Search

In this paper, we introduce PASTA (Perceptual Assessment System for explanaTion of Artificial intelligence), a novel framework for a human-centric evaluation of XAI techniques in computer vision. Our first key contribution is a human evaluation of XAI explanations on four diverse datasets (COCO, Pascal Parts, Cats Dogs Cars, and MonumAI) which constitutes the first large-scale benchmark dataset for XAI, with annotations at both the image and concept levels. This dataset allows for robust evaluation and comparison across various XAI methods. Our second major contribution is a data-based metric for assessing the interpretability of explanations. It mimics human preferences, based on a database of human evaluations of explanations in the PASTA-dataset. With its dataset and metric, the PASTA framework provides consistent and reliable comparisons between XAI techniques, in a way that is scalable but still aligned with human evaluations. Additionally, our benchmark allows for comparisons between explanations across different modalities, an aspect previously unaddressed. Our findings indicate that humans tend to prefer saliency maps over other explanation types. Moreover, we provide evidence that human assessments show a low correlation with existing XAI metrics that are numerically simulated by probing the model., Comment: https://github.com/ENSTA-U2IS-AI/Dataset_XAI

Quantitative evaluation of echocardiography is essential for precise assessment of cardiac condition, monitoring disease progression, and guiding treatment decisions. The diverse nature of echo images, including variations in probe types, manufacturers, and pathologies, poses challenges for developing artificial intelligent models that can generalize across different clinical practice. We introduce EchoApex, the first general-purpose vision foundation model echocardiography with applications on a variety of clinical practice. Leveraging self-supervised learning, EchoApex is pretrained on over 20 million echo images from 11 clinical centres. By incorporating task-specific decoders and adapter modules, we demonstrate the effectiveness of EchoApex on 4 different kind of clinical applications with 28 sub-tasks, including view classification, interactive structure segmentation, left ventricle hypertrophy detection and automated ejection fraction estimation from view sequences. Compared to state-of-the-art task-specific models, EchoApex attains improved performance with a unified image encoding architecture, demonstrating the benefits of model pretraining at scale with in-domain data. Furthermore, EchoApex illustrates the potential for developing a general-purpose vision foundation model tailored specifically for echocardiography, capable of addressing a diverse range of clinical applications with high efficiency and efficacy.

Algorithmic Recourse (AR) aims to provide users with actionable steps to overturn unfavourable decisions made by machine learning predictors. However, these actions often take time to implement (e.g., getting a degree can take years), and their effects may vary as the world evolves. Thus, it is natural to ask for recourse that remains valid in a dynamic environment. In this paper, we study the robustness of algorithmic recourse over time by casting the problem through the lens of causality. We demonstrate theoretically and empirically that (even robust) causal AR methods can fail over time except in the - unlikely - case that the world is stationary. Even more critically, unless the world is fully deterministic, counterfactual AR cannot be solved optimally. To account for this, we propose a simple yet effective algorithm for temporal AR that explicitly accounts for time. Our simulations on synthetic and realistic datasets show how considering time produces more resilient solutions to potential trends in the data distribution.

In an Oberbeck-Boussinesq model, rigorously derived, which includes compressibility, one could expect that the onset of convection for the B\'enard problem occurs at a higher critical Rayleigh number. Since of the difficulties related to the new partial differential equations, with non constant coefficients and a non divergence-free velocity field, we show the increased stability of the rest state by exploring the related Lorenz approximation system.

Graph Neural Networks (GNNs) have emerged as the predominant paradigm for learning from graph-structured data, offering a wide range of applications from social network analysis to bioinformatics. Despite their versatility, GNNs face challenges such as lack of generalization and poor interpretability, which hinder their wider adoption and reliability in critical applications. Dropping has emerged as an effective paradigm for improving the generalization capabilities of GNNs. However, existing approaches often rely on random or heuristic-based selection criteria, lacking a principled method to identify and exclude nodes that contribute to noise and over-complexity in the model. In this work, we argue that explainability should be a key indicator of a model's quality throughout its training phase. To this end, we introduce xAI-Drop, a novel topological-level dropping regularizer that leverages explainability to pinpoint noisy network elements to be excluded from the GNN propagation mechanism. An empirical evaluation on diverse real-world datasets demonstrates that our method outperforms current state-of-the-art dropping approaches in accuracy, and improves explanation quality.

