Year: 2025
Type: conference-paper
Abstract: Smart government platforms face challenges in handling heterogeneous data and missing information in knowledge graphs. This paper presents MSPA, an efficient time series prediction model combining multi-scale pooling and attention mechanisms. The model converts 1D time series to 2D format, applies multi-scale pooling (2×2, 3×3, 4×4 kernels) for feature compression, and uses single-head attention to capture global dependencies while reducing computational complexity from O(H×L²) to O(0.42L²). MSPA effectively predicts missing data dynamics to improve knowledge graph integrity. Experiments demonstrate superior prediction accuracy and sixfold efficiency improvement over existing methods, providing a novel solution for intelligent government decision-making and urban governance.
Authors: Wenhan Song, Enguang Zuo, Ziwei Yan, Junyu Zhu
Institutions: Xinjiang University, Tsinghua University
Language: English
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4414931908