graph_zoom: GraphZoom Main Function with Optional Direct Feature Graph

In bbuchsbaum/graphweights: Constructing Adjacency Matrices Based on Spatial and Feature Similarity

GraphZoom Main Function with Optional Direct Feature Graph

Description

Perform GraphZoom pipeline as described in the paper.

If 'A_feat' is provided, it will be used directly for graph fusion instead of computing a feature graph from X. If both 'A_feat' and 'X' are provided, 'A_feat' takes precedence.

Usage

graph_zoom(
  A,
  X = NULL,
  A_feat = NULL,
  levels = 2,
  t = 10,
  kNN = 10,
  beta = 1,
  embed_dim = 16,
  filter_power = 2,
  sigma = 0.001,
  verbose = FALSE
)

Arguments

A	Sparse adjacency matrix (N x N) of the original graph topology.
X	Node feature matrix (N x K). If NULL, no fusion from features is done unless A_feat is provided.
A_feat	Optional precomputed feature graph (N x N). If provided, used directly for fusion.
levels	Number of coarsening levels.
t	Dimension for local spectral embedding during coarsening.
kNN	Number of nearest neighbors for feature-based kNN (ignored if A_feat is given).
beta	Weight to balance topology and feature graph in fusion.
embed_dim	Dimension of embeddings at coarsest level.
filter_power	The power k in Eq.(7) for refinement.
sigma	Self-loop factor in refinement filter.
verbose	If TRUE, print progress messages.

Value

A list:

E_original	Refined embeddings for the original graph (N x embed_dim).
graphs	List of graphs from level 0 to level L (fused and coarsened).
H_list	List of mapping operators.
E_coarse	Embeddings at the coarsest level.

References

GraphZoom Paper (Sections 3.1-3.4)

Examples

library(Matrix)
library(RcppHNSW)
set.seed(42)
N <- 100
A <- rsparsematrix(N, N, 0.01)
A <- (A + t(A))/2
diag(A) <- 0
X <- matrix(rnorm(N*20), nrow=N, ncol=20)
# Run with computed feature graph:
result <- graph_zoom(A, X, levels=2, embed_dim=8, kNN=5)

# If you have your own A_feat:
# Suppose A_feat is a NxN matrix you constructed from features externally:
# result <- graph_zoom(A, A_feat=my_feature_graph, levels=2, embed_dim=8)

E <- result$E_original

bbuchsbaum/graphweights documentation built on Dec. 14, 2024, 1:13 a.m.