Expert Systems最新文献

The application of artificial intelligence (AI) in healthcare has been witnessing an increasing interest. Particularly, federated learning (FL) has become favourable due to its potential for enhancing model quality whilst maintaining data privacy and security. However, the effectiveness of present FL methodologies could underperform under non-IID conditions, characterised by divergent data distributions across clients. The globally constructed FL model may suffer potent issues by allowing the least-performing models to equal participation. Thus, we propose a new accuracy-based FL approach (FedAcc) which only takes into account the clients' validation accuracy to consider their participation during global aggregation, also called Smart Healthcare Amplified (SHA). However, with limited supervised data it is challenging to increase the model performance thus concept of transfer learning (TL) is used. TL enables the global model to integrate knowledge from precomputed systems, resulting in an efficient model. However, the complexity of the global system is amplified by these TL models, leading to challenges related to vanishing gradients, particularly when dealing with a substantial number of layers. To mitigate this, we present a Transfer Learning Domain Adaptation Model (TLDAM). TLDAM employs a two-layered sequentially trained TL model, which contains approximately 50% fewer layers compared to traditional TL models. TLDAM is trained on multiple datasets such as MNIST and CIFAR10, to enhance its knowledge and make it domain-adaptive. Moreover, experimental results conducted on the UCI-HAR dataset reveal the supremacy of our proposed framework with an accuracy of 94.2990%, F-score of 94.2820%, precision of 94.3058%, and recall of 94.2993% over traditional FL techniques and state-of-the-art techniques.

Recent 3D point cloud classification has predominantly focused on local spatial attention, neglecting distant contextual relationships due to the inherent sparsity of LiDAR-generated data over longer distances. Existing 3D object detection methods prioritize local features, hindering the extraction of semantic information. Despite attempts with transformers, methods often reduce computations through local spatial attention, neglecting content class and scarcely establishing connections among distant global points. Our proposed point class-adaptive transformer (PCaT) addresses these limitations by establishing long-range feature dependencies while significantly reducing computations. PCaT includes three key modules: the class-adaptive transformer (CaT), which utilizes local self-attention and global self-attention based on class similarity to facilitate an efficient trade-off between capturing extended-global dependencies and managing computational challenges; nested binary clustering (NbC), which dynamically partitions queries into multiple clusters based on content features in each Transformer block; and the AfA, which aggregates high-dimensional features using max-pooling alongside a residual MLP component and low-dimensional features using average pooling and a CaT block. Additionally, PCaT incorporates point cloud segmentation via local–global feature aggregation (PcSeg) to facilitate effective point cloud segmentation. Extensive experimentation on the ModelNet40, ScanObjectNN, and S3DIS datasets demonstrates the superior performance and reasonable stability of PCaT compared with existing methods. PCaT achieves 94.2% overall accuracy (OA) and mIoU scores of 89.2% and 86.2% for the ScanObjectNN and S3DIS datasets, respectively.

The application of machine learning, particularly federated learning, in collaborative model training, has demonstrated significant potential for enhancing diversity and efficiency in outcomes. In the healthcare domain, particularly healthcare with disabilities, the sensitive nature of data presents a significant challenge as sharing even the computation on these data can risk exposing personal health information. This research addresses the problem of enabling shared model training for healthcare data—particularly with disabilities decreasing the risk of leaking or compromising sensitive information. Technologies such as federated learning provide solution for decentralised model training but fall short in addressing concerns related to trust building, accountability and control over participation and data. We propose a framework that integrates federated learning with advanced identity management as well as privacy and trust management technologies. Our framework called Theme (Trusted Healthcare Machine Learning Environment) leverages digital identities (e.g., W3C decentralised identifiers and verified credentials) and policy enforcements to regulate participation. This is to ensure that only authorised and trusted entities can contribute to the model training. Additionally, we introduce the mechanisms to track contributions per participant and offer the flexibility for participants to opt out of model training at any point. Participants can choose to be either contributors (providers) or consumers (model users) or both, and they can also choose to participate in subset of activities. This is particularly important in healthcare settings, where individuals and healthcare institutions have the flexibility to control how their data are used without compromising the benefits. In summary, this research work contributes to privacy preserving shared model training leveraging federated learning without exposing sensitive data; trust and accountability mechanisms; contribution tracking per participant for accountability and back-tracking; and fine-grained control and autonomy per participant. By addressing the specific needs of healthcare data for people with disabilities or such institutions, the Theme framework offers a robust solution to balance the benefits of shared machine learning with critical need to protecting sensitive data.

Worldwide terrorist activities continue to pose a significant threat to global security and stability. The unpredictable nature of these acts necessitates advanced analytical approaches to enhance prevention and response strategies. This study examines undetectable word extensions across multiple datasets, using terrorism-related datasets as a case study. This research aims to overcome constraints in current predictive models associated with terrorist attack prediction. While many studies have used the GTD for predicting global terrorist attacks, this study expands beyond GTD by evaluating a corpus of terrorism incidents to enhance predictive analysis through lexical usage. The study employs several machine learning algorithms including Decision Tree (DT), Bootstrap Aggregating (BA), Random Forest (RF), Extra Trees (ET) and XGBoost (XG) algorithms for evaluation. Our approach integrates multiple datasets to reduce dependence on GTD alone. Findings indicate that RF performs best on the GTD database, with 90.20% accuracy in predicting worldwide terrorist attacks. DT achieves 90.40% accuracy when applied to the TF–IDF dataset. XG demonstrates superior performance across various aggregation settings and feature sets, achieving 95.77% accuracy in forecasting worldwide terrorist acts. XG's consistent and effective performance across various contexts highlights its versatility. Its high adaptability and robust performance position it as the preferred algorithm for conducting predictive research on global terrorist acts using the available datasets. Our research findings underscore the importance of incorporating diverse datasets to enhance understanding of terrorist activities and improve predictive capabilities.