee of several international conferences. She is a member of Women4Cyber.

Aims and Learning Objectives

With the increasing integration of Large Language Models (LLMs) into public and private systems (e.g., from conversational assistants to autonomous driving and healthcare), it is imperative to develop LLM-based decision support systems accounting fortext corpora contamination and to evaluate the robustness of these models against possible malicious use. Adversarial machine learning has established itself as a fundamental tool for understanding and mitigating vulnerabilities of AI-based systems. Attacks originally developed against computer vision classifiers have now evolved into prompt injection, jailbreaks, and training data contamination, all aimed at manipulating LLM outputs. Starting with an introduction toadversarial learning and foundational work in computer vision, this tutorial will provide an overview of adversarial vulnerabilities that may affect LLMs. It will map these algorithms to the classical taxonomy of adversarial machine learning, present some current state-of-the-art attacks against LLMs, and discuss defensive strategies to improve model robustness to adversarial threats. This will be done also accounting for the objectives of the project Sistemi cyber-fisici INTElligenti e SIcuri per la sanità, l'industria e la società (SINTESI).

Scope

The tutorial will provide an overview of adversarial vulnerabilities affecting large language models (LLMs), as well as the opportunities and challenges of defending against them. Adversarial attacks on LLMs, such as prompt injection, data poisoning, and model poisoning, are recognized as critical security risks in the development of AI applications and are highlighted in the OWASP Top 10 for Large Language Model Applications. In particular, the tutorial will introduce the foundational taxonomy of adversarial machine learning algorithms categorized by attacker goal, timing, and knowledge. It will present adversarial algorithms specifically targeting LLMs through input manipulations and discuss some defense mechanisms. Finally, the tutorial will explore current research challenges and emerging directions in securing LLM-based models to design trustworthy, robust and reliable generative AI systems. Overall, the tutorial aims to provide a comprehensive overview of the vulnerabilities affecting LLMs and the state-of-the-artapproaches for defending against them.

Length

1 hour

Full Description

As Large Language Models (LLMs) are increasingly integrated into several production systems (e.g.,chatbots, autonomous agents, code assistants, healthcare), their failures can produce catastrophic consequences. Although their use yields significant results in automating several tasks, they are susceptible to adversarial attacks. Adversarial attacks are inputs that are specifically designed to drive machine learning algorithms to produce erroneous outputs for malicious purposes [1]. These are samples created with, e.g., slight, intentional perturbations of original samples, designed to fool AI-based models into making incorrect predictions [2]. A foundational taxonomy of this adversarial Machine Learning categorizes these algorithms based on (1) attacker goal, (2) attack timing, and (3)attacker knowledge. Concerning the attacker's goal, adversarial attacks can be categorized into target attacks, where the adversarial sample is created to be misclassified into a specific class, and (2) untarget attacks, where the adversarial sample is created to be misclassified into any class different from the original class. Concerning the attack timing, these algorithms can be used (1) to perform attacks at training time (poisoning) by injecting adversarial samples into the training data to change the boundaries learned by the target, or (2) to perform attacks at testing time (evasion) with the goal of carefully creating adversarial samples against an already-trained model. Finally, based on the knowledge of attackers of the target model, adversarial attacks can be divided into (1) White box attacks, where the attackers generate adversarial samples based on the parameter settings of the attacked model, and (2) Blackbox attacks, where attackers do not know any information about the target model (the more common real-world situation). A similar categorization can also be defined in the specific approaches to evade LLMs. On the other hand, understanding adversarial attacks in the context of LLMs poses several challenges [3]. LLMs are large, complex models that perform generative tasks by maintaining context. In addition, they are often multimodal and are being integrated into complex ecosystems.

This tutorial will start with an introduction to adversarial learning as a strategy to decrease the reliability and trustworthiness of AI-based systems. Based on the categorization of adversarial attacks, the tutorial will provide an overview of the most common input-manipulation attacks on LLMs, such as prompt injection, jailbreak, or DAN [4, 5]. Also, the tutorial will introduce some defensive strategies tomake models more robust to adversarial manipulation, such as input filtering and sanitization [6] and adversarial training [7].

