Skip to contents

Create Train/Validation/Test Split for Graph Data

graph_split.Rd

Creates splits for graph neural networks following rsample's structure. Unlike traditional ML, GCNs use the full graph during training but only compute loss on labeled (training) nodes.

Usage

graph_split(data, prop = c(0.6, 0.2, 0.2), seed = NULL)

Arguments

data

Data frame or adjacency matrix. The full dataset to split.

prop

Numeric vector of length 2 or 3. Proportions for splits.

  • Length 2: c(train, test) - creates train/test split

  • Length 3: c(train, val, test) - creates train/val/test split Must sum to 1.0. Default: c(0.6, 0.2, 0.2)

seed

Integer or NULL. Random seed for reproducibility. Default: NULL

Value

A graph_split object (list) containing:

  • data: Original data

  • train_id: Integer vector of training indices

  • val_id: Integer vector of validation indices (or NULL if length(prop) == 2)

  • test_id: Integer vector of test indices

Details

The proportions must sum to 1.0. The function creates non-overlapping splits where each row belongs to exactly one split.

For GCN training, you use:

  • Full X and A_sparse for all forward passes

  • IDs to select which predictions to use for loss computation

Examples

if (FALSE) { # \dontrun{
# Standard 60/20/20 split
split <- graph_split(A_sparse, seed = 42)

# Custom split (70/15/15)
split <- graph_split(A_sparse, prop = c(0.7, 0.15, 0.15))

# Two-way split (80/20 train/test)
split <- graph_split(A_sparse, prop = c(0.8, 0.2))

# Use in training
predictions <- model(X, A_sparse)
train_loss <- nnf_mse_loss(predictions[split$train_id], y[split$train_id])
} # }