Many machine learning algorithms for tabular data produce black-box models, which prevent users from understanding the rationale behind the model predictions. In their unconstrained form, graph neural networks fall into this category, and they have further limited abilities to handle heterogeneous data. To overcome these limitations, an approach is proposed, called IGNH (Interpretable Graph Neural Network for Heterogeneous tabular data), which handles both categorical and numerical features, while constraining the learning process to generate exact feature attributions together with the predictions. A large-scale empirical investigation is presented, showing that the feature attributions provided by IGNH align with Shapley values that are computed post hoc. Furthermore, the results show that IGNH outperforms two powerful machine learning algorithms for tabular data, Random Forests and TabNet, while competing favourably with XGBoost.
Current deep learning-based models for image classification are effective at making decisions, but their lack of transparency can be a significant concern in high-stakes settings. To address this issue, many state-of-the-art methods define (post-hoc) explanations as visual heatmaps or image segments deemed responsible for the classifiers’ outputs. However, the static nature of these explanations often fails to align with human explanatory practices. To obtain human-oriented explanations, we propose an alternative, novel form of dialogue-based interactive explanations for image classifiers: visual debates between two fictional players who interact to argue for and against the classifiers’ outputs. Specifically, in our method, the players propose arguments, which are (abstract) features drawn from classifiers’ latent representations and these arguments are countered by the opposing player. We present a realization of visual debates based on quantization for extracting arguments, recurrent networks for modelling player behaviour, and network dissection for argument visualization. Experimentally, we show that our visual debates satisfy the desiderata of dialecticity, convergence, and faithfulness.
Most techniques for explainable machine learning focus on a single modality for the explanations, e.g., using either feature or example attribution. A novel approach, called Hybrid Attribution Network (HAN), is proposed for providing multimodal explanations for image classification. The technique first extracts embeddings from a deep neural network (DNN), which are subsequently used by a KNN classifier to form predictions; example attributions can then be derived from the latter. Based on the example attributions, pixel attributions are further generated to provide complementary feature-level explanations. Results from an empirical investigation show that HAN may provide highly concentrated example attributions, i.e., the predictions can be explained with few training examples, without compromising predictive performance relative to the original deep neural network. Moreover, the pixel attributions are shown to enhance the interpretability of the predictions, by highlighting key pixels in the example attributions. An important finding from the empirical investigation is that the choice of layer to use for the embeddings may have a large impact on both the predictive performance and the generated explanations.
With the rise of neural network approaches for machine learning problems, the focus has shifted to learning deep concepts across multiple layers. Although single-hidden-layer networks are sufficient to arbitrarily approximate any function, multilayer networks have proven to be superior in predictive performance for many applications. Unlike for neural networks, in symbolic machine learning approaches like decision tree or rule learning algorithms, the benefits of hidden layers of learned intermediate concepts remain uncertain. In this work, we empirically investigate the potential gains of deep concepts for symbolic approaches from three perspectives. First, we compare the number of possible flat and deep Boolean expressions with limited complexity, underlining the higher expressive power of deep models in such a setting. Second, we use logic minimization algorithms to generate minimal flat and deep Boolean formulas for artificial Boolean classification problems with different numbers of attributes and training examples, showing under which circumstances the use of deep concepts can lead to noticeably less complex models. Third, we compare the predictive performance of flat and deep models with a fixed maximum complexity on these datasets. We interpret these results as evidence that encourages further investigation of algorithms for learning complexity-bounded deep rule sets.
The growing deployment of Machine Learning models has increased the demand for interpretability, leading to the development of various explainability methods. Yet, each method has its strengths and limitations, making it challenging to identify a one-size-fits-all solution. This paper introduces FIRE360, a multi-faceted and fast local explanation approach for tabular data. FIRE360 incorporates various explanation types (feature importances, rules, counterfactuals, exemplars, and counter-exemplars), offering a 360-degree model interpretability. FIRE360 also includes a dashboard with indicators for prediction reliability and explanation quality, such as fidelity and robustness. To ensure efficiency, FIRE360 uses GAN-generated synthetic data to approximate the original dataset and selects similar records to the target instance, avoiding costly on-demand neighborhood generation. Multiple local surrogate models are trained to capture different aspects of the data and black-box behavior. FIRE360 achieves higher fidelity than state-of-the-art methods, with improvements up to 95
The emergence of simulators powered by Large Language Models (LLMs) has enabled the realistic modeling of complex social phenomena while significantly reducing the costs and challenges of real-world data collection. Despite their promise, assessing the reliability of these simulators remains an open challenge. Existing validation methods often focus on isolated domains and operate at fixed levels of granularity, limiting their generalizability. In this work, we introduce simvale (simulator validation with latent embeddings), a generalizable, multidomain framework for quantitatively assessing LLM-based social simulators. simvale leverages both interpretable features and latent representations to yield global and local assessments about the fidelity of a simulator to real-world dynamics, or about the effects of controlled interventions. We demonstrate the effectiveness of simvale through a case study evaluating a simulator's ability to reproduce Online Social Network behavioral patterns and capture the impact of moderation interventions.
