Special sessions

The increasing availability of large-scale biomedical and healthcare data, including electronic health records (EHRs), medical imaging, physiological signals, genomics, epigenomics and other omics data, offers unprecedented opportunities for scientific discovery and precision medicine. This special session focuses on machine learning, neural networks, computational intelligence and data mining methods that support knowledge discovery from complex biomedical data through effective collaboration between AI systems and domain experts. Topics of interest include multimodal learning, clinical data analytics, medical image analysis, AI-assisted decision support, disease diagnosis and prognosis, representation learning, explainable and trustworthy AI, visual analytics, interactive machine learning, and human-in-the-loop approaches. Particular emphasis will be placed on methodologies that not only improve predictive performance but also facilitate scientific insight generation, hypothesis discovery and actionable knowledge extraction in biomedical and healthcare applications.

Topics of Interest:

Biomedical and Clinical AI

• Machine Learning for Healthcare
• Clinical Data Mining
• Electronic Health Records (EHR) Analytics
• Clinical Decision Support Systems
• AI-Assisted Triage and Risk Stratification
• Predictive Modeling in Healthcare

Medical Imaging

• Medical Image Analysis
• Computer-Aided Diagnosis
• Radiomics and Pathomics
• Deep Learning for Imaging Biomarkers
• Multimodal Imaging Analysis

AI Methods for Omics and Precision Medicine

• Genomics and Epigenomics
• Multi-omics Integration
• Precision and Personalized Medicine
• Biomarker Discovery
• Systems Biology and Computational Medicine

Knowledge Discovery

• Knowledge Discovery from Biomedical Data
• Hypothesis Generation using AI
• Data Mining and Exploratory Analytics
• Scientific Discovery Systems
• Representation Learning for Biomedical Data

Human-AI Collaboration

• Interactive Machine Learning
• Human-in-the-Loop AI
• Explainable and Trustworthy AI
• Visual Analytics
• Expert-Guided Learning
• Trustworthy Human-AI Interaction

Emerging Topics

• Foundation Models for Healthcare
• Graph Neural Networks in Biomedicine
• Federated Learning
• Privacy-Preserving AI
• Causal Learning in Healthcare
   

Graph Neural Networks have become a leading framework for learning from relational and graph-structured data. However, the growing scale and complexity of real-world graphs are exposing the limitations of conventional message-passing architectures and standard training methods. This special session will focus on emerging graph learning models that exploit topological, geometrical, spectral, physical, and dynamical principles. It will also cover novel training paradigms aimed at improving scalability, data efficiency, memory usage, and energy consumption. Topics include but are not limited to

• Graph representation learning;
• Advanced Graph Neural Architectures;
• Backpropagation-free graph learning;
• Graph structure learning and relational inference;
• Theory of graph neural networks (e.g., expressive power, learnability, negative results);
• Explainability in Graph Learning;
• Learning on complex graphs (e.g., dynamic graphs and heterogeneous graphs);
• Randomized neural networks for graphs (e.g., reservoir computing);
• Recurrent, recursive, and contextual models;
• Scalability, data efficiency, and training techniques of graph neural networks;
• Architectures for foundation models operating on graphs;
• Graph datasets and benchmarks.

The session aims to connect architectural innovation, theoretical understanding, and resource-efficient learning to identify promising directions for the next generation of Graph Neural Networks.
 

Since its introduction, Deep learning (DL) has achieved significant success. In many application areas, DL models constitute effective solutions, but their development still requires vast amounts of annotated data. In many contexts, due either to the annotation process or to the expertise required to generate high-quality labels, annotated data are scarce, and sometimes even absent. The design of DL methods that are less dependent on annotations is therefore one of the hottest topics in DL.  

The goal of this special session on learning with few annotations is to gather (i) researchers who aim to alleviate the dependence of DL models in annotations, as well as (ii) researchers working in application domains where annotations are scarce. Special attention should be given to challenging datasets, such as domains with high imbalance between classes, very large number of classes, temporal data, etc. 

