Machine learning (ML) holds great promise for transforming healthcare, by enhancing diagnostic accuracy, enabling earlier interventions, and refining treatment strategies. Yet, the opacity of many ML models remains a major barrier to clinical adoption. Explainable Artificial Intelligence (XAI) is often proposed as a way to address this challenge, but we still know little about how different stakeholders perceive and engage with XAI in practice. This study explored how three XAI techniques, Local SHAP, Global SHAP, and Attention Mechanism, shape trust, usability, and decision-making across diverse user groups. We conducted semi-structured interviews with 20 participants, including clinical experts, data scientists, and members of the public, using real-world ML prediction tasks in two healthcare domains: sepsis-related mortality and psychology. Participants reviewed visualizations of model explanations and shared their perspectives in guided discussions. Through thematic analysis, five key themes emerged. Participants stressed that explainability must be paired with perceived reliability, contextual relevance, and human oversight to build trust. Local SHAP was generally seen as the most intuitive method, while Global SHAP and Attention Mechanism were often perceived as too abstract or technical. Across all groups, AI was viewed as a decision-support tool, not a replacement for clinical judgment. Successful integration of XAI was seen as dependent on adaptable explanations, integration into existing workflows, appropriate training, and interdisciplinary collaboration. Our findings highlight the need for human-centered design and institutional readiness in the development of trustworthy, usable XAI systems for healthcare.
Evaluating machine learning (ML) model bias is key to building trustworthy and robust ML systems. Counterfactual Fairness (CF) audits allow the measurement of bias of ML models with a causal framework, yet their conclusions rely on a single causal graph that is rarely known with certainty in real-world scenarios. We propose CF with Graph Uncertainty (CF-GU), a bias evaluation procedure that incorporates the uncertainty of specifying a causal graph into CF. CF-GU (i) bootstraps a Causal Discovery algorithm under domain knowledge constraints to produce a bag of plausible Directed Acyclic Graphs (DAGs), (ii) quantifies graph uncertainty with the normalized Shannon entropy, and (iii) provides confidence bounds on CF metrics. Experiments on synthetic data show how contrasting domain knowledge assumptions support or refute audits of CF, while experiments on real-world data (COMPAS and Adult datasets) pinpoint well-known biases with high confidence, even when supplied with minimal domain knowledge constraints.
Missing data represents one of the most ubiquitous data quality issues, and also one of the most impactful on machine learning (ML) pipelines. Indeed, not only most commonly applied ML methods cannot directly employ incomplete data, but also the techniques employed to manage this issue can impact on the performance and evaluation of ML models. Among such techniques to manage missing data, imputation, that is filling in the missing values using information from the observed data, remains among the most popular and effective in practice. Yet, from a theoretical point of view, it is still not clear under which conditions it is possible to learn effectively after imputation. In this article we address this gap by studying learnability under imputation in the framework of statistical learning theory. After giving a general definition of learnability under imputation, we show three main contributions: 1) we introduce a novel stability condition, called noise risk stability, which we prove to be both sufficient and, under weak assumptions, necessary for learnability under imputation; 2) we show that a large class of ML models (including linear and kernel methods) satisfies noise risk stability; 3) we characterize the learning-theoretic properties of two common imputation methods (constant and regression imputation). Our results set the stage for a rigorous study of imputation and missing data management in the framework of statistical learning theory, by also describing relevant open questions.
The goal of learning to hash (L2H) is to derive data-dependent hash functions from a given data distribution in order to map data from the input space to a binary coding space. Despite the success of L2H, two observations have cast doubt on the source of the power of L2H, i.e., learning. First, a recent study shows that even using a version of locality sensitive hashing functions without learning achieves binary representations that have comparable accuracy as those of L2H, but with less time cost. Second, existing L2H methods are constrained to three types of hash functions: thresholding, hyperspheres, and hyperplanes only. In this paper, we unveil the potential of Voronoi diagrams in hashing. Voronoi diagram is a suitable candidate because of its three properties. This discovery has led us to propose a simple and efficient no-learning binary hashing method, called Voronoi Diagram Encoded Hashing (VDeH), which constructs a set of hash functions through a data-dependent similarity measure and produces independent binary bits through encoded hashing. We demonstrate through experiments on several benchmark datasets that VDeH achieves superior performance and lower computational cost compared to existing state-of-the-art methods under the same bit length. Our code is available at: https://github.com/Phantom-Det /VDeH/tree/main.
