Abstract Chest X-rays are an inexpensive and widely available imaging modality for diagnosing or monitoring a variety of medical conditions. Given their abundance, healthcare providers could greatly benefit from automated systems capable of screening healthy patients and supporting the diagnosis of pathological cases. Deep learning has become central to such decision-support systems, offering accurate and efficient image classification that can improve clinical workflows and reduce radiologist workload. However, despite the rapid evolution of general-purpose neural architectures, particularly attention-based models, their application to medical imaging remains constrained by limited incorporation of medical domain knowledge. Most existing attention mechanisms optimize only task-specific losses, disregarding crucial anatomical and lesion-location priors, which can hinder generalization and interpretability. In this work, we introduce a fully automated, attention-guided classification framework that integrates medical priors through an on-the-fly segmentation of the lungs, followed by a spatially aware attention loss that directs the network’s focus toward clinically relevant regions. The method requires minimal physician input—only a single annotated X-ray indicating potential lesion areas at initialization and generalizes effectively across patients without relying on absolute bounding-box coordinates. Gradient-based activation mapping is further employed to ensure alignment between attention and lesion-specific regions. Our approach is architecture-agnostic and integrates seamlessly into end-to-end pipelines. Experiments on two medical image datasets demonstrate that the proposed segmentation-enhanced attention loss improves both classification accuracy and representation interpretability compared to the standard cross-entropy loss. The code is available at: https://github.com/rcorizzo/cxr-segmentation-attention/ .

Abstract Anomaly detection provides an effective decision support capability in several real-world domains. One limitation of conventional approaches is their inability to preserve knowledge as models are constantly updated with recent data, leading to catastrophic forgetting. Continual learning approaches overcome this limitation by providing strategies that provide a trade-off between model stability and plasticity. However, to deal with concept-agnostic scenarios, transitions between tasks/concepts must be detected and provided as auxiliary information to the models. While change point detection methods are a natural fit, the most effective ones for complex and evolving data rely on choosing an appropriate distance measure. However, a fundamental knowledge gap in current research stands in how distance measures for change point detection impact models’ ability to adapt and perform over time as new concepts emerge from evolving data. In this paper, we address this issue by proposing a modular approach to identify transitions in concept-agnostic scenarios and investigating how different distances in change detection affect the predictive performance of anomaly detection models in continual learning scenarios. We perform experiments with different continual learning strategies and compare them with concept-incremental scenarios across multiple real-world datasets. Our key results highlight that it is feasible to perform concept-agnostic learning with a small decline in anomaly detection performance compared to concept-incremental. Moreover, this decline can be mitigated with proper selection of the distance measure for change detection. Finally, our results reveal that even moderately accurate identification of changes can lead to competitive anomaly detection performance.

The recently proposed xLSTM is a powerful model that leverages expressive multiplicative gating and residual connections, providing the temporal capacity needed for long-horizon forecasting and representation learning. This architecture has demonstrated success in time series forecasting, lossless compression, and even large-scale language modeling tasks, where its linear memory footprint and fast inference make it a viable alternative to Transformers. Despite its growing popularity, no prior work has explored xLSTM for anomaly detection. In this work, we fill this gap by proposing xLSTMAD, the first anomaly detection method that integrates a full encoder-decoder xLSTM architecture, purpose-built for multivariate time series data. Our encoder processes input sequences to capture historical context, while the decoder is devised in two separate variants of the method. In the forecasting approach, the decoder iteratively generates forecasted future values xLSTMAD-F, while the reconstruction approach reconstructs the input time series from its encoded counterpart xLSTMAD-R. We investigate the performance of two loss functions: Mean Squared Error (MSE), and Soft Dynamic Time Warping (SoftDTW) to consider local reconstruction fidelity and global sequence alignment, respectively. We evaluate our method on the comprehensive TSB-AD-M benchmark, which spans 17 real-world datasets, using state-of-the-art challenging metrics such as VUS-PR. In our results, xLSTM showcases state-of-the-art accuracy, outperforming 23 popular anomaly detection baselines. Our paper is the first work revealing the powerful modeling capabilities of xLSTM for anomaly detection, paving the way for exciting new developments on this subject. Our code is available at: https://github.com/Nyderx/xlstmad.

