Numerous social movements (SMs) around the world help support the UN's Sustainable Development Goals (SDGs). Understanding how key events shape SMs is key to the achievement of the SDGs. We have developed SMART (Social Media Analysis & Reasoning Tool) to track social movements related to the SDGs. SMART was designed by a multidisciplinary team of AI researchers, journalists, communications scholars and legal experts. This paper describes SMART's transformer-based multivariate time series Discourse Evolution Engine for Predictions about Social Movements (DEEP) to predict the volume of future articles/posts and the emotions expressed. DEEP outputs probabilistic forecasts with uncertainty estimates, providing critical support for editorial planning and strategic decision-making. We evaluate DEEP with a case study of the #MeToo movement by creating a novel longitudinal dataset (433K Reddit posts and 121K news articles) from September 2024 to June 2025, which is publicly available for research purposes.

Abstract Named Entity Recognition (NER) in specialized domains like biomedicine suffers from acute data scarcity, requiring expensive expert annotations. While data augmentation offers a promising solution, it inevitably introduces noisy and mislabeled samples that can degrade model performance. This problem is amplified in few-shot scenarios where every training example matters. We introduce PALAUNER (Policy-based Active Learning to Augment Named Entity Recognition), a reinforcement learning framework that learns to select high-quality samples from augmented data pools. Using a deep Q-network, our agent evaluates samples based on content features and model predictions, deciding which examples will improve NER performance. Experiments across five BioNER benchmarks demonstrate that PALAUNER consistently enhances diverse augmentation methods, from simple perturbations to GPT-based generation. Average F1 improvements are of 0.5−7.1 points in few-shot settings. PALAUNER’s modular design enables seamless integration with emerging augmentation techniques, providing a generalizable solution for training data quality enhancement. We publicly release our code on GitHub: ( https://github.com/picuslab/palauner ).

Abstract Smart cities, as a constantly evolving field, offer numerous opportunities for advancement and innovation. In particular, more and more specific sectors within smart cities currently require increased attention and development. Notably, the integration of emerging technologies holds promise for significant progress in efficiency and management across various urban issues, with urban waste management being a key focus of this research. In this paper, we propose a new methodology for intelligent waste management combining graph-based modeling and advanced machine learning techniques. A case study of a municipality in the Melbourne area (Australia) is examined, where smart bins equipped with fill-level sensors have been deployed, and their data made publicly accessible. The core of this investigation lies in utilizing sophisticated machine learning models, including Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), for predicting bin fill levels. Additionally, we introduce a graph-based spatial representation in which bins are modeled as nodes of a weighted graph, and graph embedding techniques are employed to capture and encode spatial dependencies among them. The findings validate the effectiveness of the proposed methodology, achieving a Mean Absolute Error (MAE) of 0.16, a Root Mean Squared Error (RMSE) of 0.21, a Mean Squared Error (MSE) of 0.044, and an R 2 of 0.83, which outperforms the reference baseline. Our results highlight how the integration of spatial information through graph embedding holds promise for improving predictions in smart waste management. Furthermore, the study suggests a transferable methodological framework, applicable at the architectural level to smart city challenges where spatially distributed entities exhibit interdependent temporal dynamics, pending empirical validation across further domains and cities.

As Large Language Models (LLMs) generate increasingly sophisticated text-image pairs, the risk of LLMs being used for phishing is growing, posing significant challenges for Internet security. We conduct a detailed, IRB-approved study of pairs (uc,pc) where uc is a user property (e.g., demographic or behavioral) and pc is a property of the post content (e.g., topic, emotion, and recency). Unlike past research that focuses either on user properties alone or post content alone, we conduct a comprehensive statistical analysis of more than 1,400 hypotheses involving such (uc,pc) pairs and identify the conditions on user-content pairs that have a significantly higher click probability compared with pairs that do not satisfy the condition. The results include highly nuanced findings connecting susceptibility to phishing with the combination of user properties and post content properties. They also cast some light on contradictory findings from prior work on understanding human susceptibility to phishing.

