Scholarlink

Effective mobile network traffic forecasting demands datasets that integrate multivariate features while preserving temporal and spatial dependencies and avoiding feature leakage. We propose a spatiotemporal, causality-preserving preprocessing framework that converts raw measurements into temporally consistent and spatially informed inputs. Its four modules: Multiscale Temporal Feature Construction, Spatiotemporal Enhancement, Interaction and Encoding Layer, and Feature Selection, capture nonlinear, time-dependent, and spatial correlations. Experiments with Hist Gradient Boosting Regressor (HGBR) and Long Short-Term Memory (LSTM) models demonstrate improved model stability, generalization, and interpretability, while ablation studies confirm the unique contribution of each module. The framework enables robust and reproducible throughput prediction from real-world cellular data.

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

Causality-Preserving Spatiotemporal Feature Engineering for Cellular Network Throughput Forecasting

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Pan Ruifeng,
  • Chun Wang,
  • Jianjun Cai,
  • Mengsheng Wang,
  • Wang Xu An

摘要

Effective mobile network traffic forecasting demands datasets that integrate multivariate features while preserving temporal and spatial dependencies and avoiding feature leakage. We propose a spatiotemporal, causality-preserving preprocessing framework that converts raw measurements into temporally consistent and spatially informed inputs. Its four modules: Multiscale Temporal Feature Construction, Spatiotemporal Enhancement, Interaction and Encoding Layer, and Feature Selection, capture nonlinear, time-dependent, and spatial correlations. Experiments with Hist Gradient Boosting Regressor (HGBR) and Long Short-Term Memory (LSTM) models demonstrate improved model stability, generalization, and interpretability, while ablation studies confirm the unique contribution of each module. The framework enables robust and reproducible throughput prediction from real-world cellular data.