Memes possess a humorous intent, yet they can also be used for malicious purposes. Analysing meme data has the potential to enhance content monitoring, identify emerging topics, and support content moderation in online platforms. Memes also represent an interesting use case for multimodal machine learning, as they combine text and image data. In this study, we explored the linguistic characteristics and analysed the convergent themes of five meme classes through common word extraction. Moreover, we compared the effectiveness of various machine learning models, i.e., unimodal (text or image) and multimodal (early fusion, late fusion) in binary and multiclass meme classification tasks. Our results on a large meme dataset showed that memes heavily adhered to current affairs, demonstrated by the high frequency of topical words across meme classes. Regarding model accuracy, early fusion achieved superior accuracy over late fusion in meme classification. Binary models outperformed multi-class classification methods. However, fusion models did not consistently surpass the accuracy of independent text or image-based models.

Social media content is increasingly subject to hate speech towards specific demographic groups. In this context, machine learning approaches appear relevant to detect malicious contents and support moderators in mitigation initiatives. However, many of the existing approaches are exclusively focused on the analysis of textual contents. On the other hand, studies that address multiple data modalities often rely on a single feature representation and a fully supervised learning setting. In this paper, we tackle multimodal hate speech detection resorting to different learning settings (one-class learning and binary classification). We also investigate the effectiveness of multiple deep learning model backbones and language models to extract embedding feature representations for text and image modalities. Our experiments with a real-world hate speech dataset show that there is a significant performance gap between one-class learning and binary classification, and that the choice of embedding representations for image and text modalities can impact the detection performance for different predictive models.

Social media platforms are pivotal in information dissemination but also contribute to the rapid spread of misinformation, especially during high-impact events like natural disasters, terrorist attacks, and political unrest. While recent advances in multi-modal learning have enhanced misinformation detection by integrating features from various modalities (e.g., text, images), certain areas remain under-explored, particularly the use of event-based multi-modal data. This paper introduces a novel approach to misinformation detection on social media using an event-based multi-modal learning framework. Our method extends beyond traditional techniques by employing latent variable modeling to capture non-linear associations in event-based multi-modal data and to generate joint features between events for classification. This approach enhances misinformation detection and enables the contextual understanding of terms across different events. We provide a detailed analysis of our dataset preparation, methodology, and results, demonstrating the effectiveness of our framework on a widely-used dataset of tweets from high-impact events. The paper concludes with insights into potential enhancements and future directions in multi-modal misinformation detection.

Accurately identifying patients with Right Ventricular Dysfunction (RVD) is critical for timely diagnosis and treatment, yet it remains a significant challenge in clinical practice due to the complexity and variability of echocardiography text reports. In this study, we address this crucial problem by evaluating the effectiveness of recurrent neural networks (RNNs) in classifying patients with and without RVD from echocardiography text reports. We demonstrate that our model exhibits strong performance, with high recall and precision in identifying RVD, underscoring its potential utility in clinical settings. Furthermore, we enhance the interpretability of the framework by leveraging the Local Interpretable Model-Agnostic Explanations (LIME) framework. This approach allows us to extract and validate meaningful semantic information related to RVD, providing deeper insights into the model’s local predictive capabilities and improving transparency in decision-making.

Online hate speech proliferation has created a difficult problem for social media platforms. A particular challenge relates to the use of coded language by groups interested in both creating a sense of belonging for its users and evading detection. Coded language evolves quickly and its use varies over time. This paper proposes a methodology for detecting emerging coded hate-laden terminology. The methodology is tested in the context of online antisemitic discourse. The approach considers posts scraped from social media platforms, often used by extremist users. The posts are scraped using seed expressions related to previously known discourse of hatred towards Jews. The method begins by identifying the expressions most representative of each post and calculating their frequency in the whole corpus. It filters out grammatically incoherent expressions as well as previously encountered ones so as to focus on emergent well-formed terminology. This is followed by an assessment of semantic similarity to known antisemitic terminology using a fine-tuned large language model, and subsequent filtering out of the expressions that are too distant from known expressions of hatred. Emergent antisemitic expressions containing terms clearly relating to Jewish topics are then removed to return only coded expressions of hatred.

