Fig. 4: Applying unsupervised Laplacian score features selection to VoPo features.
From: VoPo leverages cellular heterogeneity for predictive modeling of single-cell data
Unsupervised feature selection was applied independently to the frequency-based features engineered from each metaclustering solution. Here, we show distributions of classification accuracies (AUC) with (purple boxplots) and without (gray boxplots) feature selection applied to features extracted with VoPo. The dashed orange horizontal line shows the mean baseline classification accuracy using the features obtained without the use of the VoPo pipeline. In all three datasets VoPo’s robust feature extraction (gray) improved the results over baseline (orange boxplots). Additionally, VoPo’s robust feature selection followed by unsupervised feature selection further improved classification accuracy (purple boxplots). The boxplots show median values, interquartile range, whiskers of 1.5 times interquartile range, and all individual points.
