Research On Dynamic Brain Network Feature Dimensionality Reduction Algorithm Based On Adaptive Weighted T-SNE

Functional Magnetic Resonance Imaging(fMRI)is an important brain imaging technique.Among them,brain function network reconstruction technology based on resting state fMRI provides an effective method for studying brain function and characteristics.Most studies currently assume that the brain is stationary during the resting fMRI data collection,but in fact the brain is a whole and constantly changing activity.Therefore,some researchers have developed the brain network from the perspective of dynamic system theory.The study of dynamic performance,and the use of sliding windows to construct the whole brain brain network state observation matrix,as the matrix is up to thousands of dimensions,it is difficult to directly observe its main characteristics,which brings difficulties to further study the dynamic characteristics of brain networks.In the research process,the t-distributed stochastic neighbor embedding(t-SNE)algorithm has some effect on dimension reduction of the above matrix,but due to the similarity probability in calculating high-dimensional spatial samples.At that time,the actual distribution of the sample was not considered,so the dimensionality reduction effect was not good.This paper proposes an adaptive weighted improved t-SNE algorithm,which first normalizes the Euclidean distance of each pair of sample points in a high-dimensional space,and then performs group analysis on the different distributions of the actual sample pairs.According to the three cases of close distance,moderate distance and long distance,they are adaptively weighted and calculate the similarity probability between the sample points in the high-dimensional space.The weighted relative distance is used instead of the Euclidean absolute distance to achieve the high-dimensional sample similarity.More accurate measurement.In order to verify the effect of self-adaptive weighted t-SNE algorithm on the dimensionality reduction of brain state observation matrix,this paper conducts experiments and analysis from the following three aspects: Self-adaptive weighted t-SNE algorithm and other dimensionality reduction algorithms Reduced dimensional comparison test on the sample;reduced-dimensional comparison test of adaptive weighted t-SNE algorithm and traditional t-SNE algorithm on self-acquiredmulti-samples;adaptive weighted t-SNE algorithm and traditional t-SNE algorithm in open database Reduced dimensional comparison tests on multiple samples.The visualization of dimensionality reduction shows that the reduced-dimensional clustering visualization effect of self-adaptive weighted t-SNE is superior to the traditional t-SNE algorithm and other dimensionality reduction algorithms,and effectively eliminates the crossover and scatter in the traditional t-SNE algorithm.Clusters are not concentrated and other issues.The performance index of brain network state clustering after dimensionality reduction shows that the clustering index Davies-Bouldin Index value and Dunn Index value of the self-adaptive weighted t-SNE algorithm are significantly improved compared with the traditional t-SNE algorithm,thus further studying the dynamics of the brain network.Features provide a strong foundation.