fix clippy error

* implemented multiclass for svc
* modified the multiclass svc so it doesnt modify the current api

.github/CODEOWNERS

@@ -2,6 +2,5 @@
 # the repo. Unless a later match takes precedence,
 # Developers in this list will be requested for
 # review when someone opens a pull request.
-*       @VolodymyrOrlov
 *       @morenol
 *       @Mec-iS

.github/CONTRIBUTING.md

@@ -50,9 +50,9 @@ $ rust-code-analysis-cli -p src/algorithm/neighbour/fastpair.rs --ls 22 --le 213
 
 1. After a PR is opened maintainers are notified
 2. Probably changes will be required to comply with the workflow, these commands are run automatically and all tests shall pass:
-    * **Coverage** (optional): `tarpaulin` is used with command `cargo tarpaulin --out Lcov --all-features -- --test-threads 1`
     * **Formatting**: run `rustfmt src/*.rs` to apply automatic formatting
     * **Linting**: `clippy` is used with command `cargo clippy --all-features -- -Drust-2018-idioms -Dwarnings`
+    * **Coverage** (optional): `tarpaulin` is used with command `cargo tarpaulin --out Lcov --all-features -- --test-threads 1`
     * **Testing**: multiple test pipelines are run for different targets
 3. When everything is OK, code is merged.
 

.github/workflows/ci.yml

@@ -19,14 +19,13 @@ jobs:
             { os: "ubuntu", target: "i686-unknown-linux-gnu" },
             { os: "ubuntu", target: "wasm32-unknown-unknown" },
             { os: "macos", target: "aarch64-apple-darwin" },
-            { os: "ubuntu", target: "wasm32-wasi" },
           ]
     env:
       TZ: "/usr/share/zoneinfo/your/location"
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
       - name: Cache .cargo and target
-        uses: actions/cache@v2
+        uses: actions/cache@v4
         with:
           path: |
             ~/.cargo
@@ -36,16 +35,13 @@ jobs:
       - name: Install Rust toolchain
         uses: actions-rs/toolchain@v1
         with:
-          toolchain: 1.81 # 1.82 seems to break wasm32 tests https://github.com/rustwasm/wasm-bindgen/issues/4274
+          toolchain: stable
           target: ${{ matrix.platform.target }}
           profile: minimal
           default: true
       - name: Install test runner for wasm
         if: matrix.platform.target == 'wasm32-unknown-unknown'
         run: curl https://rustwasm.github.io/wasm-pack/installer/init.sh -sSf | sh
-      - name: Install test runner for wasi
-        if: matrix.platform.target == 'wasm32-wasi'
-        run: curl https://wasmtime.dev/install.sh -sSf | bash
       - name: Stable Build with all features
         uses: actions-rs/cargo@v1
         with:
@@ -65,12 +61,6 @@ jobs:
       - name: Tests in WASM
         if: matrix.platform.target == 'wasm32-unknown-unknown'
         run: wasm-pack test --node -- --all-features
-      - name: Tests in WASI
-        if: matrix.platform.target == 'wasm32-wasi'
-        run: |
-          export WASMTIME_HOME="$HOME/.wasmtime"
-          export PATH="$WASMTIME_HOME/bin:$PATH"
-          cargo install cargo-wasi && cargo wasi test
   
   check_features:
     runs-on: "${{ matrix.platform.os }}-latest"
@@ -81,9 +71,9 @@ jobs:
     env:
       TZ: "/usr/share/zoneinfo/your/location"
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
       - name: Cache .cargo and target
-        uses: actions/cache@v2
+        uses: actions/cache@v4
         with:
           path: |
             ~/.cargo

.github/workflows/coverage.yml

@@ -12,9 +12,9 @@ jobs:
     env:
       TZ: "/usr/share/zoneinfo/your/location"
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v4
       - name: Cache .cargo
-        uses: actions/cache@v2
+        uses: actions/cache@v4
         with:
           path: |
             ~/.cargo

.github/workflows/lint.yml

@@ -14,7 +14,7 @@ jobs:
     steps:
       - uses: actions/checkout@v2
       - name: Cache .cargo and target
-        uses: actions/cache@v2
+        uses: actions/cache@v4
         with:
           path: |
             ~/.cargo

Cargo.toml

@@ -2,7 +2,7 @@
 name = "smartcore"
 description = "Machine Learning in Rust."
 homepage = "https://smartcorelib.org"
-version = "0.4.0"
+version = "0.4.2"
 authors = ["smartcore Developers"]
 edition = "2021"
 license = "Apache-2.0"

README.md

@@ -18,4 +18,4 @@
 -----
 [![CI](https://github.com/smartcorelib/smartcore/actions/workflows/ci.yml/badge.svg)](https://github.com/smartcorelib/smartcore/actions/workflows/ci.yml)
 
-To start getting familiar with the new smartcore v0.3 API, there is now available a [**Jupyter Notebook environment repository**](https://github.com/smartcorelib/smartcore-jupyter). Please see instructions there, contributions welcome see [CONTRIBUTING](.github/CONTRIBUTING.md).
+To start getting familiar with the new smartcore v0.4 API, there is now available a [**Jupyter Notebook environment repository**](https://github.com/smartcorelib/smartcore-jupyter). Please see instructions there, contributions welcome see [CONTRIBUTING](.github/CONTRIBUTING.md).

src/algorithm/neighbour/fastpair.rs

@@ -173,6 +173,21 @@ impl<'a, T: RealNumber + FloatNumber, M: Array2<T>> FastPair<'a, T, M> {
         }
     }
 
