DRAF-Net: Dual-Branch Residual-Guided Multi-View Attention Fusion Network for Station-Level Numerical Weather Prediction Correction

Remote Sensing, Journal Year: 2025, Volume and Issue: 17(2), P. 206 - 206

Published: Jan. 8, 2025

Accurate station-level numerical weather predictions are critical for disaster prevention and mitigation, with error correction playing an essential role. However, existing models struggle to effectively handle the high-dimensional features complex dependencies inherent in meteorological data. To address these challenges, this paper proposes dual-branch residual-guided multi-view attention fusion network (DRAF-Net), a novel deep learning-based model. DRAF-Net introduces two key innovations: (1) residual structure that enhances spatial sensitivity of improves output stability by connecting raw data shallow features, respectively; (2) mechanism spatiotemporal influences, temporal dynamics, associations, significantly improving representation dependencies. The effectiveness was validated on real-world datasets comprising observations from Chinese stations. It achieved average RMSE reduction 83.44% MAE 84.21% across all eight variables, outperforming other methods. Moreover, extensive studies confirmed contributions each component, while visualization results highlighted model’s ability eliminate anomalous values improve prediction consistency. code will be made publicly available support future research development.

Language: Английский

An improved graph factorization machine based on solving unbalanced game perception

Frontiers in Neurorobotics, Journal Year: 2024, Volume and Issue: 18

Published: Dec. 4, 2024

The user perception of mobile game is crucial for improving experience and thus enhancing profitability. sparse data captured in the can lead to sporadic performance model. This paper proposes a new method, balanced graph factorization machine (BGFM), based on existing algorithms, considering imbalance important high-dimensional features. categories are first by Borderline-SMOTE oversampling, then features represented naturally graph-structured way. highlight that BGFM contains interaction mechanisms aggregating beneficial results as edges graph. Next, combines (FM) neural network strategies concatenate any sequential feature interactions with an attention mechanism assigns inter-feature weights. Experiments were conducted collected dataset. proposed was compared eight state-of-the-art models, significantly surpassing all them AUC, precision, recall, F-measure indices.

Language: Английский