Local community detection is a specialized area within the broader field of community detection, focusing on the identification of communities centered around a set of initial seed nodes. Despite increasing research interest over the past two decades, a comprehensive assessment of the diverse existing methods and their performance under varying conditions remains challenging. In this paper, we conduct a focused comparative study of several well-known local community detection algorithms, applied specifically to static undirected graphs. We implement representative algorithms and evaluate their behavior, strengths, and limitations across standard benchmark datasets. Our aim is to provide clear empirical insights into the performance and trade-offs of these algorithms, thereby guiding researchers in selecting appropriate methods for their use cases. We also present implementation details and evaluation results to promote reproducibility and facilitate further investigation.
Federated learning is a distributed paradigm that enables collaborative training across multiple clients while preserving data privacy. However, in practice, it often encounters challenges such as heterogeneous client computing capabilities and varying communication conditions. Additionally, differences in user preferences, data acquisition environments, and devices lead to non-identical data distributions and diverse research tasks across clients. To address these heterogeneity challenges, this paper proposes a constrained Federated Learning Partial Training algorithm (cFedPT), motivated by variations in client data distributions. The algorithm dynamically adjusts the participating submodels in each training round and enables clients with heterogeneous resources to engage in federated learning. Experimental results on the ISIC2019 image dataset and the ICBHI audio dataset demonstrate that cFedPT achieves superior performance compared to baseline models.
One crucial issue in Federated Learning (FL) is the fact that the participating devices, i.e., the data owners, are expected to collect non-Independent-and-Identically-Distributed (non-IID) data. This deteriorates training performance, i.e., delaying training or not reaching convergence. Consequently, additional methods, such as data augmentation, are applied to address this problem. In fact, producing synthetic data using GAN models has been proven to be an effective method against non-IID FL. However, the data owners need to participate in an additional FL training to train GANs in a privacy-preserving manner. Yet, this is not expected to be always feasible due to data owners’ resource limitations. In this position paper, we identify the issues that occur when training GANs with FL for data augmentation, and propose a lightweight alternative that utilizes Split Learning (SL) to offload the computational load into a compute node. Further, we highlight the gap of cloud-based FL-SL integration and propose a microservice architecture based on existing tools that could significantly enhance the FL-SL deployment and orchestration, while balancing the energy and the financial cost.
We study the renting servers in the cloud problem (RSiC), inspired by job allocation in cloud computing environments. In this problem, jobs arrive sequentially in an online manner, and their sizes are revealed upon arrival. If a job’s duration is also known at arrival, the scenario is classified as clairvoyant; otherwise, it is non-clairvoyant. Each job must be assigned to a server, with servers available on-demand and subject to a fixed capacity per unit of time. The number of servers that can be rented is unlimited, and the objective is to minimize the total rental time of all servers. In this paper, we evaluate the performance of nearly all existing clairvoyant and non-clairvoyant algorithms for the RSiC problem. Additionally, we introduce new algorithms, which were derived from combinations of existing algorithms. Some of the introduced algorithms outperform all previously known algorithms in our experimental evaluations. Unlike prior studies that exclusively utilized synthetic data, we use real-world Azure data from [4] in our experiments. This dataset captures large-scale virtual machine (VM) allocation across Azure’s availability zones, offering a more practical perspective on the operational effectiveness of these algorithms.
Cloud simulation tools are essential for modeling and evaluating complex cloud computing infrastructures, enabling researchers to study resource allocation strategies, scheduling algorithms, and overall system performance. CloudSim Plus is a widely used Java framework for building such simulations. However, despite its rich feature set, it offers only limited support for post-simulation analysis, providing primarily static ASCII tables that summarize final resource states. As simulation complexity grows, this output format becomes insufficient, highlighting the need for more interactive and insightful analysis tools. In this work, we present Cooper, a lightweight and modular visualization framework that extends CloudSim Plus with standardized event collection and timeline-based visualization. Cooper enables detailed temporal analysis of simulation runs, helping users understand dynamic system behavior, resource utilization patterns, and the effects of scheduling decisions. The event logging library runs entirely in memory, introduces minimal runtime overhead, and does not require external database or server infrastructure. Its design reduces integration effort and promotes reproducibility by decoupling simulation from visualization.