On the brink of the one-year anniversary since the public release of ChatGPT, scholarly research has directed their interest toward detection methodologies for machine-generated text. Different models have been proposed, including feature-based classification and detection approaches, as well as deep learning architectures, with a small portion of them integrating contextual information to enhance accurate predictions. Moreover, detection approaches explored thus far have focused primarily on English datasets, with limited attention given to the examination of similar methods in other languages. As a result, the applicability and efficacy of these methods in linguistically diverse contexts remains underexplored. In this paper, we present a one-class deep fusion model that considers both contextual text features derived from word embeddings and linguistic features to detect machine-generated texts in English and Spanish. Experimental results indicated that our model outperformed popular baseline one-class learning models in the detection task, presenting higher accuracy scores in the English dataset. Results are discussed in comparison to competing classifiers as well as the language biases found in detection models.

Social media content can present a number of threats, including misinformation and hate speech towards specific demographic groups. One challenge is to effectively discriminate between benign and malicious posts, given the massive amount of available content. In this context, predictive models for malicious content detection can be extremely valuable, leading to the automatic removal of posts and user accounts or content being flagged for subsequent moderation. However, some of the existing detection models are limited to the analysis of a single data modality. At the same time, most multi-modal approaches operate in a fully supervised learning setting that assumes the availability of labeled data for both benign and malicious content. In this paper, we fill this gap by proposing a multimodal one-class learning approach for malicious online content detection. Our approach leverages feature extraction, dimensionality reduction, and one-class learning models to analyze text and image data in online posts simultaneously. Models learn their decision function in the challenging scenario where only benign online content is used as training data, overcoming the limitations of a fully supervised setting. Our experiments with two real-world datasets containing misinformation and hate speech posts reveal the effectiveness of different combinations of one-class learning models and dimensionality reduction techniques.

Intrusion Detection System is a strategic analytical tool for the security of organizations and institutions. Among existing approaches, distributed and collaborative intrusion detection approaches are particularly effective since they combine data analysis from multiple sources to provide increased model robustness. Although many state-of-the-art approaches have the ability to adapt to evolving environments and incoming data, they are subject to catastrophic forgetting of past knowledge. At the same time, recent works in lifelong continual anomaly detection showcase the merit of simultaneous adaptation and knowledge retention. However, lifelong methods are thus far limited to the analysis of a single data source and do not provide distributed and collaborative learning capabilities. In this paper, we fill this gap by proposing a novel distributed continual learning intrusion detection framework with collaborative experience replay. The system is built from independent Detection Nodes and a Continual Learning Center. While the nodes are in charge of data selection and intrusion detection, the Continual Learning Center implements a collaborative replay strategy, performs model updates, and broadcasts the most recent model to the nodes. The separation of responsibilities allows for the decomposition of the system into task-oriented services, leading to a modular, flexible, and scalable architecture. An extensive evaluation involving popular network intrusion detection datasets shows the potential of our framework and the improvement in detection performance that can be achieved with the collaborative replay strategy.

Predictive models are essential in big data platforms to tackle the needs of several real-world applications. However, static models are known to be prone to obsolescence in dynamic environments. While concept drift detection represents a viable way to deal with this problem, it is scarcely explored in the context of big data streams. In this paper, we propose a distributed drift detection workflow based on the DDM algorithm paired with a predictive model. Our workflow updates the predictive model as soon as drifts are detected, adjusting to the most recent data characteristics. To enable the analysis of large-scale datasets, we leverage Pandas UDFs and Apache Spark, efficiently distributing this workload across multiple worker node instances. Our experiments on two real-world drift detection datasets show the positive results obtained in terms of Speedup, Scaleup, and a limited impact in detection delay in comparison to a single worker node instance.

Forecasting in real-world domains characterized by large-scale sensor data is a timely and challenging problem. Despite the widespread adoption of big data frameworks and platforms, the landscape of available scalable forecasting methods is restricted to few options, due to the inherent challenges for model training distribution that arise with more complex models, such as graph-based deep neural networks. In this paper, we propose a distributed forecasting approach for sensor networks based on a Graph Convolutional Long Short-Term Memory (GCN-LSTM) model architecture. Sub-models for each node of the sensor network are built separately, and trained independently on multiple worker nodes in a cluster environment, while preserving the spatio-temporal contribution provided by the global structure of the network. Job distribution and scheduling are carried out through a custom strategy leveraging the Horovod framework, performing synchronous training on GPUs available on multiple worker nodes. Experiments on multi-node renewable energy datasets show that our approach is effective in terms of both accuracy and scalability. Specifically, the execution of the distributed workflow does not negatively impact the forecasting performance of the model, and the speedup and scaleup achieved are close to the ideal values.