Non-degenerate directed hypergraphs, i.e., directed hypergraphs where a node cannot be both in the tail and the head of a hyperedge, model important scenarios, from contact networks for analyzing the spread of information or diseases, to bill cosponsoring graphs for studying the bipartisanship of elected representatives. Existing null models for dihypergraphs allow degeneracy, and most samples drawn from them are degenerate, even when the starting network is not, making these models unrealistic in many cases. An inappropriate null model may lead to wrongly accepting/rejecting a hypothesis when performing statistical hypothesis testing. We discuss null models for non-degenerate dihypergraphs, including some novel ones, and present DiNgHy, a suite of Markov-Chain-Monte-Carlo (MCMC) algorithms to sample from them. DiNgHy is the first correct MCMC sampler for the previously proposed configuration model for non-degenerate directed hypergraphs. The Markov chain underlying our algorithm is not irreducible in general, so we give mild sufficient conditions for irreducibility. We show that existing methods cannot be used to sample from our null models, and evaluate our algorithms on real and artificial dihypergraphs, comparing the results of hypothesis tests when using our null models versus existing ones that allow degeneracy, and measuring their empirical mixing time.
Graph learning for urban region modeling has gained significant attention for leveraging multi-modal data to generate region representations for downstream task prediction. However, existing models face two key limitations: (1) they primarily adopt a global perspective, overlooking the joint modeling of both local and global aspects, and (2) they rely on redundant, low-information nodes, leading to suboptimal region representations. To address these challenges, we propose GraphJCL, a dual-perspective framework that models both local and global perspectives. Specifically, GraphJCL first constructs local graphs for individual regions and a global graph encompassing all regions, integrating POI, taxi flow, remote sensing, street view, and road network data. Additionally, GraphJCL employs specialized message-passing mechanisms to efficiently capture both local and global graph node representations. Furthermore, GraphJCL incorporates entropy-optimized graph node pruning, retaining only the most informative nodes to enhance final region representations. To ensure the effectiveness of the designed dual-perspective graph framework, GraphJCL introduces a joint contrastive learning approach, optimizing region representations through geography-driven, entropy-optimized, and mutual information-based optimization techniques. Extensive experiments on two real-world datasets across five modalities demonstrate that GraphJCL consistently outperforms state-of-the-art methods on three tasks, validating its flexibility and effectiveness.
Low visibility is a critical factor affecting aviation and transportation safety, often leading to operational disruptions, delays, and potential hazards. Weather phenomena, such as fog, rain, and snow, significantly contribute to reducing visibility, making accurate prediction essential for mitigating risks. Conventional forecasting methods with time-series visibility and meteorological data often struggle with data imbalance and censored data issues, which impact forecasting accuracy, particularly in the low visibility range. In this paper, we propose a new approach by employing Censored Quantile Regression Neural Network and Light Gradient-Boosting Machine to forecast visibilities in the low and high visibility ranges and combining the forecast values by using a probabilistic classifier model built with Logistic Regression. We show the effectiveness of the proposed approach by performing experiments with two datasets of observed and forecast meteorological data from Japan and evaluating it in terms of forecasting errors and the accuracy of forecasting of low visibility. Experimental results suggest that our approach is well-suited to forecast low visibility with high accuracy up to 24 h ahead.
