Evaluation Metrics
Understand the 24 metrics used for model evaluation across 4 categories
Filter by Category
Showing 24 of 24 metrics
NMI
Normalized Mutual Information
Measures the mutual information between predicted clusters and true labels, normalized by the entropy of the two distributions
ARI
Adjusted Rand Index
Similarity between predicted and true clusters adjusted for chance
ASW
Silhouette Score (Average)
Average silhouette coefficient measuring cluster cohesion and separation
DAV
Davies-Bouldin Index
Average similarity between each cluster and its most similar neighboring cluster
CAL
Calinski-Harabasz Index
Ratio of between-cluster to within-cluster dispersion
COR
Correlation-Based Distance
Measure based on Pearson correlation of latent representations
UMAP_Dist
UMAP Distance Correlation
Spearman correlation between latent space and UMAP-reduced pairwise distance matrices
UMAP_Q_local
UMAP Local Structure Quality
Average coranking quality for local neighborhoods in UMAP space
UMAP_Q_global
UMAP Global Structure Quality
Average coranking quality for global relationships in UMAP space
UMAP_Overall
UMAP Overall Embedding Quality
Comprehensive UMAP quality combining distance correlation, local and global preservation
tSNE_Dist
t-SNE Distance Correlation
Spearman correlation between latent space and t-SNE-reduced pairwise distance matrices
tSNE_Q_local
t-SNE Local Structure Quality
Average coranking quality for local neighborhoods in t-SNE space
tSNE_Q_global
t-SNE Global Structure Quality
Average coranking quality for global relationships in t-SNE space
tSNE_Overall
t-SNE Overall Embedding Quality
Comprehensive t-SNE quality combining distance correlation, local and global preservation
Manifold_Dim
Manifold Dimensionality Efficiency
Multi-method dimensionality efficiency score combining variance thresholds, Kaiser criterion, elbow detection, and spectral decay
Spectral_Decay
Spectral Decay Rate
Rate of eigenvalue decay indicating information concentration in leading dimensions
Part_Ratio
Participation Ratio Score
Effective dimensionality measure from eigenvalue distribution (trajectory: lower is better, steady-state: higher is better)
Anisotropy
Anisotropy Score
Multi-method anisotropy combining log-ellipticity, condition numbers, ratio variance, entropy, dominance, and effective dimensionality
Traj_Dir
Trajectory Directionality
Dominance of primary developmental axis relative to other directions
Noise_Resil
Noise Resilience
Signal-to-noise ratio based on leading vs. trailing PCA components
Core_Quality
Core Latent Space Quality
Fundamental quality score combining manifold, spectral, participation, and anisotropy metrics
Overall_LSE
Overall Intrinsic Quality
Comprehensive latent space quality with data-type-aware weighting
Train_Time
Training Time
Total time to train the model on the dataset
Inference_Time
Inference Time
Time to embed all cells through the trained encoder
Metric Categories
🎯 Clustering & Cell Type Discovery
Supervised metrics comparing predicted clusters to ground truth labels
6 metrics
📊 Embedding Quality (UMAP & t-SNE)
Visualization quality via coranking analysis (4 metrics × 2 methods)
8 metrics
🔬 Intrinsic Latent Space (LSE)
Unsupervised geometric, spectral, and topological properties
8 metrics
âš¡ Computational Efficiency
Training and inference performance
2 metrics