As Graph Neural Networks (GNNs) become more pervasive, it becomes paramount to build robust tools for computing explanations of their predictions. A key desideratum is that these explanations are faithful, i.e., that they portray an accurate picture of the GNN's reasoning process. A number of different faithfulness metrics exist, begging the question of what faithfulness is exactly, and what its properties are. We begin by showing that existing metrics are not interchangeable -- i.e., explanations attaining high faithfulness according to one metric may be unfaithful according to others -- and can be systematically insensitive to important properties of the explanation, and suggest how to address these issues. We proceed to show that, surprisingly, optimizing for faithfulness is not always a sensible design goal. Specifically, we show that for injective regular GNN architectures, perfectly faithful explanations are completely uninformative. The situation is different for modular GNNs, such as self-explainable and domain-invariant architectures, where optimizing faithfulness does not compromise informativeness, and is also unexpectedly tied to out-of-distribution generalization.

The advent of powerful neural classifiers has increased interest in problems that require both learning and reasoning. These problems are critical for understanding important properties of models, such as trustworthiness, generalization, interpretability, and compliance to safety and structural constraints. However, recent research observed that tasks requiring both learning and reasoning on background knowledge often suffer from reasoning shortcuts (RSs): predictors can solve the downstream reasoning task without associating the correct concepts to the high-dimensional data. To address this issue, we introduce rsbench, a comprehensive benchmark suite designed to systematically evaluate the impact of RSs on models by providing easy access to highly customizable tasks affected by RSs. Furthermore, rsbench implements common metrics for evaluating concept quality and introduces novel formal verification procedures for assessing the presence of RSs in learning tasks. Using rsbench, we highlight that obtaining high quality concepts in both purely neural and neuro-symbolic models is a far-from-solved problem. rsbench is available at: https://unitn-sml.github.io/rsbench.

Structured output prediction problems are ubiquitous in machine learning. The prominent approach leverages neural networks as powerful feature extractors, otherwise assuming the independence of the outputs. These outputs, however, jointly encode an object, e.g. a path in a graph, and are therefore related through the structure underlying the output space. We discuss the semantic loss, which injects knowledge about such structure, defined symbolically, into training by minimizing the network's violation of such dependencies, steering the network towards predicting distributions satisfying the underlying structure. At the same time, it is agnostic to the arrangement of the symbols, and depends only on the semantics expressed thereby, while also enabling efficient end-to-end training and inference. We also discuss key improvements and applications of the semantic loss. One limitations of the semantic loss is that it does not exploit the association of every data point with certain features certifying its membership in a target class. We should therefore prefer minimum-entropy distributions over valid structures, which we obtain by additionally minimizing the neuro-symbolic entropy. We empirically demonstrate the benefits of this more refined formulation. Moreover, the semantic loss is designed to be modular and can be combined with both discriminative and generative neural models. This is illustrated by integrating it into generative adversarial networks, yielding constrained adversarial networks, a novel class of deep generative models able to efficiently synthesize complex objects obeying the structure of the underlying domain., Comment: Preprint of Ch. 22 "Semantic Loss Functions for Neuro-Symbolic Structured Prediction" in "Compendium of Neurosymbolic Artificial Intelligence", https://ebooks.iospress.nl/ISBN/978-1-64368-406-2. arXiv admin note: substantial text overlap with arXiv:2201.11250, arXiv:2007.13197

Algorithmic Recourse aims to provide actionable explanations, or recourse plans, to overturn potentially unfavourable decisions taken by automated machine learning models. In this paper, we propose an interaction paradigm based on a guided interaction pattern aimed at both eliciting the users' preferences and heading them toward effective recourse interventions. In a fictional task of money lending, we compare this approach with an exploratory interaction pattern based on a combination of alternative plans and the possibility of freely changing the configurations by the users themselves. Our results suggest that users may recognize that the guided interaction paradigm improves efficiency. However, they also feel less freedom to experiment with "what-if" scenarios. Nevertheless, the time spent on the purely exploratory interface tends to be perceived as a lack of efficiency, which reduces attractiveness, perspicuity, and dependability. Conversely, for the guided interface, more time on the interface seems to increase its attractiveness, perspicuity, and dependability while not impacting the perceived efficiency. That might suggest that this type of interfaces should combine these two approaches by trying to support exploratory behavior while gently pushing toward a guided effective solution.

