HAL-X: Scalable hierarchical clustering for rapid and tunable single-cell analysis
Fig 2
Out-of-sample accuracy vs. weighted F-score (comparison to ground-truth) for various clustering algorithms.
We used out-of-the-box clustering algorithms DBSCAN, spectral clustering, K-means and Meanshift with various hyper-parameters and clustered 6 different benchmark two-dimensional datasets: Noisy Circles, Gaussian mixtures with different covariance structures, non-convex clusters (Noisy Moons) and a random noise map (No Structure). These datasets were taken from scikit-learn clustering methods as benchmark datasets (see S1 Methods for more details). The out-of-sample accuracy is computed by training a random forest classifier with 100 estimators on the clustering labels provided by the clustering algorithms. We use a 80/20 train/test split to train the classifiers and evaluate the out-of-sample error. Notice that the F-score and out-of-sample accuracy as measured are monotonically related.