Scholarlink

This paper introduces DeepSliceNet, an end-to-end deep learning framework for classifying network slice types to improve resource allocation in 6G networks. By applying a comprehensive pipeline including data augmentation and robust overfitting controls on the DeepSlice dataset, our Bidirectional LSTM (BiLSTM) model achieved a standout 96% accuracy and a macro-F1 score of 0.94. This sequential model demonstrated superior performance over five other architectures (Dense, CNN, LSTM, GRU), showing exceptional generalization for minority classes like mMTC and URLLC. To combat the limited size of telecom datasets, we propose a domain-specific synthetic data augmentation technique and overfitting countermeasures like early stopping and dropout. The high performance of the BiLSTM model underscores its suitability for capturing bidirectional feature dependencies in structured slice-type problems, enabling intelligent resource allocation and network resilience.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DeepSliceNet: A Deep Learning Framework for Network Slice Type Prediction with Augmentation and Overfitting Control

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Naween Kumar,
  • Yajnaseni Dash,
  • Ankit Dubey,
  • Shubham Sharma,
  • Rahat Iqbal,
  • Saad Amin,
  • Faiyaz Doctor

摘要

This paper introduces DeepSliceNet, an end-to-end deep learning framework for classifying network slice types to improve resource allocation in 6G networks. By applying a comprehensive pipeline including data augmentation and robust overfitting controls on the DeepSlice dataset, our Bidirectional LSTM (BiLSTM) model achieved a standout 96% accuracy and a macro-F1 score of 0.94. This sequential model demonstrated superior performance over five other architectures (Dense, CNN, LSTM, GRU), showing exceptional generalization for minority classes like mMTC and URLLC. To combat the limited size of telecom datasets, we propose a domain-specific synthetic data augmentation technique and overfitting countermeasures like early stopping and dropout. The high performance of the BiLSTM model underscores its suitability for capturing bidirectional feature dependencies in structured slice-type problems, enabling intelligent resource allocation and network resilience.