In recent years, with the continuous development of technology in the field of artificial intelligence, the field of human-computer interaction has attracted more scholars' attention. Studies have shown that the characteristic power spectrum density extracted by EEG (electroencephalogram) is sensitive to the change of mental load, but its dimension is too high, which is prone to data disaster. LLE (locally linear embedding) is a commonly used nonlinear dimensionality reduction algorithm, which makes up for the shortcoming of traditional linear dimensionality reduction algorithms that cannot find nonlinear structural relationships in data. Because of the different sparsity and distortion degree of sample distribution in different data sets, the optimal neighborhood parameters are different when using LLE to reduce dimension of different data sets. The relationship between Euclidean distance and geodesy distance between sample points was used to quantify the distortion degree of the data set. VK-LLE (variable k-locally linear embedding) algorithm of adaptive neighborhood parameters dynamically adjusts the optimal neighborhood parameters of each data set. The interference caused by different distortion degree of sample distribution to dimensionality reduction effect was solved. The experimental results show that the classification accuracy of data after VK-LLE dimensionality reduction using SVM (support vector machine) is generally higher than that after traditional LLE dimensionality reduction using SVM, and it has stronger adaptability to complex data sets.