Prompt Description

The diffusion convolution submodule serves as the core spatial feature extraction unit within the spatiotemporal dynamic modeling module. Its primary function involves utilizing the sparse dynamic adjacency matrix generated by the dynamic graph construction submodule to achieve multi-order directed diffusion and multi-scale aggregation of node features on the dynamic topology. This precisely captures the nonlinear functional connections and multi-hop spatial dependencies between electrodes. It also adapts to the directed characteristics of EEG signal neural transmission, compensating for the loss of directionality information in traditional spectral graph convolutions. This module forms a close connection with the dynamic graph construction submodule through "topology generation-feature diffusion." The dynamic graph provides sample-specific spatial association constraints, while diffusion convolution completes the deep spatial modeling of features based on these constraints. (1) Multi-Order Directed Diffusion Iteration Let the diffusion order be (a hyperparameter), the input feature tensor be (where is the batch size, is the number of feature channels, is the number of nodes/electrodes, and is the number of time steps), and the dynamic adjacency matrix be (generated by the dynamic graph construction submodule through a "prior-guided + data-driven" dual path and optimized by Top-K sparsification). The diffusion process is performed iteratively, and the feature propagation formula for the iteration (k=1,2,...,K) is: (3-9) where, is the initial input feature; is the tensor multiplication operator based on Einstein summation convention, which is used to realize the batch matrix operation of features and adjacency matrix, and its specific element calculation expression is: (3-10) This equation clearly reflects the directionality and weighting characteristics of the diffusion process: the order diffusion feature of node in the sample, channel, and time step is the weighted sum of the order features of all its predecessor nodes with the adjacency matrix element as the weight. The weight is generated by the dynamic graph construction submodule and directly represents the directed association strength of node to node (such as the probability of neural signals being transmitted from electrode to electrode ), perfectly matching the directional transmission characteristics of EEG neural signals. It is worth noting that since the dynamic graph construction submodule has removed redundant weak connections through Top-K sparsification optimization, is a sparse matrix. The above summation process only needs to calculate the Top-K strongly correlated neighbors of each node, which significantly reduces the computational complexity of multi-order diffusion and avoids noise interference from weak connections. (2) Multi-Scale Feature Fusion After diffusion iterations, feature tensors of different scales are obtained, , ,..., (each tensor has a dimension of , corresponding to the feature aggregation results of 1st to order neighbors - that is, integrates direct neighbor information, integrates 2-hop neighbor information, and so on. In order to fully integrate multi-scale spatial information and avoid the limitations of single-scale features, these feature tensors are first concatenated along the feature channel dimension to obtain the concatenated feature tensor: (3-11) where is the channel dimension concatenation operator. After concatenation, , and its number of channels is times the initial number of channels, integrating multi-range spatial feature information from local to global. In order to unify the feature channel dimension (consistent with the input channel number ) to facilitate subsequent module connection and residual connection, a 1×1 convolution is used to reduce the dimension of the concatenated features.