Federated learning (FL) has emerged as a powerful framework for privacy-preserving machine learning, especially relevant in fields like healthcare, finance, and mobile devices. Despite its success, traditional FL systems have a significant limitation: they rely on a static set of clients, forming a federation at the beginning of the training process, which remains fixed throughout the training cycle, thus limiting their scalability and adaptability in dynamic, data-rich settings. To address this, we introduce the concept of federated learning platforms (FLPs), which extend FL into a dynamic platform where client participation is continuously adapted based on their expected value and strategic incentives. In this paper, we envision FLPs as a natural extension of conventional FL that resemble dynamic, value-driven digital platforms where participants can join or leave the federation at any time. Given this dynamicity of client participation, FLPs are designed to gracefully handle changes in the client pool to uphold their value proposition. In this article, we propose a framework for implementing FLPs, outlining key components such as those for dynamic FLP governance, including client on- and offboarding as well as process monitoring. Furthermore, we demonstrate the practical viability of FLPs through a proof of concept for an exemplary use-case and discuss key challenges related to federation stability, data interoperability, as well as privacy, alongside potential solutions. Finally, we present a roadmap and future research directions, guiding the development of robust and scalable FLPs to drive innovation in FL and data interoperability. • We envision FLPs for dynamic and value-driven participation in federated learning. • We propose a framework for FLPs consisting of components for dynamic FLP governance. • We discuss key challenges like federation stability and data privacy concerns. • We provide a possible roadmap for future research directions on FLPs.

Recommender systems are powerful tools developed to mitigate information overload in e-commerce platforms. Social recommender systems leverage social relations among users to predict their preferences. Recently, graph neural networks have been utilized for social recommendations, modeling user-user social relations and user–item interactions as graph-structured data. Despite their improvement over traditional systems, most existing social recommender systems exploit only first-order social relations and overlook the importance of social influence diffusion from higher-order neighbors in social networks. Additionally, these techniques often treat all neighboring nodes equally, without highlighting the most influential ones. To address these challenges, we introduce GATE-SR, a novel model that leverages a multi-head graph attention autoencoder to capture indirect social influence from higher-order neighbors while emphasizing the most relevant users. Moreover, we incorporate implicit social connections derived from coherent communities within the network. While GATE-SR performs comparably to baseline models in rich data environments, its strength lies in excelling at cold-start scenarios—where other models often fall short. This focus on cold-start performance aligns with our goal of building a robust recommender system for real-world challenges. Through extensive experiments on three real-world datasets, we demonstrate that GATE-SR outperforms several state-of-the-art baselines in cold-start scenarios. These results highlight the crucial role of accentuating the most influential neighbors, both explicit and implicit, when modeling higher-order social connections for more accurate recommendations. • Varied attention enhances recommendations by assigning importance to neighbors. • Autoencoder’s stacked layers model high-order social relations effectively. • Community detection cuts over-individualized recommendations, optimizing complexity. • Mitigate data sparsity using implicit social connections. • Adept at mitigating cold-start probelm, emphasizing higher-order social influence.

Federated learning (FL) facilitates shared training of machine learning models while maintaining data privacy. Unfortunately, it suffers from data imbalance among participating clients, causing the performance of the shared model to drop. To diminish the negative effects of unfavorable data-specific properties, both algorithm- and data-based approaches seek to make FL more resilient against them. In this regard, data-based approaches prove to be more versatile and require less domain knowledge to be applied efficiently. Hence, they seem particularly suitable for widespread application in various FL environments. Although data-based approaches such as local data sampling have been applied to FL in the past, previous research did not provide a systematic analysis of the potential and limitations of individual data sampling strategies to improve FL. To this end, we (1) identify relevant local data sampling strategies for FL, (2) identify data-specific properties that negatively affect FL performance, and (3) provide a benchmark of local data sampling strategies regarding their effect on model performance, convergence, and training time in synthetic, real-world, and large-scale FL environments. Moreover, we propose and rigorously test a novel method for data sampling in FL that locally optimizes the choice of sampling strategy prior to FL participation. Our results show that FL can benefit from applying local data sampling in terms of performance and convergence rate, especially when data imbalance is high or the number of clients and samples is low. Furthermore, our proposed sampling strategy offers the best trade-off between model performance and training time.

In recent years, we have witnessed both artificial intelligence obtaining remarkable results in clinical decision support systems (CDSSs) and explainable artificial intelligence (XAI) improving the interpretability of these models. In turn, this fosters the adoption by medical personnel and improves trustworthiness of CDSSs. Among others, counterfactual explanations prove to be one such XAI technique particularly suitable for the healthcare domain due to its ease of interpretation, even for less technically proficient staff. However, the generation of high-quality counterfactuals relies on generative models for guidance. Unfortunately, training such models requires a huge amount of data that is beyond the means of ordinary hospitals. In this paper, we therefore propose to use federated learning to allow multiple hospitals to jointly train such generative models while maintaining full data privacy. We demonstrate the superiority of our approach compared to locally generated counterfactuals. Moreover, we prove that generative models for counterfactual generation that are trained using federated learning in a suitable environment perform only marginally worse compared to centrally trained ones while offering the benefit of data privacy preservation. Finally, we integrate our method into a prototypical CDSS for treatment recommendation for sepsis patients, thus providing a proof of concept for real-world application as well as insights and sanity checks from clinical application. • Limited availability of data limits small hospitals in generating high- quality counterfactual explanations. • Integrating federated learning mitigates this limitation and maintains data privacy. • Benefit of using federated learning depends on the degree of data imbalance among hospitals. • Proof-of-concept for clinical application for sepsis treatment recommendation.

Federated Learning (FL) is a learning paradigm that allows clients to profit from the data that is available across multiple clients to train a joint model. As FL allows to train such a joint model without explicitly sharing data, but only sharing model updates, it has attained popularity in healthcare settings where patient data is subject to strict privacy policies and needs to be locally stored at each hospital or healthcare provider. A particular challenge for FL settings is data imbalance across clients, as it has been found to be detrimental to model performance and impact the influence of each client on the learning process. Unfortunately, the healthcare domain is particularly prone to such imbalanced data due to regional differences in disease management, prescription behavior etc. In this paper, we introduce the two novel metrics Benefit and Contribution to quantify to which degree individual clients benefit from participation in FL and how they contribute to its success, respectively. Therefore, we measure Benefit and Contribution with respect to four types of imbalances present in data at each client side. Our results show that both client Benefit and Contribution are influenced by data imbalance in such a way that high imbalance in data quantity, label distribution and feature distribution reduces or nullifies clients’ Benefit while increasing their Contribution. Thus, the most valuable clients within a cohort benefit the least from their participation, exposing a critical thread to the success of clinical FL cohorts by withdrawing participation.