In today's digital world, user-generated reviews represent invaluable insights reflecting authentic experiences, preferences, and perceptions regarding products and services. Such reviews play a critical role both for consumers seeking informed purchasing decisions and businesses aiming to optimize their offerings and strategies. Recent advancements in machine learning, particularly the emergence of Large Language Models (LLMs) have significantly enhanced the processing and interpretation of this rich, unstructured textual data. Traditional platforms for comparing products and services focus primarily on structured specifications, often neglecting the detailed, experience-based information contained in user reviews. To address this limitation, we propose a novel framework that leverages Generative AI and advanced LLMs to extract and interpret user feedback, enabling more informed, experience-aware comparisons. Our approach involves three main phases: targeted review and metadata collection; topic modeling and sentiment classification using fine-tuned BERT models; and structured comparisons powered by user reviews, featuring attribute-level scores and natural language explanations generated by advanced GenAI tools such as GPT-4. Evaluated on real-world scenarios, including comparisons of similar Amazon products and nearby hotels, our framework outperforms traditional aggregation methods by generating more precise comparative scores and context-aware explanations.
Alzheimer's Disease (AD) globally represents one of the most prevalent and devastating neurodegenerative disorders. Recent advances in biomarker identification, neuroimaging, and diagnostic technologies offer promising tools to detect AD before significant symptoms appear. Artificial intelligence (AI) systems represent one of the most powerful tools for this purpose, since they are revolutionizing AD detection and management by analyzing complex patterns across multi-modal data. In this context, this paper introduces VERONECA (VErsatile and RObust approach for NEuroimaging and Clinical data-based Alzheimer's diagnosis), a novel AI method based on an advanced multi-modal boosting approach which promotes diversity among modalities through multi-armed bandits. The method offers: i) state-of-the-art performance in AD diagnosis, through effective integration of heterogeneous data; ii) specialized learners tailored to work with specific peculiarities of each modality; and iii) a robust strategy for handling missing modalities. Our experiments show that VERONECA effectively leverages complex multi-modal data to diagnose AD. Moreover, its intrinsically modular design facilitates the seamless integration of emerging modalities as well as the adaptation to evolving clinical standards. The code of the proposed method VERONECA is publicly available at https://github.com/code-paper-veroneca/veroneca.git
We study online recommendation of item sets that are high in both rating and diversity. We formalize our problem as a contextual bandit problem in which the reward of a recommended set is the rating sum of the member items plus the diversity of the set. In our problem setting, each item is represented as a feature vector (context) in a multi-dimensional metric space. Its rating is assumed to be the noisy value of some linear function of its feature vector, and the diversity of an item set is measured using the Sum diversity for the feature metric space. We develop an for this contextual bandit problem that guarantees a O(√(n)ln n)=Õ(√(n)) upper bound on (1/2)-regret. According to our experimental results on a benchmark dataset, the proposed algorithm outperforms existing algorithms in terms of not only our combined reward but also the reward of rating sum only.
Advanced Persistent Threats (APTs) represent a growing menace to modern digital infrastructure. Unlike traditional cyberattacks, APTs are stealthy, adaptive, and long-lasting, often bypassing signature-based detection systems. This paper introduces a novel framework for APT detection that unites deep learning, reinforcement learning (RL), and active learning into a cohesive, adaptive defense system. Our system combines autoencoders for latent behavioral encoding with a multi-agent ensemble of RL-based defenders, each trained to distinguish between benign and malicious process behaviors. We identify a critical challenge in existing detection systems: their static nature and inability to adapt to evolving attack strategies. To this end, our architecture includes multiple RL agents (Q-Learning, PPO, DQN, adversarial defenders), each analyzing latent vectors generated by an autoencoder. When any agent is uncertain about its decision, the system triggers an active learning loop to simulate expert feedback, thus refining decision boundaries. An ensemble voting mechanism, weighted by each agent's performance, ensures robust final predictions.