+    ///
+    /// Return order dissimilarities from closest to furthest
+    ///
+    #[allow(dead_code)]
+    pub fn ordered_pairs(&self) -> std::vec::IntoIter<&PairwiseDistance<T>> {
+        // improvement: implement this to return `impl Iterator<Item = &PairwiseDistance<T>>`
+        // need to implement trait `Iterator` for `Vec<&PairwiseDistance<T>>`
+        let mut distances = self
+            .distances
+            .values()
+            .collect::<Vec<&PairwiseDistance<T>>>();
+        distances.sort_by(|a, b| a.partial_cmp(b).unwrap());
+        distances.into_iter()
+    }
+
     //
     // Compute distances from input to all other points in data-structure.
     // input is the row index of the sample matrix
@@ -588,4 +603,103 @@ mod tests_fastpair {
 
         assert_eq!(closest, min_dissimilarity);
     }
+
+    #[test]
+    fn fastpair_ordered_pairs() {
+        let x = DenseMatrix::<f64>::from_2d_array(&[
+            &[5.1, 3.5, 1.4, 0.2],
+            &[4.9, 3.0, 1.4, 0.2],
+            &[4.7, 3.2, 1.3, 0.2],
+            &[4.6, 3.1, 1.5, 0.2],
+            &[5.0, 3.6, 1.4, 0.2],
+            &[5.4, 3.9, 1.7, 0.4],
+            &[4.9, 3.1, 1.5, 0.1],
+            &[7.0, 3.2, 4.7, 1.4],
+            &[6.4, 3.2, 4.5, 1.5],
+            &[6.9, 3.1, 4.9, 1.5],
+            &[5.5, 2.3, 4.0, 1.3],
+            &[6.5, 2.8, 4.6, 1.5],
+            &[4.6, 3.4, 1.4, 0.3],
+            &[5.0, 3.4, 1.5, 0.2],
+            &[4.4, 2.9, 1.4, 0.2],
+        ])
+        .unwrap();
+        let fastpair = FastPair::new(&x).unwrap();
+
+        let ordered = fastpair.ordered_pairs();
+
+        let mut previous: f64 = -1.0;
+        for p in ordered {
+            if previous == -1.0 {
+                previous = p.distance.unwrap();
+            } else {
+                let current = p.distance.unwrap();
+                assert!(current >= previous);
+                previous = current;
+            }
+        }
+    }
+
+    #[test]
+    fn test_empty_set() {
+        let empty_matrix = DenseMatrix::<f64>::zeros(0, 0);
+        let result = FastPair::new(&empty_matrix);
+        assert!(result.is_err());
+        if let Err(e) = result {
+            assert_eq!(
+                e,
+                Failed::because(FailedError::FindFailed, "min number of rows should be 3")
+            );
+        }
+    }
+
+    #[test]
+    fn test_single_point() {
+        let single_point = DenseMatrix::from_2d_array(&[&[1.0, 2.0, 3.0]]).unwrap();
+        let result = FastPair::new(&single_point);
+        assert!(result.is_err());
+        if let Err(e) = result {
+            assert_eq!(
+                e,
+                Failed::because(FailedError::FindFailed, "min number of rows should be 3")
+            );
+        }
+    }
+
+    #[test]
+    fn test_two_points() {
+        let two_points = DenseMatrix::from_2d_array(&[&[1.0, 2.0], &[3.0, 4.0]]).unwrap();
+        let result = FastPair::new(&two_points);
+        assert!(result.is_err());
+        if let Err(e) = result {
+            assert_eq!(
+                e,
+                Failed::because(FailedError::FindFailed, "min number of rows should be 3")
+            );
+        }
+    }
+
+    #[test]
+    fn test_three_identical_points() {
+        let identical_points =
+            DenseMatrix::from_2d_array(&[&[1.0, 1.0], &[1.0, 1.0], &[1.0, 1.0]]).unwrap();
+        let result = FastPair::new(&identical_points);
+        assert!(result.is_ok());
+        let fastpair = result.unwrap();
+        let closest_pair = fastpair.closest_pair();
+        assert_eq!(closest_pair.distance, Some(0.0));
+    }
+
+    #[test]
+    fn test_result_unwrapping() {
+        let valid_matrix =
+            DenseMatrix::from_2d_array(&[&[1.0, 2.0], &[3.0, 4.0], &[5.0, 6.0], &[7.0, 8.0]])
+                .unwrap();
+
+        let result = FastPair::new(&valid_matrix);
+        assert!(result.is_ok());
+
+        // This should not panic
+        let _fastpair = result.unwrap();
+    }
 }

