Book of Abstracts

Workshop welcome and institutional framing by Salvatore Ruggieri and Francesca Chiaromonte.

Keynote lecture by Prabhani Kuruppumullage Don (Pennsylvania State University).

Authorship Analysis (AAn) is a Natural Language Processing (NLP) task that aims to infer characteristics of the author of a linguistic text. These characteristics may include the author’s identity as well as biographical and sociolinguistic information, such as age, gender, native language, and political orientation. AAn has important applications in areas such as cultural heritage, forensic linguistics, and cybersecurity. In the latter domain, it can be used to detect, discourage, or trace criminal activities including phishing, cyberbullying, and identity theft. A large body of research has focused on applying AAn techniques to online communication, including emails, blogs, social media posts, and tweets. In particular, authorship analysis can support the monitoring of harmful or illegal content shared on social media platforms and help identify posts that violate platform policies or legal regulations. At the same time, AAn raises significant privacy concerns. Its ability to de-anonymize authors or link pseudonymous identities may endanger individuals such as whistleblowers, journalists, or political activists. As a result, increasing attention has been devoted to methods for intentionally modifying writing style in order to conceal authorial identity and personal characteristics. This task, commonly referred to as adversarial stylometry or authorship obfuscation, seeks to reduce the effectiveness of stylometric analysis. This talk provides an overview of the field, its main applications and recent developments.

Online public debate does not evolve in a neutral space: the structure of interactions and the context in which they occur can significantly shape how discussions develop. While some environments amplify polarization and conflict, others foster support and empathy. This talk explores these differences by examining online discussions across contrasting information environments, showing how patterns of interaction and language vary, and what these dynamics reveal about the evolution of public discourse in digital spaces.

Graph neural networks (GNNs) are effective for node classification when labels can leverage information from local neighborhoods. However, they can struggle when prediction depends on long-range interactions, due to well-known problems such as oversquashing. To address this issue, prior work has proposed rewiring the graph topology to improve signal propagation. In this work, we introduce RAwR, a novel and efficient rewiring method that creates a quotient graph from an equitable partition and connects it to the input graph. This enables faster communication between nodes with the same structural role -- i.e., the same Weisfeiler-Leman graph coloring -- and reduces the total effective resistance. Furthermore, an approximate definition of the equitable partition allows for controllable shrinking of the quotient graph until it collapses to a single node, thereby recovering the well-known Master Node rewiring technique. Across a broad evaluation benchmark, including standard homophilic and heterophilic datasets as well as synthetic graphs specifically designed for long-range interactions, RAwR achieves state-of-the-art results. We also analytically investigate the improvements that RAwR can achieve in an idealized teacher-student model of linear GNNs, explaining when and why role-based rewiring helps. This theoretical insight leads to the definition of Spectral Role Lift (SRL), a measure useful for identifying the approximate equitable partition that leads to the best performance.

Feature attribution is the dominant paradigm for explaining the predictions of complex machine learning models like neural networks. However, most existing methods offer little guarantee of reflecting the model's prediction-making process. We define the notion of explanatory alignment and argue that it is central to trustworthy predictive modeling: in short, it requires that explanations directly underlie predictions rather than serve as rationalizations. We present model readability as a design principle enabling alignment, and Pointwise-interpretable Networks (PiNets) as a modeling framework to pursue it in a deep learning context. PiNets combine statistical intelligence with a pseudo-linear structure that yields instance-wise linear predictions in an arbitrary feature space. We illustrate their use on image classification and segmentation tasks, demonstrating that PiNets produce explanations that are not only aligned by design but also faithful across other dimensions: meaningfulness, robustness, and sufficiency.