Large Language Models (LLMs) demonstrate an impressive capacity to recall a vast range of factual knowledge. However, understanding their underlying reasoning and internal mechanisms in exploiting this knowledge remains a key research area. This work unveils the factual information an LLM represents internally for sentence-level claim verification. We propose an end-to-end framework to decode factual knowledge embedded in token representations from a vector space to a set of ground predicates, showing its layer-wise evolution using a dynamic knowledge graph. Our framework employs activation patching, a vector-level technique that alters a token representation during inference, to extract encoded knowledge. Accordingly, we neither rely on training nor external models. Using factual and common-sense claims from two claim verification datasets, we showcase interpretability analyses at local and global levels. The local analysis highlights entity centrality in LLM reasoning, from claim-related information and multi-hop reasoning to representation errors causing erroneous evaluation. On the other hand, the global reveals trends in the underlying evolution, such as word-based knowledge evolving into claim-related facts. By interpreting semantics from LLM latent representations and enabling graph-related analyses, this work enhances the understanding of the factual knowledge resolution process., Comment: Accepted at COLM 2024

We explore the potential of Large Language Models (LLMs) to assist and potentially correct physicians in medical decision-making tasks. We evaluate several LLMs, including Meditron, Llama2, and Mistral, to analyze the ability of these models to interact effectively with physicians across different scenarios. We consider questions from PubMedQA and several tasks, ranging from binary (yes/no) responses to long answer generation, where the answer of the model is produced after an interaction with a physician. Our findings suggest that prompt design significantly influences the downstream accuracy of LLMs and that LLMs can provide valuable feedback to physicians, challenging incorrect diagnoses and contributing to more accurate decision-making. For example, when the physician is accurate 38% of the time, Mistral can produce the correct answer, improving accuracy up to 74% depending on the prompt being used, while Llama2 and Meditron models exhibit greater sensitivity to prompt choice. Our analysis also uncovers the challenges of ensuring that LLM-generated suggestions are pertinent and useful, emphasizing the need for further research in this area., Comment: Accepted for oral presentation at NAACL 2024, The 6th Clinical Natural Language Processing Workshop

There is increasing interest in developing AIs for assisting human decision-making in high-stakes tasks, such as medical diagnosis, for the purpose of improving decision quality and reducing cognitive strain. Mainstream approaches team up an expert with a machine learning model to which safer decisions are offloaded, thus letting the former focus on cases that demand their attention. his separation of responsibilities setup, however, is inadequate for high-stakes scenarios. On the one hand, the expert may end up over-relying on the machine's decisions due to anchoring bias, thus losing the human oversight that is increasingly being required by regulatory agencies to ensure trustworthy AI. On the other hand, the expert is left entirely unassisted on the (typically hardest) decisions on which the model abstained. As a remedy, we introduce learning to guide (LTG), an alternative framework in which - rather than taking control from the human expert - the machine provides guidance useful for decision making, and the human is entirely responsible for coming up with a decision. In order to ensure guidance is interpretable} and task-specific, we develop SLOG, an approach for turning any vision-language model into a capable generator of textual guidance by leveraging a modicum of human feedback. Our empirical evaluation highlights the promise of \method on a challenging, real-world medical diagnosis task.

Neuro-Symbolic (NeSy) predictors that conform to symbolic knowledge - encoding, e.g., safety constraints - can be affected by Reasoning Shortcuts (RSs): They learn concepts consistent with the symbolic knowledge by exploiting unintended semantics. RSs compromise reliability and generalization and, as we show in this paper, they are linked to NeSy models being overconfident about the predicted concepts. Unfortunately, the only trustworthy mitigation strategy requires collecting costly dense supervision over the concepts. Rather than attempting to avoid RSs altogether, we propose to ensure NeSy models are aware of the semantic ambiguity of the concepts they learn, thus enabling their users to identify and distrust low-quality concepts. Starting from three simple desiderata, we derive bears (BE Aware of Reasoning Shortcuts), an ensembling technique that calibrates the model's concept-level confidence without compromising prediction accuracy, thus encouraging NeSy architectures to be uncertain about concepts affected by RSs. We show empirically that bears improves RS-awareness of several state-of-the-art NeSy models, and also facilitates acquiring informative dense annotations for mitigation purposes.