src/cluster/agglomerative.rs

@@ -0,0 +1,315 @@
+//! # Agglomerative Hierarchical Clustering
+//!
+//! Agglomerative clustering is a "bottom-up" hierarchical clustering method. It works by placing each data point in its own cluster and then successively merging the two most similar clusters until a stopping criterion is met. This process creates a tree-based hierarchy of clusters known as a dendrogram.
+//!
+//! The similarity of two clusters is determined by a **linkage criterion**. This implementation uses **single-linkage**, where the distance between two clusters is defined as the minimum distance between any single point in the first cluster and any single point in the second cluster. The distance between points is the standard Euclidean distance.
+//!
+//! The algorithm first builds the full hierarchy of `N-1` merges. To obtain a specific number of clusters, `n_clusters`, the algorithm then effectively "cuts" the dendrogram at the point where `n_clusters` remain.
+//!
+//! ## Example:
+//!
+//! ```
+//! use smartcore::linalg::basic::matrix::DenseMatrix;
+//! use smartcore::cluster::agglomerative::{AgglomerativeClustering, AgglomerativeClusteringParameters};
+//!
+//! // A dataset with 2 distinct groups of points.
+//! let x = DenseMatrix::from_2d_array(&[
+//!         &[0.0, 0.0], &[1.0, 1.0], &[0.5, 0.5], // Cluster A
+//!         &[10.0, 10.0], &[11.0, 11.0], &[10.5, 10.5], // Cluster B
+//!     ]).unwrap();
+//!
+//! // Set parameters to find 2 clusters.
+//! let parameters = AgglomerativeClusteringParameters::default().with_n_clusters(2);
+//!
+//! // Fit the model to the data.
+//! let clustering = AgglomerativeClustering::<f64, usize, DenseMatrix<f64>, Vec<usize>>::fit(&x, parameters).unwrap();
+//!
+//! // Get the cluster assignments.
+//! let labels = clustering.labels; // e.g., [0, 0, 0, 1, 1, 1]
+//! ```
+//!
+//! ## References:
+//!
+//! * ["An Introduction to Statistical Learning", James G., Witten D., Hastie T., Tibshirani R., 10.3.2 Hierarchical Clustering](http://faculty.marshall.usc.edu/gareth-james/ISL/)
+//! * ["The Elements of Statistical Learning", Hastie T., Tibshirani R., Friedman J., 14.3.12 Hierarchical Clustering](https://hastie.su.domains/ElemStatLearn/)
+
+use std::collections::HashMap;
+use std::marker::PhantomData;
+
+use crate::api::UnsupervisedEstimator;
+use crate::error::{Failed, FailedError};
+use crate::linalg::basic::arrays::{Array1, Array2};
+use crate::numbers::basenum::Number;
+
+/// Parameters for the Agglomerative Clustering algorithm.
+#[derive(Debug, Clone, Copy)]
+pub struct AgglomerativeClusteringParameters {
+    /// The number of clusters to find.
+    pub n_clusters: usize,
+}
+
+impl AgglomerativeClusteringParameters {
+    /// Sets the number of clusters.
+    ///
+    /// # Arguments
+    /// * `n_clusters` - The desired number of clusters.
+    pub fn with_n_clusters(mut self, n_clusters: usize) -> Self {
+        self.n_clusters = n_clusters;
+        self
+    }
+}
+
+impl Default for AgglomerativeClusteringParameters {
+    fn default() -> Self {
+        AgglomerativeClusteringParameters { n_clusters: 2 }
+    }
+}
+
+/// Agglomerative Clustering model.
+///
+/// This implementation uses single-linkage clustering, which is mathematically
+/// equivalent to finding the Minimum Spanning Tree (MST) of the data points.
+/// The core logic is an efficient implementation of Kruskal's algorithm, which
+/// processes all pairwise distances in increasing order and uses a Disjoint
+/// Set Union (DSU) data structure to track cluster membership.
+#[derive(Debug)]
+pub struct AgglomerativeClustering<TX: Number, TY: Number, X: Array2<TX>, Y: Array1<TY>> {
+    /// The cluster label assigned to each sample.
+    pub labels: Vec<usize>,
+    _phantom_tx: PhantomData<TX>,
+    _phantom_ty: PhantomData<TY>,
+    _phantom_x: PhantomData<X>,
+    _phantom_y: PhantomData<Y>,
+}
+
+impl<TX: Number, TY: Number, X: Array2<TX>, Y: Array1<TY>> AgglomerativeClustering<TX, TY, X, Y> {
+    /// Fits the agglomerative clustering model to the data.
+    ///
+    /// # Arguments
+    /// * `data` - A reference to the input data matrix.
+    /// * `parameters` - The parameters for the clustering algorithm, including `n_clusters`.
+    ///
+    /// # Returns
+    /// A `Result` containing the fitted model with cluster labels, or an error if
+    pub fn fit(data: &X, parameters: AgglomerativeClusteringParameters) -> Result<Self, Failed> {
+        let (num_samples, _) = data.shape();
+        let n_clusters = parameters.n_clusters;
+        if n_clusters > num_samples {
+            return Err(Failed::because(
+                FailedError::ParametersError,
+                &format!("n_clusters: {n_clusters} cannot be greater than n_samples: {num_samples}"),
+            ));
+        }
+
+        let mut distance_pairs = Vec::new();
+        for i in 0..num_samples {
+            for j in (i + 1)..num_samples {
+                let distance: f64 = data
+                    .get_row(i)
+                    .iterator(0)
+                    .zip(data.get_row(j).iterator(0))
+                    .map(|(&a, &b)| (a.to_f64().unwrap() - b.to_f64().unwrap()).powi(2))
+                    .sum::<f64>();
+
+                distance_pairs.push((distance, i, j));
+            }
+        }
+        distance_pairs.sort_unstable_by(|a, b| b.0.partial_cmp(&a.0).unwrap());
+        let mut parent = HashMap::new();
+        let mut children = HashMap::new();
+        for i in 0..num_samples {
+            parent.insert(i, i);
+            children.insert(i, vec![i]);
+        }
+
+        let mut merge_history = Vec::new();
+        let num_merges_needed = num_samples - 1;
+
+        while merge_history.len() < num_merges_needed {
+            let (_, p1, p2) = distance_pairs.pop().unwrap();
+
+            let root1 = parent[&p1];
+            let root2 = parent[&p2];
+
+            if root1 != root2 {
+                let root2_children = children.remove(&root2).unwrap();
+                for child in root2_children.iter() {
+                    parent.insert(*child, root1);
+                }
+                let root1_children = children.get_mut(&root1).unwrap();
+                root1_children.extend(root2_children);
+                merge_history.push((root1, root2));
+            }
+        }
+