Acknowledgments

This is in fulfilment of the objectives of project "Sistemi cyber-fisici INTElligenti e SIcuri per la sanità,l'industria e la società - SINTESI" (PNRIC 21-27, DR 307 18.3.2025 - MI MI 1_93098910503_0000962,CUP B99H26000370007, CUP B92F26000470005).

References

1. C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, R. Fergus, Intriguingproperties of neural networks, 2014. URL: https://arxiv.org/abs/1312.6199. arXiv:1312.6199.
2. B. Biggio, F. Roli, Wild patterns: Ten years after the rise of adversarial machine learning, CCS '18,Association for Computing Machinery, 2018, p. 2154-2156. doi:10.1145/3243734.3264418.
3. M. Al Kuwaiti, H. Ismail, Adversarial attacks on large language models: A survey, in: V. Bhateja,M. El Barachi, A. T. Azar, D. K. Sharma (Eds.), Information System Design: Big Data Analytics andData Science, Springer Nature Singapore, Singapore, 2026, pp. 529-547.
4. F. Perez, I. Ribeiro, Ignore previous prompt: Attack techniques for language models, in: NeurIPSML Safety Workshop, 2022. URL: https://openreview.net/forum?id=qiaRo_7Zmug.
5. J. Wang, Attackeval: A systematic empirical study of prompt injection attack effectiveness againstlarge language models, 2026. arXiv:2604.03598.
6. P. Jalan, V. Abishethvarman, B. Chandna, U. Naseem, Survey on llm safety: Attacks, de-fenses, alignment, metrics, and guardrails, Machine Learning 115 (2026) 130. doi:10.1007/s10994-026-07060-8.
7. C. Anghel, M. V. Craciun, A. Cocu, A. A. Anghel, A. Istrate, Steadyeval: Robust llm exam graders viaadversarial training and distillation, Computers 15 (2026). URL: https://www.mdpi.com/2073-431X/15/1/55. doi:10.3390/computers15010055.

Abstract

This tutorial introduces YSocial, an open-source LLM-powered social media digital twin for designing, running, and analyzing controlled simulations of online social dynamics, recommender effects, misinformation, and agent interactions. Participants will learn how to configure scenarios, run simulations, analyze generated data with ySights, and explore the creation of reusable YSocial datasets through the SoBigData infrastructure.

Presenters

Giulio Rossetti. He is a Senior Researcher at ISTI-CNR and External Professor at the University of Pisa. He is also a member of the Knowledge Discovery and Data Mining Laboratory (KDD-lab), a joint research team that connects the Computer Science Department of the University of Pisa, the Institute of Information Science and Technologies of CNR, and the Scuola Normale Superiore. Within the lab, he coordinates the research activities on Complex Network Analysis. His research focuses on modeling, characterizing and forecast emerging (social) behaviors leveraging complex networks and AI/ML techniques. His current research encompasses dynamics on and of complex networks,feature-rich and higher-order network modeling.

Roberto Trasarti. He is a senior researcher and data scientist with a focus on mobility data analysis and behavioral patterns. He currently works at the National Research Council of Italy (CNR) and has contributed to numerous international research projects, specializing in the use of Big Data to understand social phenomena. With several scientific publications, Trasarti is a recognized expert in the field of complex data analysis and applied artificial intelligence, machine learning, data mining andautomatic reasoning. He is member of Knowledge Discovery and Delivery Laboratory, and coordinator of the SoBigData Research Infrastructure (www.sobigdata.eu) and its related projects: SoBigData++(H2020 GA n.871042) , SoBigData PPP (HORIZON-INFRA ESFRI N.101079043) and SoBigData.it (ItalianRRNP n.0000013).