Artificial Neural Networks (NNs) are unable to learn tasks continually using a single model, which leads to forgetting old knowledge, known as catastrophic forgetting. This is one of the shortcomings that usually plague intelligent systems based on NN models. Federated Learning (FL) is a decentralized approach to training machine learning models on multiple local clients without exchanging raw data. A paradigm that handles model learning in both settings, federated and continual, is known as Federated Continual Learning (FCL). In this work, we propose a novel FCL algorithm, called FedCluLearn, which uses a stream micro-cluster indexing scheme to deal with catastrophic forgetting. FedCluLearn interprets the federated training process as a stream clustering scenario. It stores statistics, similar to micro-clusters in stream clustering algorithms, about the learned concepts at the server and updates them at each training round to reflect the current local updates of the clients. FedCluLearn uses only active concepts in each training round to build the global model, meaning it temporarily forgets the knowledge that is not relevant to the current situation. In addition, the proposed algorithm is flexible in that it can consider the age of local updates to reflect the greater importance of more recent data. The proposed FCL approach has been benchmarked against three baseline algorithms by evaluating its performance in several control and real-world data experiments. The implementation of FedCluLearn and the experimental results are available at https://github.com/milenaangelova1/FedCluLearn .
This paper presents BellatrExplorer, a dashboard application to interactively explore random forest predictions on the individual instance level. The application is inspired by the recently proposed local interpretability toolbox Bellatrex, that exploits the internal random forest structure to extract 1-3 prototype rules that act as a surrogate model for an instance of interest. BellatrExplorer is aimed at expert users trying to better understand the behavior of their random forest in a specific application, and could allow to uncover potential biases or artifacts arising in model training. Currently, the tool supports random forests for binary classification, regression, and survival analysis tasks. It features (1) intuitive exploration of univariate predictive counterfactuals, (2) analysis of decision tree rules to the individual split level, and (3) a visualisation of the rules extracted by Bellatrex that allow to assess the local interpretation at a glance. The tool is available at https://github. com/robbedhondt/BellatrExplorer/ and a demonstration video can be found at https://itec.kuleuven-kulak.be/bellatrexplorer/
The high dimensionality and variability of Computational Fluid Dynamics (CFD) data pose a significant challenge for Machine Learning (ML) models. The only solutions in the literature addressing inference from CFD flow fields are based on expert-driven features, which consist of fluid dynamic quantities averaged on specific regions of the entire computational domain. However, using handcrafted features can limit the scalability and portability of existing methods, and result in the loss of critical flow field information that might be essential for capturing non-linear patterns inherent in the CFD data. We propose a method to replace handcrafted features with features defined on regions obtained by clustering. Our approach combines: i) physics-based clustering, to identify meaningful regions within the flow field, ii) cluster-based feature extraction, to capture localized fluid dynamics properties, and iii) set-learning models to process the extracted information. Our solution allows integrating physics-based modeling with ML, and provides a portable and flexible pipeline capable of effectively dealing with the variability and dimensionality of CFD flow fields. We validate our method on publicly available CFD datasets (from the aerospace domain) and apply it to a realistic scenario, that is, the classification of pathologies in real 3D human upper airways extracted from CT scans, acquired in collaboration with a medical hospital. Experimental results demonstrate the accuracy and scalability of our method, and highlight its potential for leveraging CFD data in ML frameworks for other scientific and engineering applications.