Fake audios, videos, and images are now proliferating widely. We developed GODDS, the Global Online Deepfake Detection system, for a specific user community, namely journalists. GODDS leverages an ensemble of deepfake detectors, along with a human in the loop, to provide a deepfake report on each submitted video/image/audio or VIA artifact submitted to the system. To date, VIA artifacts submitted by over 50 journalists from outlets such as the New York Times, Wall Street Journal, CNN, Agence France Press, and others have been run through GODDS. Unlike other deepfake detection systems, GODDS doesn't just focus on the submitted artifact but automatically derives context about the subject of the VIA artifact. Because context is not always available on all subjects, GODDS focuses on alleged deepfakes of high profile individuals, organizations, and events, where there is likely to be considerable contextual information.

Vision-Language Models (VLMs) excel in multimodal tasks, yet their effectiveness in specialized medical applications remains underexplored. Accurate interpretation of medical images and text is crucial for clinical decision support, particularly in multiple-choice question answering (MCQA). To address the lack of benchmarks in this domain, we introduce the Multilingual Multimodal Medical Exam Dataset (MMMED), designed to assess VLMs' ability to integrate visual and textual information for medical reasoning. MMMED includes 582 MCQA pairs from Spanish medical residency exams (MIR), with multilingual support (Spanish, English, Italian) and paired medical images. We benchmark state-of-theart VLMs, analyzing their strengths and limitations across languages and modalities. The dataset is publicly available on Hugging Face (https://huggingface.co/datasets/praiselab-picuslab/MMMED), with experimental code on GitHub (https://github.com/PRAISELab-PicusLab/MMMED).

Medical question answering systems face deployment challenges including hallucinations, bias, computational demands, privacy concerns, and the need for specialized expertise across diverse domains. Here, we present SOLVE-Med, a multi-agent architecture combining domain-specialized small language models for complex medical queries. The system employs a Router Agent for dynamic specialist selection, ten specialized models (1B parameters each) fine-tuned on specific medical domains, and an Orchestrator Agent that synthesizes responses. Evaluated on Italian medical forum data across ten specialties, SOLVE-Med achieves superior performance with ROUGE-1 of 0.301 and BERTScore F1 of 0.697, outperforming standalone models up to 14B parameters while enabling local deployment. Our code is publicly available on GitHub: https://github.com/PRAISELab-PicusLab/SOLVE-Med.

The rapid rise of deep generative models has enabled realistic synthetic media, posing severe risks to healthcare communication where trust and accuracy are essential. Existing deepfake detectors, mostly trained on entertainment-oriented datasets, degrade sharply in medical contexts, where domainspecific terminology and evidence-based discourse are critical. Moreover, post-hoc moderation is often too slow to mitigate the spread of health-related disinformation, underscoring the need for timely, domain-adapted solutions. We address these gaps by (i) curating a multimodal healthcare-oriented corpus of in-the wild and synthetically generated deepfakes, (ii) benchmarking state-of-the-art monomodal detectors for video, audio, and text, and (iii) introducing a late-fusion framework that integrates their predictions via ensemble methods. Our experiments show that late fusion consistently outperforms individual detectors, yielding notable gains in macro-F1 and AUC (+0.159 and +0.131, respectively), particularly under domain-shift and lowresource conditions. These findings demonstrate the importance of multimodal integration for robust detection of health-related deepfakes and contribute toward building trustworthy AI systems to safeguard medical communication. Code and data will be released for transparency and reproducibility: https://github.com/PRAISELab-PicusLab/Deepfake-Detection-in-Healthcare

Misinformation in healthcare, from vaccine hesitancy to unproven treatments, poses risks to public health and trust in medical systems. While machine learning and natural language processing have advanced automated fact-checking, validating biomedical claims remains uniquely challenging due to complex terminology, the need for domain expertise, and the critical importance of grounding in scientific evidence. We introduce CER (Combining Evidence and Reasoning), a novel framework for biomedical fact-checking that integrates scientific evidence retrieval, reasoning via large language models, and supervised veracity prediction. By integrating the text-generation capabilities of large language models with advanced retrieval techniques for high-quality biomedical scientific evidence, CER effectively mitigates the risk of hallucinations, ensuring that generated outputs are grounded in verifiable, evidence-based sources. Evaluations on expert-annotated datasets (HealthFC, BioASQ-7b, SciFact) demonstrate state-of-the-art performance and promising cross-dataset generalization. Code and data are released for transparency and reproducibility: https://github.com/PRAISELab-PicusLab/CER