In machine learning (ML) verification, the majority of procedures are non-quantitative and therefore cannot be used for verifying probabilistic models, or be applied in domains where hard guarantees are practically unachievable. The probabilistic formal verification (PFV) of ML models is in its infancy, with the existing approaches limited to specific ML models, properties, or both. This contrasts with standard formal methods techniques, whose successful adoption in real-world scenarios is also due to their support for a wide range of properties and diverse systems. We propose a unifying framework for the PFV of ML systems based on Weighted Model Integration (WMI), a relatively recent formalism for probabilistic inference with algebraic and logical constraints. Crucially, reducing the PFV of ML models to WMI enables the verification of many properties of interest over a wide range of systems, addressing multiple limitations of deterministic verification and ad-hoc algorithms. We substantiate the generality of the approach on prototypical tasks involving the verification of group fairness, monotonicity, robustness to noise, probabilistic local robustness and equivalence among predictors. We characterize the challenges related to the scalability of the approach and, through our WMI-based perspective, we show how successful scaling techniques in the ML verification literature can be generalized beyond their original scope.

The Q&U Bolometric Interferometer for Cosmology (QUBIC) is the first bolometric interferometer designed to measure the primordial B-mode polarization of the Cosmic Microwave Background (CMB). Bolometric interferometry is a novel technique that combines the sensitivity of bolometric detectors with the control of systematic effects that is typical of interferometry, both key features in the quest for the faint signal of the primordial B-modes. A unique feature is the so-called "spectral imaging", i.e., the ability to recover the sky signal in several sub-bands within the physical band during data analysis. This feature provides an in-band spectral resolution of \Delta{\nu}/{\nu} \sim 0.04 that is unattainable by a traditional imager. This is a key tool for controlling the Galactic foregrounds contamination. In this paper, we describe the principles of bolometric interferometry, the current status of the QUBIC experiment and future prospects., Comment: To appear in Proc. of the mm Universe 2023 conference, Grenoble (France), June 2023, published by F. Mayet et al. (Eds), EPJ Web of conferences, EDP Sciences

Over the last decade, several regulatory bodies have started requiring the disclosure of non-financial information from publicly listed companies, in light of the investors' increasing attention to Environmental, Social, and Governance (ESG) issues. Publicly released information on sustainability practices is often disclosed in diverse, unstructured, and multi-modal documentation. This poses a challenge in efficiently gathering and aligning the data into a unified framework to derive insights related to Corporate Social Responsibility (CSR). Thus, using Information Extraction (IE) methods becomes an intuitive choice for delivering insightful and actionable data to stakeholders. In this study, we employ Large Language Models (LLMs), In-Context Learning, and the Retrieval-Augmented Generation (RAG) paradigm to extract structured insights related to ESG aspects from companies' sustainability reports. We then leverage graph-based representations to conduct statistical analyses concerning the extracted insights. These analyses revealed that ESG criteria cover a wide range of topics, exceeding 500, often beyond those considered in existing categorizations, and are addressed by companies through a variety of initiatives. Moreover, disclosure similarities emerged among companies from the same region or sector, validating ongoing hypotheses in the ESG literature. Lastly, by incorporating additional company attributes into our analyses, we investigated which factors impact the most on companies' ESG ratings, showing that ESG disclosure affects the obtained ratings more than other financial or company data.

Existing multi-relational graph neural networks use one of two strategies for identifying informative relations: either they reduce this problem to low-level weight learning, or they rely on handcrafted chains of relational dependencies, called meta-paths. However, the former approach faces challenges in the presence of many relations (e.g., knowledge graphs), while the latter requires substantial domain expertise to identify relevant meta-paths. In this work we propose a novel approach to learn meta-paths and meta-path GNNs that are highly accurate based on a small number of informative meta-paths. Key element of our approach is a scoring function for measuring the potential informativeness of a relation in the incremental construction of the meta-path. Our experimental evaluation shows that the approach manages to correctly identify relevant meta-paths even with a large number of relations, and substantially outperforms existing multi-relational GNNs on synthetic and real-world experiments.