+        let mut clusters = HashMap::new();
+        let mut assignments = HashMap::new();
+
+        for i in 0..num_samples {
+            clusters.insert(i, vec![i]);
+            assignments.insert(i, i);
+        }
+
+        let merges_to_apply = num_samples - n_clusters;
+
+        for (root1, root2) in merge_history[0..merges_to_apply].iter() {
+            let root1_cluster = assignments[root1];
+            let root2_cluster = assignments[root2];
+
+            let root2_assignments = clusters.remove(&root2_cluster).unwrap();
+            for assignment in root2_assignments.iter() {
+                assignments.insert(*assignment, root1_cluster);
+            }
+            let root1_assignments = clusters.get_mut(&root1_cluster).unwrap();
+            root1_assignments.extend(root2_assignments);
+        }
+
+        let mut labels: Vec<usize> = (0..num_samples).map(|_| 0).collect();
+        let mut cluster_keys: Vec<&usize> = clusters.keys().collect();
+        cluster_keys.sort();
+        for (i, key) in cluster_keys.into_iter().enumerate() {
+            for index in clusters[key].iter() {
+                labels[*index] = i;
+            }
+        }
+        Ok(AgglomerativeClustering {
+            labels,
+            _phantom_tx: PhantomData,
+            _phantom_ty: PhantomData,
+            _phantom_x: PhantomData,
+            _phantom_y: PhantomData,
+        })
+    }
+}
+
+impl<TX: Number, TY: Number, X: Array2<TX>, Y: Array1<TY>>
+    UnsupervisedEstimator<X, AgglomerativeClusteringParameters>
+    for AgglomerativeClustering<TX, TY, X, Y>
+{
+    fn fit(x: &X, parameters: AgglomerativeClusteringParameters) -> Result<Self, Failed> {
+        AgglomerativeClustering::fit(x, parameters)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use crate::linalg::basic::matrix::DenseMatrix;
+    use std::collections::HashSet;
+
+    use super::*;
+
+    #[test]
+    fn test_simple_clustering() {
+        // Two distinct clusters, far apart.
+        let data = vec![
+            0.0, 0.0, 1.0, 1.0, 0.5, 0.5, // Cluster A
+            10.0, 10.0, 11.0, 11.0, 10.5, 10.5, // Cluster B
+        ];
+        let matrix = DenseMatrix::new(6, 2, data, false).unwrap();
+        let parameters = AgglomerativeClusteringParameters::default().with_n_clusters(2);
+        // Using f64 for TY as usize doesn't satisfy the Number trait bound.
+        let clustering = AgglomerativeClustering::<f64, f64, DenseMatrix<f64>, Vec<f64>>::fit(
+            &matrix, parameters,
+        )
+        .unwrap();
+
+        let labels = clustering.labels;
+
+        // Check that all points in the first group have the same label.
+        let first_group_label = labels[0];
+        assert!(labels[0..3].iter().all(|&l| l == first_group_label));
+
+        // Check that all points in the second group have the same label.
+        let second_group_label = labels[3];
+        assert!(labels[3..6].iter().all(|&l| l == second_group_label));
+
+        // Check that the two groups have different labels.
+        assert_ne!(first_group_label, second_group_label);
+    }
+
+    #[test]
+    fn test_four_clusters() {
+        // Four distinct clusters in the corners of a square.
+        let data = vec![
+            0.0, 0.0, 1.0, 1.0, // Cluster A
+            100.0, 100.0, 101.0, 101.0, // Cluster B
+            0.0, 100.0, 1.0, 101.0, // Cluster C
+            100.0, 0.0, 101.0, 1.0, // Cluster D
+        ];
+        let matrix = DenseMatrix::new(8, 2, data, false).unwrap();
+        let parameters = AgglomerativeClusteringParameters::default().with_n_clusters(4);
+        let clustering = AgglomerativeClustering::<f64, f64, DenseMatrix<f64>, Vec<f64>>::fit(
+            &matrix, parameters,
+        )
+        .unwrap();
+
+        let labels = clustering.labels;
+
+        // Verify that there are exactly 4 unique labels produced.
+        let unique_labels: HashSet<usize> = labels.iter().cloned().collect();
+        assert_eq!(unique_labels.len(), 4);
+
+        // Verify that points within each original group were assigned the same cluster label.
+        let label_a = labels[0];
+        assert_eq!(label_a, labels[1]);
+
+        let label_b = labels[2];
+        assert_eq!(label_b, labels[3]);
+
+        let label_c = labels[4];
+        assert_eq!(label_c, labels[5]);
+
+        let label_d = labels[6];
+        assert_eq!(label_d, labels[7]);
+
+        // Verify that all four groups received different labels.
+        assert_ne!(label_a, label_b);
+        assert_ne!(label_a, label_c);
+        assert_ne!(label_a, label_d);
+        assert_ne!(label_b, label_c);
+        assert_ne!(label_b, label_d);
+        assert_ne!(label_c, label_d);
+    }
+
+    #[test]
+    fn test_n_clusters_equal_to_samples() {
+        let data = vec![0.0, 0.0, 5.0, 5.0, 10.0, 10.0];
+        let matrix = DenseMatrix::new(3, 2, data, false).unwrap();
+        let parameters = AgglomerativeClusteringParameters::default().with_n_clusters(3);
+        let clustering = AgglomerativeClustering::<f64, f64, DenseMatrix<f64>, Vec<f64>>::fit(
+            &matrix, parameters,
+        )
+        .unwrap();
+
+        // Each point should be its own cluster. Sorting makes the test deterministic.
+        let mut labels = clustering.labels;
+        labels.sort();
+        assert_eq!(labels, vec![0, 1, 2]);
+    }
+
+    #[test]
+    fn test_one_cluster() {
+        let data = vec![0.0, 0.0, 5.0, 5.0, 10.0, 10.0];
+        let matrix = DenseMatrix::new(3, 2, data, false).unwrap();
+        let parameters = AgglomerativeClusteringParameters::default().with_n_clusters(1);
+        let clustering = AgglomerativeClustering::<f64, f64, DenseMatrix<f64>, Vec<f64>>::fit(
+            &matrix, parameters,
+        )
+        .unwrap();
+
+        // All points should be in the same cluster.
+        assert_eq!(clustering.labels, vec![0, 0, 0]);
+    }
+
+    #[test]
+    fn test_error_on_too_many_clusters() {
+        let data = vec![0.0, 0.0, 5.0, 5.0];
+        let matrix = DenseMatrix::new(2, 2, data, false).unwrap();
+        let parameters = AgglomerativeClusteringParameters::default().with_n_clusters(3);
+        let result = AgglomerativeClustering::<f64, f64, DenseMatrix<f64>, Vec<f64>>::fit(
+            &matrix, parameters,
+        );
+
+        assert!(result.is_err());
+    }
+}

