---
title: 'Classifying in latent space: k-NN & Random-Forest wrappers'
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{Classifying in latent space: k-NN & Random-Forest wrappers}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
params:
  family: red
css: albers.css
resource_files:
- albers.css
- albers.js
includes:
  in_header: |-
    <script src="albers.js"></script>
    <script>document.addEventListener('DOMContentLoaded',()=>document.body.classList.add('palette-red'));</script>

---

```{r setup, include = FALSE}
knitr::opts_chunk$set(
  collapse  = TRUE,
  comment   = "#>",
  fig.width = 7,
  fig.height = 4
)
library(multivarious)
library(dplyr)
library(tibble)
library(ggplot2) # Needed for plots
library(knitr)   # Needed for kable

```

```{css, echo=FALSE}
h1 {
  font-size: 1.8em; /* Adjust as needed */
  margin-top: 1.5em;
  margin-bottom: 0.8em;
  border-bottom: 1px solid #dee2e6; /* Optional: add a subtle separator */
  padding-bottom: 0.3em;
}
h2 {
  font-size: 1.4em; /* Adjust as needed */
  margin-top: 1.2em;
  margin-bottom: 0.6em;
}
/* Add rules for h3, h4 etc. if they are used and need adjustment */
```

# 1. Why classify after projection?

Once a dimensionality-reduction model (PCA, PLS, CCA, …) is fitted,
every new sample can be projected into the low-dimensional latent
space. Running a classifier there – instead of on thousands of
noisy raw variables – yields

*   fewer parameters & smaller models,
*   immunity to collinearity,
*   freedom to use partial data (ROI, missing sensors),
*   a clean separation between unsupervised decomposition and
    supervised prediction.

The `classifier()` S3 family supplied by `multiblock` provides that
glue: you hand it any projector (or `multiblock_biprojector`,
`discriminant_projector`, …) plus class labels → it returns a ready
predictor object.

# 2. Iris demo – LDA → `discriminant_projector` → k-NN

```{r data_classifier}
data(iris)
X   <- as.matrix(iris[, 1:4])
grp <- iris$Species

# Fit classical Linear DA and wrap it
if (!requireNamespace("MASS", quietly = TRUE)) {
  stop("MASS package required for LDA example")
}


# 1. Define and fit the pre-processing step using the training data
preproc_fitted <- fit(center(), X)
# 2. Transform the data
Xp <- transform(preproc_fitted, X)

# Assuming discriminant_projector, prep, center, scores are available
lda_fit <- MASS::lda(X, grouping = grp)


disc_proj <- multivarious::discriminant_projector(
  v      = lda_fit$scaling,                 # loadings (p × d)
  s      = Xp %*% lda_fit$scaling,           # scores   (n × d)
  sdev   = lda_fit$svd,                     # singular values
  labels = grp,
  preproc = preproc_fitted            # Pass the fitted pre-processor
)
print(disc_proj)
```

## 2.1 Visualise the latent space

```{r plot_latent_space}
scores_df <- as_tibble(scores(disc_proj)[, 1:2],
                       .name_repair = ~ c("LD1","LD2")) |>
  mutate(Species = iris$Species)

ggplot(scores_df, aes(LD1, LD2, colour = Species)) +
  geom_point(size = 2, alpha = .7) +
  stat_ellipse(level = .9, linewidth = .3) +
  theme_minimal() +
  ggtitle("Iris – first two LDA components")
```

## 2.2 Build a k-NN classifier on the latent scores

```{r build_knn_classifier}
set.seed(42)
train_id <- sample(seq_len(nrow(X)), size = 0.7*nrow(X))
test_id  <- setdiff(seq_len(nrow(X)), train_id)

# Assuming classifier function is available
clf_knn <- multivarious::classifier(
  x       = disc_proj,
  labels  = grp[train_id],
  new_data= X[train_id, ],  # Use training data to get reference scores
  knn     = 3
)
print(clf_knn)
```