The talk covers recent applications of machine learning in finance I have been working on in the last ten years with my PhD students: 1. The use of large language models to assess sentiment in financial news, correlate it with returns, and deploy it in sentiment-based trading strategies; 2. The use of reinforcement learning and sentiment-augmented reinforcement learning in portfolio allocation. 3. The use of deep neural networks as surrogates to speed up the pricing with stochastic models; 4. The use of deep neural networks for model calibration, i.e. the forward-looking estimation of model parameters from the market prices of European options; 5. The use of various machine learning techniques (logistic regression, support vector machines, neural networks, Bayesian regularisation, k-nearest neighbours, etc.) for credit scoring. Not all these approaches work equally well or have an edge with respect to traditional methods just because they are based on machine learning: the first has the most spectacular results, the fifth the least, the third and fourth depend on the model.

The presentation will argue that the rise of AI‑driven disinformation strategies fundamentally alters the relationship between freedom of expression and the rule of law by amplifying, accelerating, and obscuring manipulative political communication at an unprecedented scale. This development challenges the judiciary’s classical role as a guardian of free expression, raising the risk that interventions justified by opaque AI systems could be misused in contexts of democratic backsliding. Within this transformed landscape, EU instruments such as the Digital Services Act and the AI Act reconfigure the rule of law by imposing transparency, accountability, and risk‑mitigation duties on private actors whose AI technologies increasingly shape democratic discourse.

The advent of AI, particularly generative AI, raises numerous concerns regarding the protection of individuals most exposed to the risks arising from the pervasive use of new technologies across various aspects of daily life. The EU legislator has acknowledged these challenges, consequently mandating respect for the category of so-called "vulnerable" subjects. The discussion aims to address the concept of vulnerability in light of the AI Act, envisaging a governance model in which the protection of fragility becomes the fundamental benchmark for the very legitimacy of technological innovation.

The regulatory landscape in the European Union has seen two different trends in the past few years: after a wave of regulations which left us with the AI Act, the Digital Services Act, the Data Governance Act, the Data Act (just to name a few) and other landmark pieces of legislation, the new direction is going towards simplification, following the narrative that “regulation stifles innovation”. This talk will challenge this assumption and discuss responsible research practices to foster ethical innovation in the field of new technologies.

In a world pervaded by artificial intelligence, the law must maintain a predominant role in safeguarding human rights, interests, and legal certainty. However, it is increasingly difficult for regulation to keep pace with rapidly evolving technologies. A key issue concerns the interpretation of the AI Act, particularly Article 5(1)(a), which prohibits AI systems using manipulative, deceptive, or subliminal techniques that cause significant harm. Yet, the notion of “significant harm” is not defined, leaving interpreters to determine its meaning and increasing discretion and legal uncertainty. In a highly technological context, it is therefore crucial to identify the threshold beyond which harm becomes significant, in order to prevent prejudicial situations and ambiguities in enforcement. This requires analyzing European legislation on harm, its subcategories, and related concepts such as severity and legal violation. An interesting case study is that of voice-based virtual assistants, which use NLP and API techniques to provide timely responses to users. How might these systems manipulate or deceive users and lead to unconscious choices? And under what conditions could such conduct cause significant harm? This analysis aims to identify when such behaviors amount to manipulation, deception, or subliminal influence, providing guidance both ex ante for developers and ex post for affected users.

What kind of acting and, consequently, of responsibility arises in the context of artificial intelligence systems? On the one hand, an artificial intelligence system appears capable of an “acting without action” – that is, without a subject – which emerges in our very relationship with the system itself. This would call for a regime of moral responsibility different from that of fault tout court. Rather, by building upon the recognition of the vulnerabilities of the agents involved, a kind of dynamically negotiated responsibility would seem to arise. On the other hand, this appears to align well with the more general legal orientation, which tends toward liability without fault, or objective liability. Setting aside doctrinal nuances, there may be a correspondence between the two domains, namely the ethical and the legal one. Such an approach, which will be described during the speech, could do justice to already proposed solutions, such as the logging of interactions provided for in the EU Artificial Intelligence Act.