Graph Neural Networks (GNNs) have emerged as the state-of-the-art method for graph-based learning tasks. However, training GNNs at scale remains challenging, limiting the exploration of more sophisticated GNN architectures and their application to large real-world graphs. In distributed GNN training, communication overhead and waiting times have become major performance bottlenecks. To address these challenges, we propose PipeQS, an adaptive quantization and staleness-aware pipeline distributed training system for GNNs. PipeQS dynamically adjusts the bit-width of message quantization and manages staleness to reduce both communication overhead and communication waiting time. By detecting pipeline bottlenecks caused by synchronization and utilizing cached communication to bypass message delays, PipeQS significantly improves training efficiency. Experimental results validate the effectiveness of PipeQS, showing up to an 8.3 × improvement in throughput while maintaining full-graph accuracy. Furthermore, our theoretical analysis demonstrates fast convergence at a rate of O(T^ - 1/2) , where T is the total number of training epochs. PipeQS achieves a well-balanced trade-off between training speed and accuracy, significantly reducing training time without compromising performance. The code is available at https://github.com/suupahako/PipeQS-code
We present VisualTreeSearch, a fully-deployed system for visualizing and understanding web agent test-time scaling. While test-time search algorithms substantially improve web agent success rates, they remain confined to research contexts with limited practical deployment. Our system bridges this gap with three key contributions: (1) a production-ready solution with cloud-based architecture, (2) an efficient API-based state reset mechanism that reduces state reset time from 50 to 2 s, and (3) an interactive web UI that transparently demonstrates the agent's decision-making process. VisualTreeSearch provides an intuitive framework for both researchers and users to understand tree search execution in web agents.
Wearable devices allow collecting data at an individual level, which can be used to propose an unseen degree of personalization for a broad domain of applications. For instance, we focus on electrochromic frames that allow to manually change the lens' tint, or automatically, based on an ambient light sensor. We aim to use the user's interactions with his frame to adapt this automatic mode to better consider his preferences. From a technical standpoint, this is a difficult task, as prediction and estimation cannot be done separately. That is why we approach this industrial problem from a reinforcement learning perspective: a policy must control the tint class in such a way that the number of user interactions is minimized. A particularity of this problem is that there is an inherent notion of order between the finite proposed tint classes, as some are darker than others. The usual Boltzmann parametrization does not account for this. Thus, we develop and implement policy gradient methods for ordinal policies. Using a simulation setting, we show that ignoring the ordinal structure of the response variables yields a suboptimal strategy. Additionally, we tested this technique with real users in controlled conditions; as the tint-control mode updated, the number of user interactions decreased. At last, using ordinal policies can be adapted to a deep reinforcement learning context, solving classic problems with continuous actions using discretization of this space.
Graph Transformers (GTs) address the locality limitation of traditional GNNs, which aggregate only local neighbor information, by leveraging global attention. However, they suffer from two significant issues: neglecting community structures and information oversqueezing. In this paper, we first identify these two problems and propose a Community-Aware Graph Transformer (CoGT) to solve them. CoGT introduces a novel node-community-global hierarchical aggregation framework. This design preserves community-level semantics while reducing the volume of aggregated information, alleviating the oversqueezing problem. CoGT first employs a two-stage positional encoding to identify latent communities and enhance semantic consistency. Then, a hierarchical and parallel transformer computation method based on community representations facilitates global information interaction. Furthermore, we enable community-wise parallel attention computation, improving computational efficiency. Experimental results demonstrate that CoGT outperforms existing methods across multiple real-world datasets.
Gradient boosting is a widely used algorithm for fitting sparse additive models over flexible classes of basis functions. Despite its popularity, the performance of gradient boosting as an approximation algorithm to the empirical risk minimizing model with a specific number k of selected basis functions is poorly understood. We provide a theoretical lower bound of 1/2 - 1/(4k - 2) on the worst-case approximation ratio for the risk reduction that gradient boosting achieves relative to the optimal model when both are limited to k terms. This result reveals an inherent limitation in boosting's ability to approximate the best possible sparse additive model, raising the question of how tight and representative this bound is in practice. To empirically answer this question, we employ mixed integer programming (MIP) to approximate the optimal additive models on 21 real datasets. The experimental results do not show larger gaps than the theoretical analysis, indicating that the theoretical lower bound is tight. Moreover, for twelve datasets, the approximation gaps are of the same order of magnitude as the theoretical lower bound, which shows the representativeness of the theoretical bound. To that end, the study also has the practical implication that the presented MIP approach frequently offers notable improvements over gradient boosting.
