Abstract
The increase in complexity of deep learning models demands explanations that can be obtained with methods like Grad-CAM. This method computes an importance map for the last convolutional layer relative to a specific class, which is then upsampled to match the size of the input. However, this final step assumes that there is a spatial correspondence between the last feature map and the input, which may not be the case. We hypothesize that, for models with large receptive fields, the feature spatial organization is not kept during the forward pass, which may render the explanations devoid of meaning. To test this hypothesis, common architectures were applied to a medical scenario on the public VinDr-CXR dataset, to a subset of ImageNet and to datasets derived from MNIST. The results show a significant dispersion of the spatial information, which goes against the assumption of Grad-CAM, and that explainability maps are affected by this dispersion. Furthermore, we discuss several other caveats regarding Grad-CAM, such as feature map rectification, empty maps and the impact of global average pooling or flatten layers. Altogether, this work addresses some key limitations of Grad-CAM which may go unnoticed for common users, taking one step further in the pursuit for more reliable explainability methods.

Abstract
Recommendation systems play a crucial role in modern e-commerce and streaming services. However, the limited availability of public datasets hampers the rapid development of more efficient and accurate recommendation algorithms within the research community. This work introduces a stream-based data generator designed to generate user preferences for a set of items while accommodating progressive changes in user preferences. The underlying principle involves using user/item embeddings to derive preferences by exploring the proximity of these embeddings. Whether randomly generated or learned from a real finite data stream, these embeddings serve as the basis for generating new preferences. We investigate how this fundamental model can adapt to shifts in user behavior over time; in our framework, changes correspond to alterations in the structure of the tripartite graph, reflecting modifications in the underlying embeddings. Through an analysis of real-life data streams, we demonstrate that the proposed model is effective in capturing actual preferences and the changes that they can exhibit over time. Thus, we characterize these changes and develop a generalized method capable of simulating realistic data, thereby generating streams with similar yet controllable drift dynamics.

Abstract
Health system performance assessment (HSPA) is essential for health planning and to improve population health. One of the HSPA domains is related to effectiveness, which can be represented considering different dimensions. Composite indicators can be used to summarize complex constructs involving several indicators. One example of such efforts is the Healthcare Access and Quality Index from the Global Burden of Diseases Study, in which different causes of mortality amenable to health care are summarized in this index through principal component analysis and exploratory factor analysis. While these approaches use the variance of the indicators, marginal improvement is not considered, that is, the distance to the best practice frontier. In this study we propose an innovative benefit-of-the-doubt approach to combine frontier analysis and composite indicators, using amenable mortality estimates for 188 countries. In particular, we include flexible aggregating weighting schemes and a robust and conditional approach. The dual formulation gives information on the peers and the potential mortality rate reduction targets considering the background conditions. In absolute terms, Andorra and high-income countries are the most effective regarding healthcare access and quality, while sub-Saharan African and South Asian countries are the least effective. North African and Middle Eastern countries benefit the most when epidemiological patterns, geographical proximity, and country development status are considered.

Abstract
As ESG regulations gain relevance, namely following the entry into force of EU's Corporate Sustainability Reporting Directive, there is an increased requirement for value chain transparency. Organizations that are heavily based on data, Artificial Intelligence or Machine Learning, may significantly contribute to the ESG profiles of their upstream clients through the resources they spend on data storage and processing, model training or model serving. However, the lack of standardized mechanisms to report ESG-relevant indicators from these organizations hinders effective integration into their clients' ESG disclosures. This paper is motivated by an identified need to facilitate ESG indicator reporting by ML organizations in the value chain, and proposes an architecture to automate the monitoring and reporting process. This architecture is based on the principle of observability, and on the integration and effective use of open-source tools to monitor data processing pipelines and MLOps. By implementing an automated observability-based monitoring framework, ML organizations can provide actionable ESG data that align with regulatory requirements, while reducing reporting overhead for clients. We address how this architecture can be seamlessly integrated into existing operational workflows of ML providers and their clients, enhancing transparency, accountability, and compliance. The proposed architecture ensures accurate, real-time reporting and creates a scalable foundation for ESG aligned innovation in Machine Learning and adjacent domains. © 2025 IEEE.

Abstract
Interferometric observations of gravitational microlensing events offer an opportunity for precise, efficient, and direct mass and distance measurements of lensing objects, especially those of isolated neutron stars and black holes. However, such observations have previously been possible for only a handful of extremely bright events. The recent development of a dual-field interferometer, GRAVITY Wide, has made it possible to reach out to significantly fainter objects and increase the pool of microlensing events amenable to interferometric observations by 2 orders of magnitude. Here, we present the first successful observation of a microlensing event with GRAVITY Wide and the resolution of microlensed images in the event OGLE-2023-BLG-0061/KMT-2023-BLG-0496. We measure the angular Einstein radius of the lens with subpercent precision, theta E = 1.280 +/- 0.009 mas. Combined with the microlensing parallax detected from the event light curve, the mass and distance to the lens are found to be 0.472 +/- 0.012 M circle dot and 1.81 +/- 0.05 kpc, respectively. We present the procedure for the selection of targets for interferometric observations and discuss possible systematic effects affecting GRAVITY Wide data. This detection demonstrates the capabilities of the new instrument, and it opens up completely new possibilities for the follow-up of microlensing events and future routine discoveries of isolated neutron stars and black holes.

Abstract
The proliferation of disinformation on social media platforms poses a significant challenge to the reliability of online information ecosystems and the protection of public discourse. This study investigates the role of emotional sequences in detecting intentionally misleading messages disseminated on social networks. To this end, we apply a methodological pipeline that combines semantic segmentation, automatic emotion recognition, and sequential pattern mining. Emotional sequences are extracted at the subsentence level, preserving each message's temporal order of emotional cues. Comparative analyses reveal that disinformation messages exhibit a higher prevalence of negative emotions, particularly fear, anger, and sadness, interspersed with neutral segments. Moreover, false messages frequently employ complex emotional progressions-alternating between high-intensity negative emotions and emotionally neutral passages-designed to capture attention and maximize engagement. In contrast, messages from reliable sources tend to follow simpler, more linear emotional trajectories, with a greater prevalence of positive emotions such as joy. Our dataset encompasses multiple categories of disinformation, enabling a fine-grained analysis of how emotional sequencing varies across different types of misleading content. Furthermore, we validate our approach by comparing it against a publicly available disinformation dataset, demonstrating the generalizability of our findings. The results highlight the importance of analyzing temporal emotional patterns to distinguish disinformation from verified content, reinforcing the value of integrating emotional sequences into machine learning pipelines to enhance disinformation detection. This work contributes to the growing body of research emphasizing the relationship between emotional manipulation and the virality of misleading content online.