Focus in Explainable AI is shifting from explanations defined in terms of low-level elements, such as input features, to explanations encoded in terms of interpretable concepts learned from data. How to reliably acquire such concepts is, however, still fundamentally unclear. An agreed-upon notion of concept interpretability is missing, with the result that concepts used by both post-hoc explainers and concept-based neural networks are acquired through a variety of mutually incompatible strategies. Critically, most of these neglect the human side of the problem: a representation is understandable only insofar as it can be understood by the human at the receiving end. The key challenge in Human-interpretable Representation Learning (HRL) is how to model and operationalize this human element. In this work, we propose a mathematical framework for acquiring interpretable representations suitable for both post-hoc explainers and concept-based neural networks. Our formalization of HRL builds on recent advances in causal representation learning and explicitly models a human stakeholder as an external observer. This allows us to derive a principled notion of alignment between the machine representation and the vocabulary of concepts understood by the human. In doing so, we link alignment and interpretability through a simple and intuitive name transfer game, and clarify the relationship between alignment and a well-known property of representations, namely disentanglment. We also show that alignment is linked to the issue of undesirable correlations among concepts, also known as concept leakage, and to content-style separation, all through a general information-theoretic reformulation of these properties. Our conceptualization aims to bridge the gap between the human and algorithmic sides of interpretability and establish a stepping stone for new research on human-interpretable representations.

In learning to defer, a predictor identifies risky decisions and defers them to a human expert. One key issue with this setup is that the expert may end up over-relying on the machine's decisions, due to anchoring bias. At the same time, whenever the machine chooses the deferral option the expert has to take decisions entirely unassisted. As a remedy, we propose learning to guide (LTG), an alternative framework in which -- rather than suggesting ready-made decisions -- the machine provides guidance useful to guide decision-making, and the human is entirely responsible for coming up with a decision. We also introduce SLOG, an LTG implementation that leverages (a small amount of) human supervision to convert a generic large language model into a module capable of generating textual guidance, and present preliminary but promising results on a medical diagnosis task.

One population group that had to significantly adapt and change their behaviour during the COVID-19 pandemic is students. While previous studies have extensively investigated the impact of the pandemic on their psychological well-being and academic performance, limited attention has been given to their activity routines. In this work, we analyze students' behavioural changes by examining qualitative and quantitative differences in their daily routines between two distinct periods (2018 and 2020). Using an Experience Sampling Method (ESM) that captures multimodal self-reported data on students' activity, locations and sociality, we apply Non-Negative Matrix Factorization (NMF) to extract meaningful behavioural components, and quantified the variations in behaviour between students in 2018 and 2020. Surprisingly, despite the presence of COVID-19 restrictions, we find minimal changes in the activities performed by students, and the diversity of activities also remains largely unaffected. Leveraging the richness of the data at our disposal, we discovered that activities adaptation to the pandemic primarily occurred in the location and sociality dimensions.

Neuro-Symbolic (NeSy) predictive models hold the promise of improved compliance with given constraints, systematic generalization, and interpretability, as they allow to infer labels that are consistent with some prior knowledge by reasoning over high-level concepts extracted from sub-symbolic inputs. It was recently shown that NeSy predictors are affected by reasoning shortcuts: they can attain high accuracy but by leveraging concepts with unintended semantics, thus coming short of their promised advantages. Yet, a systematic characterization of reasoning shortcuts and of potential mitigation strategies is missing. This work fills this gap by characterizing them as unintended optima of the learning objective and identifying four key conditions behind their occurrence. Based on this, we derive several natural mitigation strategies, and analyze their efficacy both theoretically and empirically. Our analysis shows reasoning shortcuts are difficult to deal with, casting doubts on the trustworthiness and interpretability of existing NeSy solutions., Comment: 37th Conference on Neural Information Processing Systems (NeurIPS 2023)

We are interested in aligning how people think about objects and what machines perceive, meaning by this the fact that object recognition, as performed by a machine, should follow a process which resembles that followed by humans when thinking of an object associated with a certain concept. The ultimate goal is to build systems which can meaningfully interact with their users, describing what they perceive in the users' own terms. As from the field of Lexical Semantics, humans organize the meaning of words in hierarchies where the meaning of, e.g., a noun, is defined in terms of the meaning of a more general noun, its genus, and of one or more differentiating properties, its differentia. The main tenet of this paper is that object recognition should implement a hierarchical process which follows the hierarchical semantic structure used to define the meaning of words. We achieve this goal by implementing an algorithm which, for any object, recursively recognizes its visual genus and its visual differentia. In other words, the recognition of an object is decomposed in a sequence of steps where the locally relevant visual features are recognized. This paper presents the algorithm and a first evaluation., Comment: 10 pages, 5 figures, Accepted for publication at The second International Conference on Hybrid Human-Artificial Intelligence (HHAI2023)