src/cluster/mod.rs

@@ -3,6 +3,7 @@
 //! Clustering is the type of unsupervised learning where you divide the population or data points into a number of groups such that data points in the same groups
 //! are more similar to other data points in the same group than those in other groups. In simple words, the aim is to segregate groups with similar traits and assign them into clusters.
 
+pub mod agglomerative;
 pub mod dbscan;
 /// An iterative clustering algorithm that aims to find local maxima in each iteration.
 pub mod kmeans;

src/linalg/basic/arrays.rs

@@ -619,7 +619,7 @@ pub trait MutArrayView1<T: Debug + Display + Copy + Sized>:
         T: Number + PartialOrd,
     {
         let stack_size = 64;
-        let mut jstack = -1;
+        let mut jstack: i32 = -1;
         let mut l = 0;
         let mut istack = vec![0; stack_size];
         let mut ir = self.shape() - 1;
@@ -2190,4 +2190,29 @@ mod tests {
 
         assert_eq!(result, [65, 581, 30])
     }
+
+    #[test]
+    fn test_argsort_mut_exact_boundary() {
+        // Test index == length - 1 case
+        let boundary =
+            DenseMatrix::from_2d_array(&[&[1.0, 2.0, 3.0, f64::MAX], &[3.0, f64::MAX, 0.0, 2.0]])
+                .unwrap();
+        let mut view0: Vec<f64> = boundary.get_col(0).iterator(0).copied().collect();
+        let indices = view0.argsort_mut();
+        assert_eq!(indices.last(), Some(&1));
+        assert_eq!(indices.first(), Some(&0));
+
+        let mut view1: Vec<f64> = boundary.get_col(3).iterator(0).copied().collect();
+        let indices = view1.argsort_mut();
+        assert_eq!(indices.last(), Some(&0));
+        assert_eq!(indices.first(), Some(&1));
+    }
+
+    #[test]
+    fn test_argsort_mut_filled_array() {
+        let matrix = DenseMatrix::<f64>::rand(1000, 1000);
+        let mut view: Vec<f64> = matrix.get_col(0).iterator(0).copied().collect();
+        let sorted = view.argsort_mut();
+        assert_eq!(sorted.len(), 1000);
+    }
 }

src/linalg/basic/matrix.rs

@@ -663,6 +663,7 @@ mod tests {
     #[test]
     fn test_instantiate_err_view3() {
         let x = DenseMatrix::from_2d_array(&[&[1., 2., 3.], &[4., 5., 6.], &[7., 8., 9.]]).unwrap();
+        #[allow(clippy::reversed_empty_ranges)]
         let v = DenseMatrixView::new(&x, 0..3, 4..3);
         assert!(v.is_err());
     }

src/naive_bayes/bernoulli.rs

@@ -257,8 +257,7 @@ impl<TY: Number + Ord + Unsigned> BernoulliNBDistribution<TY> {
     /// Fits the distribution to a NxM matrix where N is number of samples and M is number of features.
     /// * `x` - training data.
     /// * `y` - vector with target values (classes) of length N.
-    /// * `priors` - Optional vector with prior probabilities of the classes. If not defined,
-    ///     priors are adjusted according to the data.
+    /// * `priors` - Optional vector with prior probabilities of the classes. If not defined, priors are adjusted according to the data.
     /// * `alpha` - Additive (Laplace/Lidstone) smoothing parameter.
     /// * `binarize` - Threshold for binarizing.
     fn fit<TX: Number + PartialOrd, X: Array2<TX>, Y: Array1<TY>>(

src/naive_bayes/gaussian.rs

@@ -174,8 +174,7 @@ impl<TY: Number + Ord + Unsigned> GaussianNBDistribution<TY> {
     /// Fits the distribution to a NxM matrix where N is number of samples and M is number of features.
     /// * `x` - training data.
     /// * `y` - vector with target values (classes) of length N.
-    /// * `priors` - Optional vector with prior probabilities of the classes. If not defined,
-    ///     priors are adjusted according to the data.
+    /// * `priors` - Optional vector with prior probabilities of the classes. If not defined, priors are adjusted according to the data.
     pub fn fit<TX: Number + RealNumber, X: Array2<TX>, Y: Array1<TY>>(
         x: &X,
         y: &Y,

src/naive_bayes/multinomial.rs

@@ -207,8 +207,7 @@ impl<TY: Number + Ord + Unsigned> MultinomialNBDistribution<TY> {
     /// Fits the distribution to a NxM matrix where N is number of samples and M is number of features.
     /// * `x` - training data.
     /// * `y` - vector with target values (classes) of length N.
-    /// * `priors` - Optional vector with prior probabilities of the classes. If not defined,
-    ///     priors are adjusted according to the data.
+    /// * `priors` - Optional vector with prior probabilities of the classes. If not defined, priors are adjusted according to the data.
     /// * `alpha` - Additive (Laplace/Lidstone) smoothing parameter.
     pub fn fit<TX: Number + Unsigned, X: Array2<TX>, Y: Array1<TY>>(
         x: &X,

