Beyond 5G (B5G) networks are designed to significantly enhance network capacity and reduce latency by utilizing higher frequency spectrum bands. These networks rely on effective network slicing—the partitioning of physical infrastructure into independent logical networks, each optimized for specific service requirements to support a wide range of emerging application. However, B5G networks face substantial challenges due to massive data generation and stringent Service-Level Agreements (SLAs). This study introduces a Network Slice Framework (NSFrame) that addresses these challenges through a novel approach combining feature selection techniques with deep learning for efficient slice classification. NSFrame integrates Mutual Information, Shapley values from cooperative game theory, and Borda Count rank aggregation to overcome high dimensionality by selecting the most relevant features. These features are then processed by a multi-head attention enhanced long short-term memory model. Evaluated on the Unicauca IP Flow Version2 and 5G-SliciNdd datasets, NSFrame achieved classification accuracies of 99.79% and 98.67%, respectively, with strong generalization confirmed through 10-fold cross-validation. The proposed approach significantly outperforms existing methods, enhancing both quality of service and quality of experience, while enabling service providers to meet strict SLAs in B5G environments where softwarization and virtualization are essential for customized service delivery.

中文翻译：

---

一种注意力增强的 LSTM 模型，用于 5G 网络之外的高效网络切片

超越 5G（B5G）网络旨在通过使用更高频谱频段，显著提升网络容量并降低延迟。这些网络依赖于有效的网络切片——将物理基础设施划分为独立的逻辑网络，每个逻辑网络针对特定服务需求进行优化，以支持广泛的新兴应用。然而，B5G 网络面临着大量数据生成和严格的服务水平协议（SLA）带来的重大挑战。本研究引入了网络切片框架（NSFrame），通过结合特征选择技术与深度学习的新颖方法，高效地实现切片分类，解决了这些挑战。NSFrame 整合了互信息、合作博弈论中的 Shapley 值和 Borda 计数秩聚合，通过选择最相关的特征克服高维度问题。这些特征随后被多头注意力增强的长期短期记忆模型处理。在 Unicauca IP Flow Version2 和 5G-SliciNdd 数据集上评估时，NSFrame 分别实现了 99.79%和 98.67%的分类准确率，并通过 10 折交叉验证验证了强强的泛化性。该方法显著优于现有方法，提升了服务质量和体验质量，同时使服务提供商能够满足 B5G 环境中严格的 SLA，软战和虚拟化对定制化服务交付至关重要。