Aims and Learning Objectives

This tutorial introduces YSocial¹, an open-source AI-driven social media simulator for studying social dynamics, recommender effects, misinformation, and online interactions in a controlled environment. YSocial combines agent-based modelling with Large Language Model agents and provides both aweb-based simulation interface and integrated analytics through ySights and Jupyter Lab. By the end of the tutorial, participants will be able to:

1. Understand the motivation and architecture of LLM-powered social media digital twins.
2. Configure a YSocial simulation scenario, including agents, simulation parameters, RSS/news sources, and recommender settings.
3. Run a small-scale social media simulation using YSocial
4. Inspect agent behaviours such as posting, commenting, reacting, sharing, following, and reading recommended content.
5. Analyze generated simulation data using ySights, including agent activity, posts, social networks,topic dynamics, and recommender-system effects.
6. Discuss methodological opportunities and limitations of LLM-based simulations for big data,computational social science, misinformation studies, and policy evaluation

Scope of the Tutorial

The tutorial is designed as a hands-on introduction to YSocial for researchers and practitioners interested in data science, big data analytics, computational social science, social simulation, recommender systems, misinformation, and LLM-based agent modelling. The session will cover:

• The conceptual foundations of social media digital twins.
• The YSocial ecosystem: web interface, admin panel, LLM agents, simulation server, and analytics.
• Scenario design: simulation parameters, agent populations, RSS feeds, recommendation mechanisms, and activity settings.
• A live demonstration of a small YSocial experiment.
• Data analysis using ySights, the Python library for extracting insights from YSocial simulations,including network analysis, agent/post exploration, recommender metrics, and visualizations.
• Best practices for reproducibility, responsible experimentation, and interpretation of LLM-agent simulations.

Length

A 2-hour format is requested because it allows both conceptual explanation and a meaningful live demonstration with analysis. A compressed 1-hour version is possible, but it would reduce the hands-onanalytics component.

Full Description

Online social media platforms are complex socio-technical systems in which users, content, recommendation algorithms, moderation rules, and platform affordances interact dynamically. Studying these systems only through observational data is often difficult because of limited access to platform internals, ethical constraints, non-repeatability of real-world events, and the challenge of testing counterfactual interventions. Digital twins and agent-based simulations provide an alternative approach: they allow researchers to build controlled, repeatable environments in which hypotheses about online interaction, information diffusion, recommender effects, and policy interventions can be tested before being deployed in real-world settings.

YSocial is an open-source AI-driven social media simulator designed for this purpose. It supports controlled simulation of social dynamics, recommender effects, and misinformation, and it is explicitly positioned for academic research, industry analysis, education, and policy-making use cases. The platform combines a social media-like web interface, an administrative interface for simulation configuration, LLM-powered agents, and integrated analytics. The public repository describes YSocial as an LLM-powered social media twin that allows users to create, configure, run, and analyze realistic simulations involving AI-driven agents and online communities.

The tutorial will begin with a short introduction to the need for simulation-based approaches in big data and data science. This is well aligned with the ITADATA2026 tutorial track, which welcomes practical knowledge, best practices, guidelines, live demonstrations, and prototypes in vertical problems of big data and data science. YSocial naturally fits this scope as a practical prototype and research platform for data-centric experimentation on online social systems.

The first part of the tutorial will introduce the YSocial architecture and modelling assumptions.YSocial agents act in an evolving online environment, producing and interacting with content through primitives inspired by microblogging platforms such as X/Twitter, Mastodon, and Bluesky. Supported actions include reading recommended posts, posting, commenting, replying, publishing news from RSSsources, sharing, reacting, receiving follow suggestions, and following or unfollowing other agents. This part will explain how such primitives support controlled experimentation on social interaction, information propagation, engagement, and network formation.

The second part will focus on scenario design. In YSocial, a scenario is the configuration of a simulation. It includes configurable parameters, RSS feeds that agents can read and share, and the recommender system used to suggest content or social connections. Participants will learn how to define a small experiment by selecting the number of agents, simulation duration, agent activity patterns, content sources, and interaction probabilities. The tutorial will emphasize practical design choices: choosing a meaningful experimental question, defining comparable simulation conditions, selecting recommender mechanisms, and documenting configuration files for reproducibility.

The third part will be a live demonstration. The presenters will show how to start YSocial locally, configure an LLM backend, launch the server, create or load a small scenario, and run a simulation. YSocial supports multiple LLM backends, including Ollama, vLLM, and custom OpenAI-compatible servers. Depending on infrastructure availability, the demo will either run live on a local machine or usea precomputed simulation database to guarantee reproducibility during the session. The demonstration will show both the administrative view and the user-facing social media interface, highlighting how human participants and AI agents can interact in the same simulated environment.