src/neighbors/mod.rs

@@ -64,7 +64,7 @@ impl KNNWeightFunction {
             KNNWeightFunction::Distance => {
                 // if there are any points that has zero distance from one or more training points,
                 // those training points are weighted as 1.0 and the other points as 0.0
-                if distances.iter().any(|&e| e == 0f64) {
+                if distances.contains(&0f64) {
                     distances
                         .iter()
                         .map(|e| if *e == 0f64 { 1f64 } else { 0f64 })

src/preprocessing/categorical.rs

@@ -24,7 +24,7 @@
 //! //    &[1.5, 1.0, 0.0, 1.5, 0.0, 0.0, 1.0, 0.0]
 //! //    &[1.5, 0.0, 1.0, 1.5, 0.0, 0.0, 0.0, 1.0]
 //! ```
-use std::iter;
+use std::iter::repeat_n;
 
 use crate::error::Failed;
 use crate::linalg::basic::arrays::Array2;
@@ -75,11 +75,7 @@ fn find_new_idxs(num_params: usize, cat_sizes: &[usize], cat_idxs: &[usize]) ->
     let offset = (0..1).chain(offset_);
 
     let new_param_idxs: Vec<usize> = (0..num_params)
-        .zip(
-            repeats
-                .zip(offset)
-                .flat_map(|(r, o)| iter::repeat(o).take(r)),
-        )
+        .zip(repeats.zip(offset).flat_map(|(r, o)| repeat_n(o, r)))
         .map(|(idx, ofst)| idx + ofst)
         .collect();
     new_param_idxs
@@ -124,7 +120,7 @@ impl OneHotEncoder {
                 let (nrows, _) = data.shape();
 
                 // col buffer to avoid allocations
-                let mut col_buf: Vec<T> = iter::repeat(T::zero()).take(nrows).collect();
+                let mut col_buf: Vec<T> = repeat_n(T::zero(), nrows).collect();
 
                 let mut res: Vec<CategoryMapper<CategoricalFloat>> = Vec::with_capacity(idxs.len());
 

src/svm/mod.rs

@@ -25,14 +25,18 @@
 /// search parameters
 pub mod svc;
 pub mod svr;
-// /// search parameters space
-// pub mod search;
+// search parameters space
+pub mod search;
 
 use core::fmt::Debug;
 
 #[cfg(feature = "serde")]
 use serde::{Deserialize, Serialize};
 
+// Only import typetag if not compiling for wasm32 and serde is enabled
+#[cfg(all(feature = "serde", not(target_arch = "wasm32")))]
+use typetag;
+
 use crate::error::{Failed, FailedError};
 use crate::linalg::basic::arrays::{Array1, ArrayView1};
 
@@ -48,197 +52,281 @@ pub trait Kernel: Debug {
     fn apply(&self, x_i: &Vec<f64>, x_j: &Vec<f64>) -> Result<f64, Failed>;
 }
 
-/// Pre-defined kernel functions
+/// A enumerator for all the kernels type to support.
+/// This allows kernel selection and parameterization ergonomic, type-safe, and ready for use in parameter structs like SVRParameters.
+/// You can construct kernels using the provided variants and builder-style methods.
+///
+/// # Examples
+///
+/// ```
+/// use smartcore::svm::Kernels;
+///
+/// let linear = Kernels::linear();
+/// let rbf = Kernels::rbf().with_gamma(0.5);
+/// let poly = Kernels::polynomial().with_degree(3.0).with_gamma(0.5).with_coef0(1.0);
+/// let sigmoid = Kernels::sigmoid().with_gamma(0.2).with_coef0(0.0);
+/// ```
 #[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
-#[derive(Debug, Clone)]
-pub struct Kernels;
+#[derive(Debug, Clone, PartialEq)]
+pub enum Kernels {
+    /// Linear kernel (default).
+    ///
+    /// Computes the standard dot product between vectors.
+    Linear,
+
+    /// Radial Basis Function (RBF) kernel.
+    ///
+    /// Formula: K(x, y) = exp(-gamma * ||x-y||²)
+    RBF {
+        /// Controls the width of the Gaussian RBF kernel.
+        ///
+        /// Larger values of gamma lead to higher bias and lower variance.
+        /// This parameter is inversely proportional to the radius of influence
+        /// of samples selected by the model as support vectors.
+        gamma: Option<f64>,
+    },
+
+    /// Polynomial kernel.
+    ///
+    /// Formula: K(x, y) = (gamma * <x, y> + coef0)^degree
+    Polynomial {
+        /// The degree of the polynomial kernel.
+        ///
+        /// Integer values are typical (2 = quadratic, 3 = cubic), but any positive real value is valid.
+        /// Higher degree values create decision boundaries with higher complexity.
+        degree: Option<f64>,
+
+        /// Kernel coefficient for the dot product.
+        ///
+        /// Controls the influence of higher-degree versus lower-degree terms in the polynomial.
+        /// If None, a default value will be used.
+        gamma: Option<f64>,
+
+        /// Independent term in the polynomial kernel.
+        ///
+        /// Controls the influence of higher-degree versus lower-degree terms.
+        /// If None, a default value of 1.0 will be used.
+        coef0: Option<f64>,
+    },
+
+    /// Sigmoid kernel.
+    ///
+    /// Formula: K(x, y) = tanh(gamma * <x, y> + coef0)
+    Sigmoid {
+        /// Kernel coefficient for the dot product.
+        ///
+        /// Controls the scaling of the dot product in the sigmoid function.
+        /// If None, a default value will be used.
+        gamma: Option<f64>,
+
+        /// Independent term in the sigmoid kernel.
+        ///
+        /// Acts as a threshold/bias term in the sigmoid function.
+        /// If None, a default value of 1.0 will be used.
+        coef0: Option<f64>,
+    },
+}
 