In this paper, we introduce Interval Real Logic (IRL), a two-sorted logic that interprets knowledge such as sequential properties (traces) and event properties using sequences of real-featured data. We interpret connectives using fuzzy logic, event durations using trapezoidal fuzzy intervals, and fuzzy temporal relations using relationships between the intervals' areas. We propose Interval Logic Tensor Networks (ILTN), a neuro-symbolic system that learns by propagating gradients through IRL. In order to support effective learning, ILTN defines smoothened versions of the fuzzy intervals and temporal relations of IRL using softplus activations. We show that ILTN can successfully leverage knowledge expressed in IRL in synthetic tasks that require reasoning about events to predict their fuzzy durations. Our results show that the system is capable of making events compliant with background temporal knowledge.

Neuro-symbolic predictors learn a mapping from sub-symbolic inputs to higher-level concepts and then carry out (probabilistic) logical inference on this intermediate representation. This setup offers clear advantages in terms of consistency to symbolic prior knowledge, and is often believed to provide interpretability benefits in that - by virtue of complying with the knowledge - the learned concepts can be better understood by human stakeholders. However, it was recently shown that this setup is affected by reasoning shortcuts whereby predictions attain high accuracy by leveraging concepts with unintended semantics, yielding poor out-of-distribution performance and compromising interpretability. In this short paper, we establish a formal link between reasoning shortcuts and the optima of the loss function, and identify situations in which reasoning shortcuts can arise. Based on this, we discuss limitations of natural mitigation strategies such as reconstruction and concept supervision.

The development of efficient exact and approximate algorithms for probabilistic inference is a long-standing goal of artificial intelligence research. Whereas substantial progress has been made in dealing with purely discrete or purely continuous domains, adapting the developed solutions to tackle hybrid domains, characterised by discrete and continuous variables and their relationships, is highly non-trivial. Weighted Model Integration (WMI) recently emerged as a unifying formalism for probabilistic inference in hybrid domains. Despite a considerable amount of recent work, allowing WMI algorithms to scale with the complexity of the hybrid problem is still a challenge. In this paper we highlight some substantial limitations of existing state-of-the-art solutions, and develop an algorithm that combines SMT-based enumeration, an efficient technique in formal verification, with an effective encoding of the problem structure. This allows our algorithm to avoid generating redundant models, resulting in drastic computational savings. Additionally, we show how SMT-based approaches can seamlessly deal with different integration techniques, both exact and approximate, significantly expanding the set of problems that can be tackled by WMI technology. An extensive experimental evaluation on both synthetic and real-world datasets confirms the substantial advantage of the proposed solution over existing alternatives. The application potential of this technology is further showcased on a prototypical task aimed at verifying the fairness of probabilistic programs.

We introduce Neuro-Symbolic Continual Learning, where a model has to solve a sequence of neuro-symbolic tasks, that is, it has to map sub-symbolic inputs to high-level concepts and compute predictions by reasoning consistently with prior knowledge. Our key observation is that neuro-symbolic tasks, although different, often share concepts whose semantics remains stable over time. Traditional approaches fall short: existing continual strategies ignore knowledge altogether, while stock neuro-symbolic architectures suffer from catastrophic forgetting. We show that leveraging prior knowledge by combining neuro-symbolic architectures with continual strategies does help avoid catastrophic forgetting, but also that doing so can yield models affected by reasoning shortcuts. These undermine the semantics of the acquired concepts, even when detailed prior knowledge is provided upfront and inference is exact, and in turn continual performance. To overcome these issues, we introduce COOL, a COncept-level cOntinual Learning strategy tailored for neuro-symbolic continual problems that acquires high-quality concepts and remembers them over time. Our experiments on three novel benchmarks highlights how COOL attains sustained high performance on neuro-symbolic continual learning tasks in which other strategies fail., Comment: 40th International Conference on Machine Learning (ICML 2023)

Graph Neural Networks (GNNs) have become the leading paradigm for learning on (static) graph-structured data. However, many real-world systems are dynamic in nature, since the graph and node/edge attributes change over time. In recent years, GNN-based models for temporal graphs have emerged as a promising area of research to extend the capabilities of GNNs. In this work, we provide the first comprehensive overview of the current state-of-the-art of temporal GNN, introducing a rigorous formalization of learning settings and tasks and a novel taxonomy categorizing existing approaches in terms of how the temporal aspect is represented and processed. We conclude the survey with a discussion of the most relevant open challenges for the field, from both research and application perspectives.