The final part will focus on analysis with ySights, the analytics library designed for YSocial simulation data. ySights provides tools for extracting insights from generated simulation data, including agent behaviours, content dynamics, network structures, and recommendation-system effects. It also includes data models for agents and posts, social network extraction, profile analysis, recommendation metrics, topic analysis, and visualization functions. Participants will learn how to load a simulation database, inspect agents and posts, extract the social graph, compute simple descriptive statistics, and visualize selected behavioural or network patterns.

An additional component of the tutorial will be dedicated to the creation of reusable simulation datasets. Participants will be shown how YSocial can be used not only as an interactive simulation environment, but also as a data generation tool for producing structured datasets describing synthetic social media interactions, including posts, comments, reactions, sharing behaviours, agent profiles, social ties, and recommender-system outputs. The tutorial will present the service offered by SoBigData2 to request the generation of dedicated datasets, enabling participants and interested researchers to access, reuse, and analyze YSocial-generated data for benchmarking, teaching, exploratory studies, and reproducible experiments in computational social science and big data analytics.

The tutorial will close with a discussion of research opportunities and methodological cautions.YSocial is introduced in the literature as an LLM-powered social media digital twin that uses LLMs to reproduce sophisticated agent behaviours and simulate user interactions, content dissemination, andnetwork dynamics. At the same time, participants will be encouraged to critically evaluate simulation outputs, distinguish exploratory simulation from empirical validation, document model assumptions, and consider ethical implications when studying misinformation, polarization, toxicity, or platform interventions.

Target Audience

The tutorial is intended for:

• Researchers and PhD students in big data, data science, computational social science, network science, AI, and social media analysis.
• Developers and data scientists interested in building reproducible pipelines for synthetic social media experimentation.
• Practitioners interested in simulation-based evaluation of online communities, recommender systems, misinformation, user engagement, or policy interventions.

No prior experience with YSocial is required. Basic familiarity with Python, Jupyter notebooks, and social network analysis will be useful but not mandatory.

Related Publications and Resources

References

1. G. Rossetti, M. Stella, R. Cazabet, K. Abramski, E. Cau, S. Citraro, A. Failla, R. Improta, V. Morini,and V. Pansanella. Y Social: an LLM-powered Social Media Digital Twin. Big Data & Society, 2026 (toappear). arXiv:2408.00818. Available at: https://arxiv.org/abs/2408.00818

Presenters

Raffaele Mattera (with Rosanna Verde). Raffaele Mattera is a Researcher in Statistics at the Department of Mathematics and Physics ofthe University of Campania "Luigi Vanvitelli". His research interests include time-series analysis, financial econometrics, statistical learning, clustering, and network modelling. His work focuses particularly on financial markets, volatility, systemic risk, and the analysis of complex economic and spatio-temporal data. He has published extensively in international journals and collaborates with researchers from several academic institutions. He teaches courses in statistical learning, experimental design, biostatistics, and time-series analysis within the Data Analytics and Data Science programmes.

Giancarlo Manzi (with Francesca De Battisti). Giancarlo Manzi is Associate Professor of Statistics at Sapienza University of Rome, within theDepartment of Methods and Models for Economics, Territory and Finance. His research interests include applied statistics, statistical modelling, missing-data imputation, big data, and computational methods. He has worked on methodological and interdisciplinary applications in economics, public services, health, epidemiology, and social sciences. He is involved in teaching courses on basic and advanced statistical methods for economics,business, society, and the environment. He has published extensively in international journals and contributed to several research projects combining statistical methodology and real-world data analysis.

Elena Siletti (with Francesca De Battisti and Marcella Lorenzini). Elena Siletti is a statistician at the Department of Economics and Statistics "Cognetti deMartiis" of the University of Turin. Her research focuses on applied statistics, statistical modelling, social and economic indicators, and the measurement of inequality and well-being. She has extensive experience in the analysis of complex and heterogeneous data, including administrative sources, surveys, social media, and other big-data environments. Her work also addresses selection bias, composite indicators, labour-market dynamics,sustainability, and interdisciplinary applications of quantitative methods.She has published in several international journals and collaborates on research projects combining methodological innovation with substantive social and economic questions.