Developing scalable machine learning solutions that meet the demand of real-world applications is crucially important in domains such as cybersecurity, smart grids, and social networks. Despite the rising interest in the analysis of graph data using big data frameworks, existing works are usually limited to the adoption of conventional machine learning models, due to their simplicity and availability as off-the-shelf algorithms in popular libraries. As a result, resorting to more sophisticated models such as deep neural networks for graph analysis in a fully distributed learning setting is still an open challenge. In this paper, we propose a distributed workflow for node classification in graphs. We focus on graph attention networks, and devise a distributed model training approach leveraging the Apache Spark, GraphX, and Horovod frameworks. Our workflow consists of feature engineering, graph partitioning, model deployment, and model training stages, which take place in a fully distributed manner. Experimental results show that leveraging graph partitioning is a feasible strategy for distributed model training on multiple workers equipped with GPUs. Specifi-cally, the randomized and informed graph partitioning strategies analyzed in our experiments present satisfactory results in terms of both accuracy and scalability with two benchmark graph datasets for node classification.

Prolonged air leak is a complication arising from a collapsed lung which can lead to serious illness such as pneumonia and empyema, and patient suffering from indwelling chest tubes. Drainage of air and liquid from chest drains can be monitored and recorded using novel digital chest drainage devices. The collected data can be analyzed by predictive models, which can provide decision support in chest tube management. Despite the promising adoption of predictive models in this context, existing approaches are still in their infancy and are mostly based on autoregressive and conventional machine learning models. In this paper, we present a LSTM-based model architecture for air leak forecasting that is able to deal with non-linear dependencies among different features and contiguous time points. We devise a post-processing procedure that leverages predictions to suggest whether the patient could have their chest tube safely removed in the upcoming hours, and evaluate the results according to a medical protocol. Experimental results show that our model is able to outperform currently adopted models, in terms of both forecasting and classification performance, suggesting the feasibility of our approach for chest tube management.

Clouds at different altitudes play different roles in Earth’s climate. Comprehensive understanding of overlapping clouds is important for climate and weather prediction. The East Pacific region is where El Niño and La Niña originate and where multi-layer clouds frequently occur. The overlap of clouds at different altitudes in this region increases the classification complexity for cloud-based climatological studies. Unlike prior work in cloud layer classification that assumes single layer or two-layer of clouds, in this work, we consider multi-layer cloud classification with 8 cloud-level classes (clear-sky, high, middle, low, high+middle, high+low, middle+low, high+middle+low). We develop and analyze machine learning models on features extracted from satellite images from the East Pacific regions collected by GOES Advanced Baseline Imager (ABI). These are used to classify CloudSat/CALIPSO observed multi-layer clouds. Due to the imbalanced nature of the data, we investigate the adoption of conventional resampling methods, as well as deep learning methods with data augmentation. In our experiments, we utilize the random forest classifier and Multilayer perceptron classifier with data augmentation methods to reduce the class imbalance during training. With these approaches, we achieve a classification accuracy of 83.6% without exploiting any ancillary information.

Anomaly detection tools present the potential to enhance defense policies and protection against different types of threats, supporting public safety and national security. Lifelong anomaly detection showcases new and challenging scenarios in which models are challenged to automatically adapt to changing conditions without forgetting past knowledge. However, the presence of anomalies in incoming data may significantly impact the robustness of models in such scenarios. Although active learning strategies could be an asset to increase model longevity and robustness, they have never been explored in this context. In this paper, we propose an active lifelong anomaly detection framework for class-incremental scenarios that supports any memory-based experience replay method, any query strategy, and any anomaly detection model. While experience replay allows models to consolidate past knowledge and simultaneously adapt to new knowledge, an active learning module reduces the number of anomalies memorized in the replay buffer. We propose two strategies that automatically identify and remove additional data points that are likely to be anomalies based on the oracle’s feedback. Our experiments on popular host-based and network-based intrusion detection datasets show that our framework can improve the anomaly detection performance of models under low labeling budget constraints.

Recently, we have witnessed that deep learning-based approaches have been widely applied. Particularly, some applications involve data that are high dimensional, sparse or imbalanced, which are different from those applications with dense data processing, such as image classification and speech recognition, where deep learning-based approaches have been extensively studied. One of the main applications is the user-centric platform that consists of great deal of users, items and user generated tabular data which are quite high-dimensional. The characteristics of such data pose unique challenges to the adoption of deep learning in these applications, including modeling, training, and online serving, etc. More and more communities from both academia and industry have initiated the endeavors to solve these challenges. This workshop will provide a venue for both the research and engineering communities to discuss and formulate the challenges, utilize opportunities, and propose new ideas in the practice and theory of deep learning on high-dimensional, sparse and imbalanced data.