Biomedical misinformation - ranging from misleading social media posts to fake news articles and deepfake videos - is increasingly pervasive across digital platforms, posing significant risks to public health and clinical decision-making. We developed CER, a comprehensive fact-checking system designed specifically for biomedical content. CER integrates specialized components for claim detection, scientific evidence retrieval, and veracity assessment, leveraging both transformer models and large language models to process textual and multimedia content. Unlike existing fact-checking systems that focus solely on structured text, CER can analyze claims from diverse sources including videos and web content through automatic transcription and text extraction. The system interfaces with PubMed for evidence retrieval, employing both sparse and dense retrieval methods to gather relevant scientific literature. Our evaluations on standard benchmarks including HealthFC, BioASQ-7, and SciFact demonstrate that CER achieves state-of-the-art performance, with F1-score improvements compared to existing approaches. To ensure reproducibility and transparency, we release the GitHub repository with the source code, within which you can reach an interactive demonstration of the system published on HuggingFace and a video demonstration of the system \https://github.com/PRAISELab-PicusLab/CER-Fact-Checking.

In last years, social lending platforms have been increasingly used as virtual environments where borrowers can directly interact with lenders without any intermediary. As a result, a reliable credit scoring strategy, i.e., assessing whether a client is able to fully repay a loan, became of utmost importance to reduce the risk of not repaying the lenders. In this context, machine learning tools are being increasingly adopted to design automatic credit scoring systems but the data imbalance problem still penalizes their predictive performance, i.e., the greatest majority of clients can afford the repayment and learning to classify ”bad” borrowers depends on few instances where the loan was not paid back. In this paper, we target the data imbalance problem and propose a novel data augmentation strategy to improve the predictive performance of credit scoring models. The proposed methodology performs data augmentation by injecting synthetic instances in the dataset generated along the decision boundary of the decision model. We assessed the effectiveness of the proposed augmentation strategy on a million-scale dataset from Lending Club, the largest Social Lending platform, and found that it improves the performance of several classification models, also in comparison to other state-of-the-art approaches.

Geopolitical news provides vast amounts of information essential for understanding international relations and political events. However, organizing this information into a coherent, structured format poses challenges due to the complexity and dynamic nature of the domain. This paper introduces a scalable system leveraging Large Language Models to build continuously updated Knowledge Graphs from Italian geopolitical news. The system features a modular architecture, including a Collector Node for scalable article extraction, a Redis-based reliable queue to manage large-scale data ingestion, and a Named Entity Recognition/Relation Extraction Engine to standardize entity-relation triples. The framework addresses key challenges, such as continuous updating and hallucination mitigation, ensuring the reliability of the graph. Our evaluations demonstrate significant improvements in scalability, uniformity of extracted triples, and graph accuracy, making this architecture particularly suitable for real-time geopolitical analysis.

Legal Artificial Intelligence has emerged as an essential field, focusing on AI technologies that facilitate various legal tasks and alleviate the workload of legal professionals. Despite advancements in Legal Artificial Intelligence, there remains a critical gap in systems that can provide both comprehensive and contextually accurate retrieval tailored to the intricate structure of EU legislation. We propose EuropeanLawAdvisor, an efficient and user-friendly legal information retrieval system designed to deliver tailored responses to legal queries. This system utilizes open-source Large Language Models within a Retrieval-Augmented Generation framework, facilitating precise and relevant information retrieval. The system employs a robust retrieval approach that integrates multi-match, k-nearest neighbors, hybrid methods, and TF-IDF search strategies across both complete documents and segmented text indexes, ensuring comprehensive retrieval for diverse query types. The implementation of the framework has demonstrated significant improvements in the accuracy and relevance of responses to EU legal queries, enhancing both the retrieval of relevant legal documents and the generation of precise responses. We show that EuropeanLawAdvisor, leveraging open-source models like Phi3-mini-3B and LLaMa-3-8B, achieves competitive Faithfulness and Relevance compared to GPT-4-Turbo. The performance gap narrows significantly in zero-shot scenarios, and our approach outperforms GPT-4-Turbo in the percentage of answered questions. We publicly release our code on GitHub: https://github.com/raffaele-russo/EuropeanLawAdvisor.