Jacopo Chiapparino (with Alessandro Bitetto and Paola Cerchiello). Jacopo Chiapparino is a highly skilled economist and data scientist with expertise in Python, R, machine learning, and financial econometrics, currently pursuing a PhD in Artificial Intelligence applied to finance. Experienced in AI engineering, risk management, and quantitative modelling, with a strong academic background and ongoing research in high-frequency trading and pattern recognition. Recognized for innovative contributions, including publications and awards in financial and technical analysis fields.

Aims and Learning Objectives

The tutorial aims to showcase recent contributions in the field of data analysis, where statistical modeling and machine learning intersect to address real-world challenges. Emphasis will be placed on the interpretability and practical implications of the proposed techniques, particularly in the analysis of high dimensional, complex, and multi-source datasets. Participants will gain insights into how advanced statistical tools and AI-driven models can enhance understanding and decision-making across domains such as healthcare, finance, and social dynamics.

Scope of the Tutorial

This tutorial offers an overview of innovative approaches to constructing and analysing complex data across statistical, historical, bibliometric, and environmental research. They address the preservation of uncertainty in interval-valued observations, the reproducible transformation of unstructured historical records into reliable datasets, and the analysis of publication waiting times in leading statistics journals. They also investigate dynamic environmental index construction through factor models and deep autoencoders. Overall, the studies emphasise reproducibility, methodological transparency, robust validation, explainability, and the integration of advanced computational techniques with domain-specific interpretation.

Length

1 hours

Full Description

The tutorial is organised in four presentations. Each presentation will take 15 minutes. The contributions that will be presented at the tutorial are as follows:

Presenters: Raffaele Mattera, Rosanna Verde
Title: Learning from interval-valued data: recent perspectives
Abstract: Interval-valued data represent observations with an intrinsically set-valued structure, whose uncertainty must be handled differently from that associated with classical point-valued data. As increasingly complex interval-valued data become available, new methodological challenges emerge for statistical learning. We discuss recent developments in clustering and dependence modelling, both temporal and spatial, for interval-valued data and propose future directions for new methodologies preserving the uncertainty inherent in interval observations.

Presenters: Francesca De Battisti, Marcella Lorenzini, Elena Siletti
Title: From Historical Records to Research Data: Designing a Reproducible Workflow for Early Modern Notarial Sources
Abstract: Early modern notarial records represent a unique source for investigating economic practices, social relationships, and the role of women in financial transactions. Yet, their heterogeneous and largely unstructured nature poses significant challenges for the systematic creation of research datasets. This contribution presents the conceptual and methodological framework of an ongoing project to design a reproducible workflow to transform historical notarial records into structured data suitable for quantitative and qualitative analyses. The proposed approach combines complementary information extraction strategies within a modular pipeline, with particular attention to transparency, reproducibility, and future validation procedures. Rather than presenting a completed system, the contribution discusses the rationale for the project, the design choices that guided the development of the workflow, and the challenges of balancing computational methods with historical interpretation. The contribution aims to foster discussion of methodological issues in creating reliable historical datasets for digital historical research.

Presenters: Francesca De Battisti, Giancarlo Manzi
Title: En attendant Godot: A dataset for modeling waiting times for academic publications
Abstract: This paper introduces a novel dataset designed to identify the response variables and covariates that influence the academic peer-review process, with a specific focus on publication waiting times across top-tier statistics and probability journals. Utilizing bibliometric sources, journal websites, and original PDF articles, we constructed two distinct datasets: the first comprises papers from eight leading statistics journals published between 2011 and 2026, while the second encompasses articles from twenty-six top statistics journals spanning 2011 to 2022. Web scraping techniques were employed to extract submission andacceptance dates. Furthermore, we discuss potential modelling approaches; for instance, these datasets can be utilized in longitudinal analyses to identify the primary determinants of review durations. Finally, Latent Dirichlet Allocation (LDA) can be applied to categorize articles by topic, thereby generating additional covariates for subsequent statistical models. Ultimately, this dataset offers researchers valuable insights into journal efficiencies, enabling more informed and strategic decisions when selecting publication venues.