Knowledge Graph Completion (KGC) aims to predict missing triples in a graph based on known relationships between entities. However, most KGC methods face the challenge of diversification representations among entities, making it difficult for models to link entities effectively. This article proposes a Joint Knowledge (Structure-Semantics) Diffusion Model (JKDM) to capture entity diversification relationships. By leveraging the probabilistic generative capabilities of diffusion models, JKDM generates diversification outputs that align with the distribution of target entities rather than producing a single deterministic result. Considering the insufficient structural information of sparse entities, which leads to their representations tending toward a smooth distribution, making it difficult for diffusion models to learn their probability distributions, we jointly enhance sparse entity representations using structural and semantic information. Structurally, a Dual-channel Graph Attention Network (DGAT) is introduced to capture structural embeddings of entities from different perspectives. Semantically, a contextual path strategy is applied to pre-trained language models (PLMs) to enrich entity semantics. Under the condition of joint embeddings, JKDM gradually denoises to generate the probability distribution of target entities. Experiments demonstrate that JKDM outperforms SOTA methods on the FB15k-237, WN18RR, and UMLS datasets, achieving improvements of 2.3
Scattering networks are deep convolutional architectures that use predefined wavelets for feature extraction and representation. They are mathematically well-understood, and have proven effective for classification tasks in limited training data scenarios, where traditional deep learning methods struggle. However, the opposite holds in larger data regimes, resulting in a performance gap between well-understood learning architectures and non-transparent yet highly effective paradigms. Our work addresses this gap on the domain of graphs by adapting the choice of diffusion operator that constructs the scattering network to the data, allowing better task-wise geometric representation. The resulting architecture preserves stability guarantees with respect to input perturbations. Continuous diffusion is applied in the learning process for more refined weight updates. Numerical experiments on benchmark datasets show that our approach consistently outperforms traditional graph scattering with predefined wavelets, expanding the scenarios where interpretable scattering architectures are competitive or superior to deep learning methods, and further reducing their aforementioned performance disparity.
Multimodal Learning is one of the many fields in Machine Learning where models leverage the combination of various modalities to enhance learning outcomes. However, modalities may differ in data representation and complexity, which can lead to learning imbalances during the training process. The time it takes for a certain modality to converge during training is a crucial metric to determine modality imbalance. Given differences in convergence rates, different modalities may harmfully interfere with each other's learning process when simultaneously trained, as is commonly done in a multimodal scenario. To mitigate this negative impact, we propose Alternating Multimodal Skip Training (AMST) where the training frequency is adjusted for each specific modality. This novel method not only improves performance in conventional multimodal models that learn with fused modalities but also enhances alternating models that train each modality separately. Additionally, it outperforms state-of-the-art models while reducing training times.
The compatibility of Federated Learning (FL) models with unseen Out-Of-Federation (OOF) centers remains a critical yet underexplored challenge, particularly when dealing with heterogeneous data. To address this gap, this study proposes a data-driven approach to assess the feasibility of applying an FL model to OOF centers. The case study explored is the prediction of diabetic retinopathy from multiple real-world, highly heterogeneous electronic health records. An FL XGBoost model (FL-XGB) is trained across five in-federation (IF) centers, showing an average test Area Under the ROC Curve (AUC) of 75.27 ≪ 0.001). The average discrepancy between the predicted and observed AUC values was 6 https://github.com/geronimaw/OFA4FL .
Stock prediction is hindered by data scarcity, and although existing data augmentation techniques have made significant strides, they often overlook the dynamic inter-stock interactions crucial for robust modeling. To address these challenges, we propose InterDiff, a diffusion-based framework that synthesizes realistic financial time series by dynamically modeling both intra- and inter-stock correlations. InterDiff employs hierarchical transformers to learn these correlations, encoding them into a guidance vector that steers a diffusion model via classifier-free guidance. This approach ensures that the synthetic data preserves fidelity while introducing controlled variability. Evaluations on CSI300 and CSI800 show that models trained on InterDiff-augmented data boost the information coefficient by 1.13–4.70