Innovation emerges from complex collaboration patterns - among inventors, firms, or institutions. However, not much is known about the overall mesoscopic structure around which inventive activity self-organizes. Here, we tackle this problem by employing patent data to analyze both individual (co-inventorship) and organization (co-ownership) networks in three strategic domains (artificial intelligence, biotechnology and semiconductors). We characterize the mesoscale structure (in terms of clusters) of each domain by comparing two alternative methods: a standard baseline - modularity maximization - and one based on the minimization of the Bayesian Information Criterion, within the Stochastic Block Model and its degree-corrected variant. We find that, across sectors, inventor networks are denser and more clustered than organization ones - consistent with the presence of small recurrent teams embedded into broader institutional hierarchies - whereas organization networks have neater hierarchical role-based structures, with few bridging firms coordinating the most peripheral ones. We also find that the discovered meso-structures are connected to innovation output. In particular, Lorenz curves of forward citations show a pervasive inequality in technological influence: across sectors and methods, both inventor (especially) and organization networks consistently show high levels of concentration of citations in a few of the discovered clusters. Our results demonstrate that the baseline modularity-based method may not be capable of fully capturing the way collaborations drive the spreading of inventive impact across technological domains. This is due to the presence of local hierarchies that call for more refined tools based on Bayesian inference.

This work investigates the application of network topology and machine learning for systemic risk prediction in equity markets. Using daily returns from S&P 500 constituents, we construct dynamic correlation networks to extract high-dimensional topological features, including eigenvalue-based metrics (absorption ratio, network entropy) and graph-theoretic centralities. We evaluate a comparative suite of predictive architectures—ranging from Gradient Boosted Decision Trees to Graph Neural Networks (GraphSAGE, GAT) and LSTMs—validated through walk-forward cross-validation with purging. Our findings reveal that network topology exhibits statistically significant anomalies an average of 67 days prior to crisis onset, with lead-lag analysis confirming predictive causality over traditional implied volatility measures (VIX). Furthermore, we examine market microstructure through the lens of herding behavior (CCK framework) and volatility spillovers (Diebold-Yilmaz), ultimately translating these signals into economically significant trading strategies such as Kelly-style probability scaling.

Network embedding is a fundamental technique to project a network into a lower-dimensional space while preserving similarities among nodes. Traditional network embeddings primarily capture node proximity, making them effective for community detection but insufficient for identifying roles, i.e., patterns of interaction beyond local neighborhoods. To address this limitation, we introduce a simple and efficient embedding technique based on approximate variants of equitable partitions. Our approach, called ε-BE, introduces a user-tunable tolerance parameter relaxing the otherwise strict condition for exact equitable partitions that can be hardly found in real-world networks. We exploit a relationship between equitable partitions and equivalence relations for Markov chains and ordinary differential equations to develop a partition refinement algorithm for computing an approximate equitable partition in polynomial time. We extend this framework to weighted and directed networks, ensuring applicability to a more general class of graphs and filling a gap in the literature where few approaches are present. We compare our method against state-of-the-art embedding techniques on synthetic and real-world networks. We report comparable, when not superior, performance for visualization, classification, clustering, and regression tasks with smaller running times, enabling the embedding of large-scale networks that could not be efficiently handled by most of the competing techniques. These results and the capability to handle weighted and directed networks make our approach a compelling alternative for structural network embedding.

Bipartite networks provide a fundamental insight into the organisation of complex real-world systems. A key challenge in modeling these systems is devising a monopartite projection that preserves the intricate information encoded within the original bipartite structure. We propose an unsupervised algorithm to obtain statistically validated projections of bipartite signed networks, according to which any two nodes sharing a statistically significant number of concordant (discordant) motifs are connected by a positive (negative) edge. By assessing statistical significance through four distinct Exponential Random Graph Models (ERGMs), we generate link-specific p-values filtered via multiple testing correction. After validating the method on synthetic configurations from a fully controllable generative model, we apply it to three real-world social networks. In all cases, the algorithm detects non-trivial mesoscopic structures that cannot be explained by the constraints of the null models, thus unveiling the authentic signed complexity of the underlying system. Finally, we show how the inherent flexibility of our framework allows for easy extensions to more sophisticated null models and different complex systems.

Final takeaway and workshop wrap-up by Andrea Vandin.