Presenters: Jacopo Chiapparino, Alessandro Bitetto, Paola Cerchiello
Title: Toward a Dynamic Data-Driven Environmental Index: Extending DDEPI via DFM and DeepAutoencoders
Abstract: In this paper we revisit the Data-Driven Environmental Performance Index (DDEPI) ofAgosto et al. as a dynamic index-construction problem. The original DDEPI provides atransparent ranking based on Robust PCA, but its yearly construction remains essentially static and does not explicitly exploit the temporal structure of the panel. We therefore build avalidation framework for a dynamic environmental composite indicator that preserves thedata-driven spirit of DDEPI while allowing for latent temporal dependence. The framework combines an explicit audit of the input representation, candidate Dynamic Factor Models and deep autoencoder architectures, researcher-specified synthetic scenarios, and a real-data-calibrated digital twin with known latent truth. The resulting evidence is intentionally regime-dependent: level-based views recover the static DDEPI structure, tuned linear baselines remain strong in quasi-linear and quasi-static settings, and the potential value of deep models is assessed in nonlinear, heterogeneous, and genuinely dynamic regimes. Finally, we pair index construction with an explainability and robustness layer based on SHAP, attention-derived importance, permutation-based checks, bootstrap confidence intervals, and sensitivity analyses across seeds and validation schemes.

Target Audience

Researchers, PhD students, and students in the fields of statistics, machine learning and artificial intelligence who are particularly interested in proposing solutions to aspects of explainability and interpretation.

Abstract

Data-driven AI is transforming how intelligent systems are designed, deployed, and used, reshaping research, industry, and society. As these systems increasingly inform decisions and operate in complex environments, their dependence on data introduces risks that extend beyond predictive accuracy. Data quality, changing operating conditions, and the interaction between AI and other system components raise questions about reliability, transparency, explainability, and responsibility. Understanding these risks and establishing trust are essential to responsible adoption.

This panel brings together diverse perspectives to discuss security, privacy, fairness, robustness, and accountability of data-driven AI, and to examine how trust in intelligent systems can be assessed and sustained throughout their lifecycle. The discussion will explore emerging challenges and opportunities, considering how research, technological innovation, and governance can shape the future of reliable and responsible AI.

Donato Malerba — University of Bari Aldo Moro

• Annalisa Appice — University of Bari Aldo Moro
• Luigi Romano — University of Naples Parthenope
• Barbara Alicino — Lutech S.p.A.
• Davide Dalle Carbonare — Engineering S.p.A.
• Richard Lloyd Stevens — IDC Italia; Steering Board member, Big Data Value Association (BDVA)

Abstract

How can Europe turn fragmented data infrastructures into trustworthy ecosystems that strengthen societal resilience? This panel synthesizes the approaches and cross-domain lessons emerging from the workshop "Building Resilient Data Ecosystems," connecting perspectives from healthcare, critical infrastructures, energy, and maritime services.

The discussion spans five complementary application areas: trustworthy AI and interoperable health data infrastructures for cancer care; risk assessment and decision support for critical infrastructures and essential services; federated Digital Twins for energy systems and communities; sovereign data sharing across the maritime value chain; and integrated human, animal and environmental health through digital innovation ecosystems.

Across these domains, a common insight emerges: resilience depends not only on advanced analytics, but also on the ability to make heterogeneous data, models and services interoperable within accountable governance frameworks. Data Spaces, privacy-preserving approaches and Digital Twins offer complementary foundations for collaboration, while explainability, data sovereignty and integration into operational workflows are essential conditions for adoption.

Building on this synthesis, the panel examines shared challenges in balancing interoperability with local control, translating AI outputs into actionable decisions, and sustaining collaboration among research institutions, industry, public authorities and healthcare stakeholders. Particular attention is given to how technical architectures and governance arrangements can jointly support preparedness, coordinated responses and adaptation.

The panel brings these perspectives together to identify common research priorities, transferable practices and opportunities for cross-domain collaboration, outlining an agenda for trustworthy, sustainable and resilient data ecosystems across Europe.

Luigi Romano - University of Naples "Parthenope"

• Vincenzo Croce, Engineering Ingegneria Informatica
• Manlio Serio, Engineering Ingegneria Informatica
• Castino Andrea, Fincantieri
• Corsano Marianna, Fincantieri