Recent advances in medical imaging and deep learning have enabled the efficient analysis of large databases of images. Notable examples include the analysis of computed tomography (CT), magnetic resonance imaging (MRI), and X-ray. While the automatic classification of images has proven successful, adopting such a paradigm in the medical healthcare setting is unfeasible. Indeed, the physician in charge of the detailed medical assessment and diagnosis of patients cannot trust a deep learning model’s decisions without further explanations or insights about their classification outcome. In this study, rather than relying on classification, we propose a new method that leverages deep neural networks to extract a representation of images and further analyze them through clustering, dimensionality reduction for visualization, and class activation mapping. Thus, the system does not make decisions on behalf of physicians. Instead, it helps them make a diagnosis. Experimental results on lung images affected by Pneumonia and Covid-19 lesions show the potential of our method as a tool for decision support in a medical setting. It allows the physician to identify groups of similar images and highlight regions of the input that the model deemed important for its predictions.

Detecting relevant changes in dynamic time series data in a timely manner is crucially important for many data analysis tasks in real-world settings. Change point detection methods have the ability to discover changes in an unsupervised fashion, which represents a desirable property in the analysis of unbounded and unlabeled data streams. However, one limitation of most of the existing approaches is represented by their limited ability to handle multivariate and high-dimensional data, which is frequently observed in modern applications such as traffic flow prediction, human activity recognition, and smart grids monitoring. In this paper, we attempt to fill this gap by proposing WATCH, a novel Wasserstein distance-based change point detection approach that models an initial distribution and monitors its behavior while processing new data points, providing accurate and robust detection of change points in dynamic high-dimensional data. An extensive experimental evaluation involving a large number of benchmark datasets shows that WATCH is capable of accurately identifying change points and outperforming state-of-the-art methods.

Structural concept complexity, class overlap, and data scarcity are some of the most important factors influencing the performance of classifiers under class imbalance conditions. When these effects were uncovered in the early 2000s, understandably, the classifiers on which they were demonstrated belonged to the classical rather than Deep Learning categories of approaches. As Deep Learning is gaining ground over classical machine learning and is beginning to be used in critical applied settings, it is important to assess systematically how well they respond to the kind of challenges their classical counterparts have struggled with in the past two decades. The purpose of this paper is to study the behavior of deep learning systems in settings that have previously been deemed challenging to classical machine learning systems to find out whether the depth of the systems is an asset in such settings. The results in both artificial and real-world image datasets show that these settings remain mostly challenging for Deep Learning systems. Deeper architectures help with structural concept complexity but not with data scarcity and class overlap.

Predicting the output power of renewable energy production plants distributed on a wide territory is a really valuable goal, both for marketing and energy management purposes. Vi-POC (Virtual Power Operating Center) project aims at designing and implementing a prototype which is able to achieve this goal. Due to the heterogeneity and the high volume of data, it is necessary to exploit suitable Big Data analysis techniques in order to perform a quick and secure access to data that cannot be obtained with traditional approaches for data management. In this paper, we describe Vi-POC -- a distributed system for storing huge amounts of data, gathered from energy production plants and weather prediction services. We use HBase over Hadoop framework on a cluster of commodity servers in order to provide a system that can be used as a basis for running machine learning algorithms. Indeed, we perform one-day ahead forecast of PV energy production based on Artificial Neural Networks in two learning settings, that is, structured and non-structured output prediction. Preliminary experimental results confirm the validity of the approach, also when compared with a baseline approach.

The problem of accurately predicting the energy production from renewable sources has recently received an increasing attention from both the industrial and the research communities. It presents several challenges, such as facing with the rate data are provided by sensors, the heterogeneity of the data collected, power plants efficiency, as well as uncontrollable factors, such as weather conditions and user consumption profiles. In this paper we describe Vi-POC (Virtual Power Operating Center), a project conceived to assist energy producers and decision makers in the energy market. In this paper we present the Vi-POC project and how we face with challenges posed by the specific application. The solutions we propose have roots both in big data management and in stream data mining.