 impl Kernels {
-    /// Return a default linear
-    pub fn linear() -> LinearKernel {
-        LinearKernel
+    /// Create a linear kernel.
+    ///
+    /// The linear kernel computes the dot product between two vectors:
+    /// K(x, y) = <x, y>
+    pub fn linear() -> Self {
+        Kernels::Linear
     }
-    /// Return a default RBF
-    pub fn rbf() -> RBFKernel {
-        RBFKernel::default()
+
+    /// Create an RBF kernel with unspecified gamma.
+    ///
+    /// The RBF kernel is defined as:
+    /// K(x, y) = exp(-gamma * ||x-y||²)
+    ///
+    /// You should specify gamma using `with_gamma()` before using this kernel.
+    pub fn rbf() -> Self {
+        Kernels::RBF { gamma: None }
     }
-    /// Return a default polynomial
-    pub fn polynomial() -> PolynomialKernel {
-        PolynomialKernel::default()
+
+    /// Create a polynomial kernel with default parameters.
+    ///
+    /// The polynomial kernel is defined as:
+    /// K(x, y) = (gamma * <x, y> + coef0)^degree
+    ///
+    /// Default values:
+    /// - gamma: None (must be specified)
+    /// - degree: None (must be specified)
+    /// - coef0: 1.0
+    pub fn polynomial() -> Self {
+        Kernels::Polynomial {
+            gamma: None,
+            degree: None,
+            coef0: Some(1.0),
+        }
     }
-    /// Return a default sigmoid
-    pub fn sigmoid() -> SigmoidKernel {
-        SigmoidKernel::default()
+
+    /// Create a sigmoid kernel with default parameters.
+    ///
+    /// The sigmoid kernel is defined as:
+    /// K(x, y) = tanh(gamma * <x, y> + coef0)
+    ///
+    /// Default values:
+    /// - gamma: None (must be specified)
+    /// - coef0: 1.0
+    ///
+    pub fn sigmoid() -> Self {
+        Kernels::Sigmoid {
+            gamma: None,
+            coef0: Some(1.0),
+        }
     }
-}
 
-/// Linear Kernel
-#[allow(clippy::derive_partial_eq_without_eq)]
-#[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
-#[derive(Debug, Clone, PartialEq, Eq, Default)]
-pub struct LinearKernel;
-
-/// Radial basis function (Gaussian) kernel
-#[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
-#[derive(Debug, Default, Clone, PartialEq)]
-pub struct RBFKernel {
-    /// kernel coefficient
-    pub gamma: Option<f64>,
-}
-
-#[allow(dead_code)]
-impl RBFKernel {
-    /// assign gamma parameter to kernel (required)
-    /// ```rust
-    /// use smartcore::svm::RBFKernel;
-    /// let knl = RBFKernel::default().with_gamma(0.7);
-    /// ```
-    pub fn with_gamma(mut self, gamma: f64) -> Self {
-        self.gamma = Some(gamma);
-        self
+    /// Set the `gamma` parameter for RBF, polynomial, or sigmoid kernels.
+    ///
+    /// The gamma parameter has different interpretations depending on the kernel:
+    /// - For RBF: Controls the width of the Gaussian. Larger values mean tighter fit.
+    /// - For Polynomial: Scaling factor for the dot product.
+    /// - For Sigmoid: Scaling factor for the dot product.
+    ///
+    pub fn with_gamma(self, gamma: f64) -> Self {
+        match self {
+            Kernels::RBF { .. } => Kernels::RBF { gamma: Some(gamma) },
+            Kernels::Polynomial { degree, coef0, .. } => Kernels::Polynomial {
+                gamma: Some(gamma),
+                degree,
+                coef0,
+            },
+            Kernels::Sigmoid { coef0, .. } => Kernels::Sigmoid {
+                gamma: Some(gamma),
+                coef0,
+            },
+            other => other,
+        }
     }
-}
 
-/// Polynomial kernel
-#[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
-#[derive(Debug, Clone, PartialEq)]
-pub struct PolynomialKernel {
-    /// degree of the polynomial
-    pub degree: Option<f64>,
-    /// kernel coefficient
-    pub gamma: Option<f64>,
-    /// independent term in kernel function
-    pub coef0: Option<f64>,
-}
+    /// Set the `degree` parameter for the polynomial kernel.
+    ///
+    /// The degree parameter controls the flexibility of the decision boundary.
+    /// Higher degrees create more complex boundaries but may lead to overfitting.
+    ///
+    pub fn with_degree(self, degree: f64) -> Self {
+        match self {
+            Kernels::Polynomial { gamma, coef0, .. } => Kernels::Polynomial {
+                degree: Some(degree),
+                gamma,
+                coef0,
+            },
+            other => other,
+        }
+    }
 
-impl Default for PolynomialKernel {
-    fn default() -> Self {
-        Self {
-            gamma: Option::None,
-            degree: Option::None,
-            coef0: Some(1f64),
+    /// Set the `coef0` parameter for polynomial or sigmoid kernels.
+    ///
+    /// The coef0 parameter is the independent term in the kernel function:
+    /// - For Polynomial: Controls the influence of higher-degree vs. lower-degree terms.
+    /// - For Sigmoid: Acts as a threshold/bias term.
+    ///
+    pub fn with_coef0(self, coef0: f64) -> Self {
+        match self {
+            Kernels::Polynomial { degree, gamma, .. } => Kernels::Polynomial {
+                degree,
+                gamma,
+                coef0: Some(coef0),
+            },
+            Kernels::Sigmoid { gamma, .. } => Kernels::Sigmoid {
+                gamma,
+                coef0: Some(coef0),
+            },
+            other => other,
         }
     }
 }
 