## 2.3 Predict and evaluate

```{r predict_knn}
pred_knn <- predict(clf_knn, new_data = X[test_id, ],
                    metric = "cosine", prob_type = "knn_proportion")

head(pred_knn$prob, 3)
print(paste("Overall Accuracy:", mean(pred_knn$class == grp[test_id])))

# Assuming rank_score and topk are available
rk  <- rank_score(pred_knn$prob, grp[test_id])
tk2 <- topk      (pred_knn$prob, grp[test_id], k = 2)

tibble(
  prank_mean = mean(rk$prank),
  top2_acc   = mean(tk2$topk)
)
```

## 2.4 Confusion-matrix on the test set

```{r plot_confusion_matrix}
cm <- table(
  Truth     = grp[test_id],
  Predicted = pred_knn$class
)

# Heat-map
cm_df <- as.data.frame(cm)
ggplot(cm_df, aes(Truth, Predicted, fill = Freq)) +
  geom_tile(colour = "grey80") +
  geom_text(aes(label = Freq), colour = "white", size = 4) +
  scale_fill_gradient(low = "#4575b4", high = "#d73027", name="Count", limits = c(0, 15)) +
  scale_y_discrete(limits = rev(levels(cm_df$Predicted))) +
  theme_minimal(base_size = 12) + coord_equal() +
  ggtitle("k-NN (k = 3) confusion matrix – test set") +
  theme(axis.text.x = element_text(angle = 45, hjust = 1))


# Pretty table as well
knitr::kable(cm, caption = "Confusion matrix (counts)")
```

# 3. Random-Forest on the same latent space

```{r build_rf_classifier}
# Check if randomForest is installed
if (requireNamespace("randomForest", quietly = TRUE)) {

  # Assuming rf_classifier.projector method is available
  rf_clf <- rf_classifier( # Using the generic here
    x       = disc_proj,
    labels  = grp[train_id],
    # Pass scores directly if method requires it, or let it call scores(x)
    scores  = scores(disc_proj)[train_id, ]
  )

  pred_rf <- predict(rf_clf, new_data = X[test_id, ])
  print(paste("RF Accuracy:", mean(pred_rf$class == grp[test_id])))
} else {
  cat("randomForest package not installed. Skipping RF example.\n")
}

```

The RF sees exactly three input variables (the LDA components) – that
keeps trees shallow and speeds-up training.

# 4. Partial-feature prediction: sepal block only

Assume that in deployment we measure only Sepal variables
(cols 1–2). A partial projection keeps the classifier happy:

```{r predict_partial}
sepal_cols <- 1:2

# Create a classifier using reference scores from Sepal columns only
clf_knn_sepal <- multivarious::classifier(
  x       = disc_proj,
  labels  = grp[train_id],
  new_data= X[train_id, sepal_cols],  # Use training data subset
  colind  = sepal_cols,             # Indicate which columns were used
  knn     = 3
)

# Predict using the dedicated sepal classifier
pred_sepal <- predict(
  clf_knn_sepal,                  # Use the sepal-specific classifier
  new_data = X[test_id, sepal_cols]
  # No need for colind here as clf_knn_sepal expects sepal data
)

print(paste("Accuracy (Sepal only):", mean(pred_sepal$class == grp[test_id])))
```
Accuracy drops a bit – as expected when using fewer features.

# 5. Which component block matters most?

`feature_importance()` can rank variable groups via a simple
"leave-block-out" score drop.

```{r calc_feature_importance}
blocks <- list(
  Sepal = 1:2,
  Petal = 3:4
)

# Assuming feature_importance is available
fi <- feature_importance(
  clf_knn,
  new_data  = X[test_id, ],
  true_labels = grp[test_id], # Pass the correct test set labels
  blocks    = blocks,
  fun       = rank_score, # Use rank_score as the performance metric
  fun_direction = "lower_is_better",
  approach  = "marginal" # Calculate marginal drop when block is removed
)
print(fi)
```