Nowadays, our security and privacy are strongly threatened by malware programs which aim to steal our confidential data and make our systems out of service, among other things. While traditional signature-based malware detection methods or statistical analysis have proven to be ineffective and time-consuming, recently data-driven Artificial Intelligence (AI) techniques, i.e. Machine Learning (ML) and Deep Learning (DL) approaches, have been successfully applied leveraging the behaviour of malware in terms of API calls, and achieving promising performances. However, their black-box behavior leads to a lack of explainability thus preventing their application in real world scenarios. In light of this, eXplainable Artificial Intelligence (XAI) methodologies and tools can be effectively embedded within an AI-based malware detection process in order to make more understandable the produced results. In this paper, we propose a XAI framework for behavioral malware detection problems and evaluate the usefulness of four XAI methods (SHAP, LIME, LRP and Attention mechanism) on three datasets with different size, sequence length and number of classes, by which we could evaluate the strengths and weaknesses – from effectiveness and efficiency point of views – of recurrent deep architectures (i.e. Long-Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) models), and their applicability in the modern Cyber Security (CS) scenarios.

Despite the vast potential for insights and value present in Electronic Health Records (EHRs), it is challenging to fully leverage all the available information, particularly that contained in the free-text data written by clinicians describing the health status of patients. The utilization of Named Entity Recognition and Linking tools allows not only for the structuring of information contained within free-text data, but also for the integration with medical ontologies, which may prove highly beneficial for the analysis of patient medical histories with the aim of forecasting future medical outcomes, such as the diagnosis of a new disorder. In this paper, we propose MedTKG, a Temporal Knowledge Graph (TKG) framework that incorporates both the dynamic information of patient clinical histories and the static information of medical ontologies. The TKG is used to model a medical history as a series of snapshots at different points in time, effectively capturing the dynamic nature of the patient's health status, while a static graph is used to model the hierarchies of concepts extracted from domain ontologies. The proposed method aims to predict future disorders by identifying missing objects in the quadruple 〈s, r, ?, t 〉, where s and r denote the patient and the disorder relation type, respectively, and t is the timestamp of the query. The method is evaluated on clinical notes extracted from MIMIC-III and demonstrates the effectiveness of the TKG framework in predicting future disorders and of medical ontologies in improving its performance.

Deep learning, owing to its transformative influence across a myriad of sectors, has recently made its foray into the legal domain, instigated by the surge in digitization. Among the multitude of applications in this space, legal document classification emerges as a pivotal yet complex undertaking. Legal texts, characterized by unique domain-centric semantics and intricate linguistic patterns, necessitate precision-driven classification systems for numerous practical implications. This paper illuminates the challenges and opportunities in automating the classification of European Union (EU) legal documents, emphasizing the interrelationships among statutes and the hierarchical nature of legal references. In this context, we introduce a novel graph data modeling technique that adeptly marries content-centric indicators with the relational dynamics inherent among diverse legal documents. Central to our approach is a framework that melds text embeddings with graph neural networks for the classification of legal documents aligned with their subject-based directories. Empirical evaluations on the EU law dataset underline the efficacy of our model across varying granularities, from general thematic categories to intricate subtopics. This endeavor not only augments the comprehensibility and accessibility of EU jurisprudence but also holds significant implications across regulatory compliance, legal research, and policy formulation, underscoring the potential of deep learning in reshaping legal paradigms.

