Prompt Description

You are an experienced scientific illustration designer. Please carefully read the literature information I provide, fully understand the research content, and generate a research paradigm diagram that can be used for scientific publication. Style: EEG-based emotion recognition 3.2.2.3 Spatio-temporal Grouping Fusion Submodule. The spatio-temporal grouping fusion submodule serves as the core interaction unit of the spatio-temporal dynamic modeling module. Its primary goal is to address the key issues of "non-stationarity in the time dimension, dynamic spatial dependence, and separation of spatio-temporal features" in EEG signals. This module employs a progressive architecture of "dynamic attention grouping - spatio-temporal joint modeling (supported by IDGCN) - gated interaction update - tree-like multi-branch fusion" to deeply couple temporal local patterns with multi-order spatial dependencies, while adapting to the spatio-temporal dynamic changes of EEG signals under different emotional states. The module is closely connected to the dynamic graph construction submodule and the diffusion convolution submodule, relying on the improved dynamic graph convolutional network (IDGCN) to achieve the organic fusion of spatial features guided by dynamic topology and temporal sequence features, providing highly discriminative spatio-temporal coupling features for subsequent multi-branch hierarchical integration. (a) Dynamic Attention Grouping Unit This unit adaptively groups features based on attention weights, breaking through the limitations of fixed window partitioning, enabling the grouping results to accurately match the spatio-temporal distribution characteristics of the data, and screening high-value feature regions for subsequent spatio-temporal joint modeling by IDGCN. (a) Attention Weight Calculation First, a lightweight convolutional network is used to compress the dimension and perform nonlinear transformation on the input features, adaptively learning the grouping attention weights. The formula is: (3-16) Where: is the input spatio-temporal feature tensor ( is the batch size, is the number of feature channels, is the number of electrode nodes, and is the time step); is the channel dimensionality reduction convolution, which reduces the number of channels from to , reducing computational overhead while retaining key features; is the weight prediction convolution, which further compresses the number of channels to 2, outputting the initial weights of two groups; is normalized along the channel dimension, so that the sum of the weights of the two groups is 1, ensuring feature energy conservation, and finally obtaining the attention weight tensor , whose element represents the weight of the feature of the electrode and the time step in the sample belonging to the first group. (b) Dynamic Group Generation Based on the learned attention weights, the original features are weighted and grouped to obtain two complementary subspace features. The formulas are: (3-17) (3-18) Where, and are the first and second group features, respectively. The core advantage of this grouping method is that the attention weights dynamically change with the input features, which can automatically focus on emotion-related key spatio-temporal regions and suppress noise interference such as electrooculography and electromyography, providing a high signal-to-noise ratio input for the subsequent efficient modeling of IDGCN. (b) IDGCN Unit: The Core Carrier of Spatio-temporal Joint Modeling This unit is the core computing unit of the module. Its design goal is to simultaneously complete "time dependence capture - dynamic spatial dependence modeling - group interaction update" for a single group of features. By integrating temporal convolution, dynamic graph convolution (DGCN), gating mechanism, and residual adjustment, it achieves deep coupling and refinement of spatio-temporal features.