Reorders features by hierarchical clustering on feature-feature correlation
distance, then fills a grid in row-major order. Features that co-vary end up
as neighbors in the image, helping the CNN's local receptive fields exploit
correlation structure.
Must be FIT on training data only to avoid leakage.
Usage
fit_corr_order(X_train, method = "average")
Arguments
- X_train
Numeric matrix (n_train × p features) for fitting the order
- method
Linkage method for hclust (default: "average")
Value
Integer vector of length p, the feature order to use for encoding
References
Sharma A et al. (2019). DeepInsight: A methodology to transform a
non-image data to an image for convolution neural network architecture.
Scientific Reports, 9, 11399.
Examples
X_train <- matrix(rnorm(30), nrow = 6, ncol = 5)
fit_corr_order(X_train)
#> [1] 3 1 2 4 5