Methods
14 encoders · 5 graph constructions · 13 baselines
Graph Encoders
14 encoder variants
Graph Attention Network (Veličković 2018)
Dynamic attention Graph Attention v2 (Brody 2022) — scCCVGBen extension
TransformerConv — Attention is All You Need graph variant (Shi 2020)
Self-supervised edge prediction GAT (Kim 2020) — scCCVGBen extension
Graph Convolutional Network (Kipf 2017)
GraphSAGE inductive (Hamilton 2017)
GraphConv — general message-passing variant
ChebNet — Chebyshev polynomial filters (Defferrard 2016)
Topology-Adaptive Graph Conv (Du 2017)
ARMA filter Conv (Bianchi 2021)
Simplified Graph Conv (Wu 2019)
Simple Spectral Graph Conv (Zhu 2021)
Graph Isomorphism Network (Xu 2019) — scCCVGBen extension
Dynamic Edge Conv (Wang 2019) — scCCVGBen extension
Graph Constructions
5 graph strategies
Standard k-NN with Euclidean distance (k=15) — scCCVGBen benchmark default
k-NN with cosine similarity — rewards direction, invariant to magnitude
Shared Nearest Neighbour — 2 cells connected if they share a fraction of neighbours
Mutual k-NN — only edges where both cells are in each other's k-NN list; stricter connectivity
Gaussian heat-kernel weights w=exp(-d²/(2σ²)); edges pruned at threshold 0.9
Comparators / Baselines
13 baseline methods
Linear PCA — sklearn.decomposition.PCA
Kernel PCA (RBF) — sklearn.decomposition.KernelPCA
Independent Component Analysis (FastICA) — sklearn.decomposition.FastICA
Factor Analysis — sklearn.decomposition.FactorAnalysis
Non-negative Matrix Factorisation — sklearn.decomposition.NMF
Truncated SVD — sklearn.decomposition.TruncatedSVD
Dictionary Learning — sklearn.decomposition.DictionaryLearning
Single-cell Variational Inference — Lopez 2018
DIP-VAE disentangled autoencoder — Kumar 2017
InfoVAE — Zhao 2017
β-TCVAE — Chen 2018
Hyper-parameterised VAE with high β (β=100)
Core scCCVGBen reference row; encoder and graph-axis variants are labelled separately