Recent years have seen an exponential growth (+98% in 2022 w.r.t. the previous year) of the number of research articles in the few-shot learning field, which aims at training machine learning models with extremely limited available data. The research interest toward few-shot learning systems for Named Entity Recognition (NER) is thus at the same time increasing. NER consists in identifying mentions of pre-defined entities from unstructured text, and serves as a fundamental step in many downstream tasks, such as the construction of Knowledge Graphs, or Question Answering. The need for a NER system able to be trained with few-annotated examples comes in all its urgency in domains where the annotation process requires time, knowledge and expertise (e.g., healthcare, finance, legal), and in low-resource languages. In this survey, starting from a clear definition and description of the few-shot NER (FS-NER) problem, we take stock of the current state-of-the-art and propose a taxonomy which divides algorithms in two macro-categories according to the underlying mechanisms: model-centric and data-centric. For each category, we line-up works as a story to show how the field is moving toward new research directions. Eventually, techniques, limitations, and key aspects are deeply analyzed to facilitate future studies.

In Biomedical Named Entity Recognition (BioNER), the use of current cutting-edge deep learning-based methods, such as deep bidirectional transformers (e.g. BERT, GPT-3), can be substantially hampered by the absence of publicly accessible annotated datasets. When the BioNER system is required to annotate multiple entity types, various challenges arise because the majority of current publicly available datasets contain annotations for just one entity type: for example, mentions of disease entities may not be annotated in a dataset specialized in the recognition ofdrugs, resulting in a poor ground truth when using the two datasets to train a single multi-task model. In this work, we proposeTaughtNet, a knowledge distillationbased framework allowing us to fine-tune a single multi-task student model by leveraging both the ground truth and the knowledge of single-taskteachers. Our experiments on the recognition of mentions of diseases, chemical compounds and genes show the appropriateness and relevance of our approach w.r.t. strong state-of-the-art baselines in terms of precision, recall and F1 scores. Moreover, TaughtNet allows us to train smaller and lighter student models, which may be easier to be used in real-world scenarios, where they have to be deployed on limitedmemory hardware devices and guarantee fast inferences, and shows a high potential to provide explainability. We publicly release both our code on github1 and our multi-task model on the huggingface repository.

In recent years, we have witnessed an ever wider spread of multimedia streaming platforms (e.g., Netflix, Spotify, and Amazon). Hence, it has become more and more essential to provide such systems with advanced recommendation facilities, in order to support users in browsing these massive collections of multimedia data according to their preferences and needs. In this context, the modeling of entities and their complex relationships (e.g., users listening to topic-based songs or authors creating different releases of their lyrics) represents the key challenge to improve the recommendation and maximize the users' satisfaction. To this end, this is the first study to leverage the high representative power of hypergraph data structures in combination with modern graph machine learning techniques in the context of music recommendation. Specifically, we propose hypergraph embeddings for music recommendation (HEMR), a novel framework for song recommendation based on hypergraph embedding. The hypergraph data model allows us to represent seamlessly all the possible and complex interactions between users and songs with the related characteristics; meanwhile, embedding techniques provide a powerful way to infer the user-song similarities by vector mapping. We have experimented the effectiveness and efficiency of our approach with respect to the state-of-the-art most recent music recommender systems, exploiting the Million Song dataset. The results show that HEMR significantly outperforms other state-of-the-art techniques, especially in scenarios where the cold-start problem arises, thus making our system a suitable solution to embed within a music streaming platform.

With the recent COVID-19 outbreak, we have assisted to the development of new epidemic models or the application of existing methodologies to predict the virus spread and to analyze how the different lock-down strategies can effectively influence the epidemic diffusion. In this paper, we propose a novel machine learning based framework able to estimate the parameters of any epidemiological model, such as contact rates and recovery rates, based on static and dynamic features of places. In particular, we model mobility data through a graph series whose spatial and temporal features are investigated by combining Graph Convolutional Neural Networks (GCNs) and Long short-term memories (LSTMs) in order to infer the parameters of SIR and SIRD models. We evaluate the proposed approach using data related to the COVID-19 dynamics in Italy and we compare the forecasts of the trained model with available data about the epidemic spread.