fdars_core/classification/
mod.rs

1//! Functional classification with mixed scalar/functional predictors.
2//!
3//! Implements supervised classification for functional data using:
4//! - [`fclassif_lda`] / [`fclassif_qda`] — FPC + LDA/QDA pipeline
5//! - [`fclassif_knn`] — FPC + k-NN classifier
6//! - [`fclassif_kernel`] — Nonparametric kernel classifier with mixed predictors
7//! - [`fclassif_dd`] — Depth-based DD-classifier
8//! - [`fclassif_cv`] — Cross-validated error rate
9//!
10//! ## Parameter ordering convention
11//! All classifiers: `(data, y, [argvals,] [scalar_covariates,] method-params)`
12
13use crate::error::FdarError;
14use crate::matrix::FdMatrix;
15use crate::regression::fdata_to_pc_1d;
16
17pub mod cv;
18pub mod dd;
19pub mod fit;
20pub mod kernel;
21pub mod knn;
22pub mod lda;
23pub mod qda;
24
25#[cfg(test)]
26mod tests;
27
28// ---------------------------------------------------------------------------
29// Shared types
30// ---------------------------------------------------------------------------
31
32/// Classification result.
33#[derive(Debug, Clone, PartialEq)]
34#[cfg_attr(feature = "serde", derive(serde::Serialize, serde::Deserialize))]
35#[non_exhaustive]
36pub struct ClassifResult {
37    /// Predicted class labels (length n)
38    pub predicted: Vec<usize>,
39    /// Posterior/membership probabilities (n x G) — if available
40    pub probabilities: Option<FdMatrix>,
41    /// Training accuracy
42    pub accuracy: f64,
43    /// Confusion matrix (G x G): row = true, col = predicted
44    pub confusion: Vec<Vec<usize>>,
45    /// Number of classes
46    pub n_classes: usize,
47    /// Number of FPC components used
48    pub ncomp: usize,
49}
50
51/// Cross-validation result.
52#[derive(Debug, Clone, PartialEq)]
53#[non_exhaustive]
54pub struct ClassifCvResult {
55    /// Mean error rate across folds
56    pub error_rate: f64,
57    /// Per-fold error rates
58    pub fold_errors: Vec<f64>,
59    /// Best ncomp (if tuned)
60    pub best_ncomp: usize,
61}
62
63// ---------------------------------------------------------------------------
64// Utility helpers
65// ---------------------------------------------------------------------------
66
67/// Count distinct classes and remap labels to 0..G-1.
68pub(crate) fn remap_labels(y: &[usize]) -> (Vec<usize>, usize) {
69    let mut labels: Vec<usize> = y.to_vec();
70    let mut unique: Vec<usize> = y.to_vec();
71    unique.sort_unstable();
72    unique.dedup();
73    let g = unique.len();
74    for label in &mut labels {
75        *label = unique.iter().position(|&u| u == *label).unwrap_or(0);
76    }
77    (labels, g)
78}
79
80/// Build confusion matrix (G x G).
81fn confusion_matrix(true_labels: &[usize], pred_labels: &[usize], g: usize) -> Vec<Vec<usize>> {
82    let mut cm = vec![vec![0usize; g]; g];
83    for (&t, &p) in true_labels.iter().zip(pred_labels.iter()) {
84        if t < g && p < g {
85            cm[t][p] += 1;
86        }
87    }
88    cm
89}
90
91/// Compute per-class means, counts, and priors from labeled features.
92pub(crate) fn class_means_and_priors(
93    features: &FdMatrix,
94    labels: &[usize],
95    g: usize,
96) -> (Vec<Vec<f64>>, Vec<usize>, Vec<f64>) {
97    let n = features.nrows();
98    let d = features.ncols();
99    let mut counts = vec![0usize; g];
100    let mut class_means = vec![vec![0.0; d]; g];
101    for i in 0..n {
102        let c = labels[i];
103        counts[c] += 1;
104        for j in 0..d {
105            class_means[c][j] += features[(i, j)];
106        }
107    }
108    for c in 0..g {
109        if counts[c] > 0 {
110            for j in 0..d {
111                class_means[c][j] /= counts[c] as f64;
112            }
113        }
114    }
115    let priors: Vec<f64> = counts.iter().map(|&c| c as f64 / n as f64).collect();
116    (class_means, counts, priors)
117}
118
119/// Accuracy from labels.
120fn compute_accuracy(true_labels: &[usize], pred_labels: &[usize]) -> f64 {
121    let n = true_labels.len();
122    if n == 0 {
123        return 0.0;
124    }
125    let correct = true_labels
126        .iter()
127        .zip(pred_labels.iter())
128        .filter(|(&t, &p)| t == p)
129        .count();
130    correct as f64 / n as f64
131}
132
133/// Extract FPC scores and append optional scalar covariates.
134pub(crate) fn build_feature_matrix(
135    data: &FdMatrix,
136    scalar_covariates: Option<&FdMatrix>,
137    ncomp: usize,
138) -> Result<(FdMatrix, Vec<f64>, FdMatrix, Vec<f64>), FdarError> {
139    let m = data.ncols();
140    let argvals: Vec<f64> = (0..m).map(|j| j as f64 / (m - 1).max(1) as f64).collect();
141    let fpca = fdata_to_pc_1d(data, ncomp, &argvals)?;
142    let n = data.nrows();
143    let d_pc = fpca.scores.ncols();
144    let d_cov = scalar_covariates.map_or(0, super::matrix::FdMatrix::ncols);
145    let d = d_pc + d_cov;
146
147    let mut features = FdMatrix::zeros(n, d);
148    for i in 0..n {
149        for j in 0..d_pc {
150            features[(i, j)] = fpca.scores[(i, j)];
151        }
152        if let Some(cov) = scalar_covariates {
153            for j in 0..d_cov {
154                features[(i, d_pc + j)] = cov[(i, j)];
155            }
156        }
157    }
158
159    Ok((features, fpca.mean, fpca.rotation, fpca.weights))
160}
161
162// ---------------------------------------------------------------------------
163// Re-exports — preserves the external API
164// ---------------------------------------------------------------------------
165
166pub use cv::fclassif_cv;
167pub use dd::fclassif_dd;
168pub(crate) use fit::classif_predict_probs;
169pub use fit::{
170    fclassif_cv_with_config, fclassif_knn_fit, fclassif_lda_fit, fclassif_qda_fit, ClassifCvConfig,
171    ClassifFit, ClassifMethod,
172};
173pub use kernel::{fclassif_kernel, kernel_classify_from_distances};
174pub use knn::{fclassif_knn, knn_classify_from_distances};
175pub use lda::fclassif_lda;
176pub use qda::fclassif_qda;
fdars_core/classification/mod.rs

fdars_core/classification/
mod.rs