Anomaly detection in real-world data streams often struggles with concept drift, where evolving data distributions challenge algorithms to maintain a balance between accuracy, speed, stability, and plasticity. We present Adaptive Isolation Forest (AIF), a novel anomaly detection algorithm designed to effectively adapt to such changes in a resource-efficient and balanced manner. AIF's novelty can be found in the combination of a smart model update mechanism together with a newly developed MinTreeMaxMass (MTMM) criterion, which scores individual trees for replacement. Extensive evaluations on various benchmark datasets demonstrate that AIF significantly outperforms existing state-of-the-art streaming anomaly detection algorithms in terms of detection accuracy. Moreover, AIF achieves linear time and space complexities, providing a robust solution that maintains high accuracy and efficiency, balancing stability and plasticity in dynamic data streams.
Spectral clustering is a widely adopted method capable of identifying complicated cluster boundaries. However, traditional spectral clustering requires the definition of a predefined similarity metric for constructing the Laplacian matrix, a requirement that limits flexibility and adaptability. Instead of predefining this metric upfront as a fixed parametric function, we introduce a novel approach that learns the optimal parameters of a similarity function through parameter optimization. This optimizes a similarity function to assign high similarity values to data pairs with shared discriminative features and low values to those without such features. Previous methods that adapt similarity measures typically treat their parameters as hyperparameters or rely on non-convex optimization strategies. However, these approaches are not well-suited for unsupervised scenarios, as they depend heavily on initial conditions and require labeled data for validation, which is unavailable in such settings. In contrast, our method employs convex optimization to learn the parameters of the similarity metrics directly, rather than treating them as hyperparameters. This enables robust and reliable unsupervised learning, making our approach particularly well-suited for spectral clustering. We validate the effectiveness and adaptability of our method on several benchmark datasets, demonstrating superior performance compared to existing techniques.
A subgroup discovery-based method has recently been proposed to understand the behavior of models in the (original) feature space. The subgroups identified represent areas of feature space where the model obtains better or worse predictive performance when compared to the average test performance. For instance, in the marketing domain, the approach extracts subgroups such as: in groups of customers with higher income and who are younger, the random forest achieves higher accuracy than on average. Here, we propose a complementary method, Meta Subspace Analysis (MSA), MSA uses metalearning to analyze these subgroups in the metafeature space. We use association rules to relate metafeatures of the feature space represented by the subgroups to the improvement or degradation of the performance of models. For instance, in the same domain, the approach extracts rules such as: when the class entropy decreases and the mutual information increases in the subgroup data, the random forest achieves lower accuracy. While the subgroups in the original feature space are useful for the end user and the data scientist developing the corresponding model, the meta-level rules provide a domain-independent perspective on the behavior of the model that is suitable for the same data scientist but also for ML researchers, to understand the behavior of algorithms. We illustrate the approach with the results of two well-known algorithms, naive Bayes and random forest, on the Adult dataset. The results confirm some expected behavior of algorithms. However, and most interestingly, some unexpected behaviors are also obtained, requiring additional investigation. In general, the empirical study demonstrates the usefulness of the approach to obtain additional knowledge about the behavior of models.
Misinformation propagation in online networks involves multifaceted interactions between users, contents, and engagement mechanisms (likes, shares, comments). Addressing this issue entails both understanding how information spreads and identifying influential users driving the dissemination process. To tackle these challenges, this paper proposes a framework based on a Graph Attention Network model, applied to a heterogeneous graph representing social interactions and context-aware dynamics. Targeting the binary classification of real vs fake news, it offers insights into both propagation patterns and influential users in the dissemination process. A core contribution is the adoption of two post-hoc mechanisms for uncovering such users: uncertainty-based Active learning-like and GNN-Explainer. A detailed comparative analysis reveals that nodes where the model exhibits the highest confidence often lack rich content information; nevertheless, combining both high-confidence and content-rich nodes grasps complementary aspects and better aligns with influential users in information propagation. The framework is benchmarked against traditional centrality measures, widely used to identify influential users in social networks. A comparative evaluation on two heterogeneous, real-world, social networks confirms that the proposed method both achieves compelling accuracy in finding influential nodes and shows a potential to scale-up to densely-connected graphs on which classic approaches may fail.
Missing data is a common occurrence in data streams, where the feature values and label information are not assumed to be available. While acquiring feature values or labels may be possible, doing so is often associated with monetary and time costs. For instance, acquiring labels is time consuming and may require specialized experts or testing equipment to be available. Similarly, acquiring feature values may require access to restricted and or limited sources with limited availability. If one would be able to maximize performance and only acquire necessary information, budget could be saved. In this work we present a combined approach to acquiring feature values and labels under latency in a stream setting. Our method is tested on four data sets under various latencies and missingness conditions to evaluate its usefulness. We also present an ablation study on the impact of individual factors and their impact on the task performance.