Topics of interest include (but are not limited to): 

• self-supervised and semi-supervised learning
• pseudo-labeling methods
• prototypical learning
• few-shot learning
• data augmentation
• benchmarking 
   

For much of their history, kernel methods and neural networks were treated as competing paradigms, the former prized for their solid theoretical footing and convex training, the latter for their empirical power. A long line of work has dissolved this divide, resulting into new synergies. Infinitely wide networks correspond to kernels and Gaussian processes, and composing these limits layer by layer yields deep, infinite-dimensional kernels whose covariance is built recursively from the layer below, with explicit forms for common activations.

The neural tangent kernel extends the same idea to the training dynamics of wide networks. Together these results give a principled, infinite-dimensional kernel description of what deep networks compute.

More recent work aims to make these infinite-dimensional kernels learnable and expressive rather than fixed, for instance by treating the kernel itself as an object to be inferred from data. This connects naturally to deep and restricted kernel machines, to graph kernels and graph neural networks, and to scalable approximations that make such methods practical.

This special session brings these threads together.

We welcome novel theoretical, methodological, and applied contributions linking kernel learning, Gaussian processes, and neural networks, with a particular focus on the infinite-dimensional and deep kernel perspective.

We encourage the submission of papers on topics including but not limited to:

• Infinite-dimensional and infinite-width kernels: arc-cosine, compositional, and deep kernels in the line of Cho and Saul
• Correspondences between infinitely wide networks, Gaussian processes, and neural tangent kernels
• Deep kernel processes and learning the kernel: representation learning in kernel space
• Heavy-tailed and stable generalizations of the infinite-width limit, beyond bounded-variance priors
• Deep kernel machines, restricted kernel machines, and primal-dual links between kernels and deep architectures
• Graph kernels, graph neural networks, and the expressivity of message passing
• Scalable and approximate methods for infinite-dimensional and deep kernels: random features, Nystrom, low-rank and sketching
• Indefinite, non-metric, and structured kernels
• Gaussian processes for deep models, uncertainty quantification, and calibration
• Applications demonstrating these methods, for example in molecular and graph data and the sciences

Machine learning systems are increasingly deployed in dynamic environments, where data distributions evolve over time, across locations, or between data sources. Existing work often treats drift as a technical challenge for maintaining predictive performance. However, understanding and describing drift are particularly important for generating suitable explanations for humans, improving the trustworthiness and fairness of AI systems, and realizing informed model adaptation. The goal of this session is to exploit the knowledge of change by bringing together approaches that characterize meaningful changes, detect, localize, and explain drift, promote human understanding, inform downstream tasks, and enable targeted model adaptation and decision-making.

More concretely, we are looking for submissions focusing on the following issues in streaming data, federated learning, and when collecting data across multiple sites:

• Definitions of meaningful change
• Structured drift models and mechanisms
• Expected vs unexpected change
• Monitoring meaningful system changes
• Uncertainty-aware drift detection
• Evaluation protocols for drift detectors
• Drift characterization, attribution and localization
• Realistic drift benchmarks
• Explainable AI for stream learning, concept drift, and changing decision behavior, including correlation- and causality-based approaches
• Human-centered evaluation of explanations in dynamic settings
• Trust calibration, fairness, reliance, actionability, and user understanding under non-stationarity
• Benchmarks, metrics, and reporting standards for explainable stream learning
• Targeted model adaptation based on drift
• Local model updates and specialization
• Memory and data management strategies 

The session explores the interfaces between Machine Learning and Network Science including aspects like network analyses and processing, from both, the methods development and applicational, viewpoints. Based on early graph and tree representation methods, abstracting networks from data, analyzing networks, machine learning techniques involving some form of graph representations have been an emerging field in machine learning. On the applicational side, medical network-centered analyses, maps of science in scientometrics, analysis of economic networks, and computational social science as well as the forensic analysis of criminal networks or graph representations of molecules have grown to established fields, while recently receiving a methodical boost through deep learning, large language models, and generative AI.

Welcoming both theory and applications of respective machine learning approaches, we especially encourage contributions combining both, method development domain-specifically motivated from specific data structure and validating new methodical approaches on real-world data and scenarios. Theoretically motivated new methodologies are also highly welcome.