A dynamic graph neural network classification method incorporating machine vision features
TL;DR
A dynamic graph neural network classification method integrating machine vision mapping and spatiotemporal evolution that effectively improves the generalization accuracy and anti-interference capability of heterogeneous network entity classification models.
Abstract
This paper addresses the limitations of node classification feature dimensionality and insufficient robustness of models in extreme scenarios within complex heterogeneous interactive networks. A dynamic graph neural network classification method integrating machine vision mapping and spatiotemporal evolution is proposed. This method introduces a multilayer self-attention-based visual feature extraction module, converting discrete high-frequency time series into a two-dimensional grid to extract morphological representations. Gated graph convolutional units are used to achieve cross-modal aggregation of local features and dynamic topology. The system incorporates a nulled neural network continuous evolver to optimize the convergence trajectory of the classification loss function under distributional bias. Quantitative testing shows that, under a distributed GPU computing architecture, the model achieves a comprehensive classification accuracy of 94.2% across all categories, with a single-batch inference latency controlled within 12.5ms, and a resilience index of 95.1 in structural perturbation environments. This research effectively improves the generalization accuracy and anti-interference capability of heterogeneous network entity classification models.
