A0543
Title: The Kriging-correlation score for feature importance in high dimensional clustering
Authors: Cheng-Han Chua - Tunghai University (Taiwan) [presenting]
Abstract: A Kriging-correlation (KC) score is proposed to identify the most critical clustering features from high-dimensional datasets. The KC score ranks feature importance through a three-stage procedure. In the first stage, an appropriate distance metric is learned from the data, and the original data is projected into a two-dimensional space using the t-distributed stochastic neighbor embedding (t-SNE) method. The resulting 2D coordinates are treated as essential clustering variables in the reduced space. In the second stage, each original feature is reconstructed from the reduced space using the AutoFRK method. In the final stage, feature importance is assessed by measuring the dependence between each original feature and its reconstructed version; the number of significant features is determined using MANOVA. The KC score is applied to four single-cell datasets, and its performance is compared with the Laplacian score, a widely used unsupervised feature ranking method. The results demonstrate that the KC score can identify a smaller set of features while achieving equal or better classification accuracy compared to the Laplacian score.