Numerous studies have demonstrated various techniques for creating adversarial samples to exploit inherent or flawed behaviors in intelligent algorithms. However, many of these methods are intricate and may necessitate model- or data-specific information. This work introduces a novel approach to crafting adversarial samples, utilizing blind optimization algorithms combined with efficient population initialization methods. We apply this approach to evaluate the resilience of well-known recommendation algorithms in adversarial scenarios. Specifically, we conducted experiments using two recommendation systems featured in the Netflix Prize competition. Our primary objective was to inflate the ratings of a target movie, increasing the likelihood of it being recommended, by creating adversarial interactions (adversarial samples) within the network. Our evaluation results reveal that while the tested recommendation algorithms exhibit resilience to blind adversarial attacks, optimization-based methods can induce moderate shifts in predicted ratings, exposing subtle but exploitable vulnerabilities.
Self-supervised learning (SSL) offers a promising solution to the problem of label scarcity by leveraging large amounts of unlabeled data to learn transferable representations. However, using extensive unlabeled datasets can introduce substantial computational costs during pretext training and may include noisy or unrepresentative samples that degrade learning. In this work, we explore whether selecting a subset of the unlabeled examples available for the pretext task can reduce computation time while maintaining satisfactory performance of SSL methods for tabular data. In particular, we investigate whether estimates of the uncertainty with which predictions are made for unlabeled data (based on the available labeled data) can inform sampling and lead to superior embeddings. To answer these questions, we carry out large-scale experiments across 28 tabular benchmark datasets with TabNet and SCARF, on varying amounts of labeled data, and with a number of strategies for sampling unlabeled data. Our results show that reducing the pool of unlabeled data yields significant computational gains with only marginal reductions in performance.
Hierarchical document classification is essential for structuring large-scale textual corpora in domains such as digital libraries and academic repositories. While recent advances in large language models (LLMs) have opened new possibilities for text classification, their applicability to hierarchical settings under real-world constraints remains underexplored. This study investigates both generative and discriminative transformer-based models, evaluating their effectiveness across multiple inference strategies: zero-shot baseline, local fine-tuning, and a global approach using category-specific models. Experiments on two real-world hierarchical datasets provide a comprehensive comparison of classification accuracy, F1-macro scores, and inference times. The results highlight that, although generative LLMs can deliver competitive (yet variable) performance at higher levels of the hierarchy, their high inference costs hinder their use in time-sensitive applications. In contrast, fine-tuned discriminative models-particularly BERT-based architectures-consistently offer a more favorable trade-off between performance and efficiency.
Counterfactual explanations have been proposed to provide actionable insights for competitive black-box classifiers by suggesting modifications to alter undesirable prediction outcomes into desired ones. Despite the rapid development of counterfactual methods that provide insightful “what-if” explanations in various applications, most existing approaches fail to generate counterfactuals considering whether the contextual integrity of the original instance –such as demographic consistency or user-specific constraints– is explicitly controlled. This article introduces Context-Aware Counterfactual Explanations (CACTUS), a novel framework for generating feasible counterfactuals that can either preserve or change the user-defined contextual features of the original instance. We illustrate CACTUS by conducting the counterfactual search in a context-conditional latent space, using a composite β -VAE model that can learn enhanced disentanglement of context-related factors. Our experimental evaluation indicates that the context-preserving CACTUS outperforms state-of-the-art methods in maintaining contextual consistency, while remaining competitive in compactness, proximity, and validity metrics. Moreover, our qualitative analysis highlights that CACTUS generates plausible counterfactual explanations that can explicitly preserve or modify user-defined contexts in real-world scenarios.
Ensuring robustness in time series classification remains a critical challenge for safety-sensitive domains like clinical decision systems. While current evaluation practices focus on accuracy measures, they fail to address model stability under semantically meaningful input deformations. We propose tsMIST (Time Series Model Sensitivity Test), a novel morphing-based framework that systematically evaluates classifier resilience through controlled interpolation between adversarial class prototypes. By calculating the switchThreshold – defined as the minimal morphing distance required to flip predictions – our method reveals critical stability patterns across synthetic benchmarks with tunable class separation and 17 medical time series datasets. Key findings show convolutional architectures (ROCKET) maintain optimal thresholds near 50 48.2 ± 3.1% ), while feature-based models (Catch22) exhibit premature decision flips at 22.7 ≤ 12
