Skip to contents

Evaluate classification performance of factor embeddings

Source: R/classifier_metrics.R
classify_embedding.Rd

Trains a k-nearest-neighbor classifier on factor embeddings and evaluates on held-out test samples. Returns a comprehensive metrics object.

Usage

classify_embedding(
  embedding,
  labels,
  test_fraction = 0.2,
  test_idx = NULL,
  k = 5L,
  seed = NULL,
  distance = c("euclidean", "cosine")
)

Arguments

embedding

Numeric matrix where rows are samples and columns are features (e.g., t(result$H) for sample embeddings).

labels

Integer or factor vector of class labels. Length must equal nrow(embedding).

test_fraction

Fraction of samples held out for testing (default 0.2).

test_idx

Optional integer vector of test indices. If provided, test_fraction is ignored.

k

Number of nearest neighbors (default 5).

seed

Random seed for train/test split reproducibility.

distance

Distance metric: "euclidean" (default) or "cosine".

Value

An fn_classifier_eval object with fields:

accuracy

Overall test accuracy

per_class

Data frame with per-class precision, recall, F1, support

macro_precision

Macro-averaged precision

macro_recall

Macro-averaged recall

macro_f1

Macro-averaged F1

weighted_f1

Support-weighted F1

auc

Macro-averaged one-vs-rest AUC (from neighbor vote fractions)

confusion

Confusion matrix (rows = true, cols = predicted)

predictions

Test set predictions

test_labels

True test labels

train_idx

Training sample indices

test_idx

Test sample indices

k

Number of neighbors used

See also

classify_logistic, classify_rf

Examples

# \donttest{
data(digits)
model <- nmf(digits, 10, maxit = 20, seed = 1, verbose = FALSE)
labels <- attr(digits, "target")
eval <- classify_embedding(model$w, labels, test_fraction = 0.2, k = 5, seed = 42)
print(eval)
#> Classification Evaluation (5-NN, euclidean distance)
#>   Samples: 1438 train, 359 test, 10 classes
#>   Accuracy:  0.9499
#>   Macro F1:  0.9491  (P=0.9516, R=0.9499)
#>   Weighted F1: 0.9494
#>   AUC (macro): 0.9949
#> 
#> Per-class:
#>  class precision recall     f1 support
#>      0    0.9355 1.0000 0.9667      29
#>      1    0.8500 0.9444 0.8947      36
#>      2    0.9070 1.0000 0.9512      39
#>      3    0.9429 0.9429 0.9429      35
#>      4    1.0000 0.9500 0.9744      40
#>      5    1.0000 0.9706 0.9851      34
#>      6    1.0000 1.0000 1.0000      39
#>      7    0.9444 1.0000 0.9714      34
#>      8    0.9643 0.7941 0.8710      34
#>      9    0.9722 0.8974 0.9333      39
# }