-impl PolynomialKernel {
-    /// set parameters for kernel
-    /// ```rust
-    /// use smartcore::svm::PolynomialKernel;
-    /// let knl = PolynomialKernel::default().with_params(3.0, 0.7, 1.0);
-    /// ```
-    pub fn with_params(mut self, degree: f64, gamma: f64, coef0: f64) -> Self {
-        self.degree = Some(degree);
-        self.gamma = Some(gamma);
-        self.coef0 = Some(coef0);
-        self
-    }
-    /// set gamma parameter for kernel
-    /// ```rust
-    /// use smartcore::svm::PolynomialKernel;
-    /// let knl = PolynomialKernel::default().with_gamma(0.7);
-    /// ```
-    pub fn with_gamma(mut self, gamma: f64) -> Self {
-        self.gamma = Some(gamma);
-        self
-    }
-    /// set degree parameter for kernel
-    /// ```rust
-    /// use smartcore::svm::PolynomialKernel;
-    /// let knl = PolynomialKernel::default().with_degree(3.0, 100);
-    /// ```
-    pub fn with_degree(self, degree: f64, n_features: usize) -> Self {
-        self.with_params(degree, 1f64, 1f64 / n_features as f64)
-    }
-}
-
-/// Sigmoid (hyperbolic tangent) kernel
-#[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
-#[derive(Debug, Clone, PartialEq)]
-pub struct SigmoidKernel {
-    /// kernel coefficient
-    pub gamma: Option<f64>,
-    /// independent term in kernel function
-    pub coef0: Option<f64>,
-}
-
-impl Default for SigmoidKernel {
-    fn default() -> Self {
-        Self {
-            gamma: Option::None,
-            coef0: Some(1f64),
-        }
-    }
-}
-
-impl SigmoidKernel {
-    /// set parameters for kernel
-    /// ```rust
-    /// use smartcore::svm::SigmoidKernel;
-    /// let knl = SigmoidKernel::default().with_params(0.7, 1.0);
-    /// ```
-    pub fn with_params(mut self, gamma: f64, coef0: f64) -> Self {
-        self.gamma = Some(gamma);
-        self.coef0 = Some(coef0);
-        self
-    }
-    /// set gamma parameter for kernel
-    /// ```rust
-    /// use smartcore::svm::SigmoidKernel;
-    /// let knl = SigmoidKernel::default().with_gamma(0.7);
-    /// ```
-    pub fn with_gamma(mut self, gamma: f64) -> Self {
-        self.gamma = Some(gamma);
-        self
-    }
-}
-
+/// Implementation of the [`Kernel`] trait for the [`Kernels`] enum in smartcore.
+///
+/// This method computes the value of the kernel function between two feature vectors `x_i` and `x_j`,
+/// according to the variant and parameters of the [`Kernels`] enum. This enables flexible and type-safe
+/// selection of kernel functions for SVM and SVR models in smartcore.
+///
+/// # Supported Kernels
+///
+/// - [`Kernels::Linear`]: Computes the standard dot product between `x_i` and `x_j`.
+/// - [`Kernels::RBF`]: Computes the Radial Basis Function (Gaussian) kernel. Requires `gamma`.
+/// - [`Kernels::Polynomial`]: Computes the polynomial kernel. Requires `degree`, `gamma`, and `coef0`.
+/// - [`Kernels::Sigmoid`]: Computes the sigmoid kernel. Requires `gamma` and `coef0`.
+///
+/// # Parameters
+///
+/// - `x_i`: First input vector (feature vector).
+/// - `x_j`: Second input vector (feature vector).
+///
+/// # Returns
+///
+/// - `Ok(f64)`: The computed kernel value.
+/// - `Err(Failed)`: If any required kernel parameter is missing.
+///
+/// # Errors
+///
+/// Returns `Err(Failed)` if a required parameter (such as `gamma`, `degree`, or `coef0`)
+/// is `None` for the selected kernel variant.
+///
+/// # Example
+///
+/// ```
+/// use smartcore::svm::Kernels;
+/// use smartcore::svm::Kernel;
+///
+/// let x = vec![1.0, 2.0, 3.0];
+/// let y = vec![4.0, 5.0, 6.0];
+/// let kernel = Kernels::rbf().with_gamma(0.5);
+/// let value = kernel.apply(&x, &y).unwrap();
+/// ```
+///
+/// # Notes
+///
+/// - This implementation follows smartcore's philosophy: pure Rust, no macros, no unsafe code,
+///   and an accessible, pythonic API surface for both ML practitioners and Rust beginners.
+/// - All kernel parameters must be set before calling `apply`; missing parameters will result in an error.
+///
+/// See the [`Kernels`] enum documentation for more details on each kernel type and its parameters.
 #[cfg_attr(all(feature = "serde", not(target_arch = "wasm32")), typetag::serde)]
-impl Kernel for LinearKernel {
+impl Kernel for Kernels {
     fn apply(&self, x_i: &Vec<f64>, x_j: &Vec<f64>) -> Result<f64, Failed> {
-        Ok(x_i.dot(x_j))
-    }
-}
-
-#[cfg_attr(all(feature = "serde", not(target_arch = "wasm32")), typetag::serde)]
-impl Kernel for RBFKernel {
-    fn apply(&self, x_i: &Vec<f64>, x_j: &Vec<f64>) -> Result<f64, Failed> {
-        if self.gamma.is_none() {
-            return Err(Failed::because(
-                FailedError::ParametersError,
-                "gamma should be set, use {Kernel}::default().with_gamma(..)",
-            ));
+        match self {
+            Kernels::Linear => Ok(x_i.dot(x_j)),
+            Kernels::RBF { gamma } => {
+                let gamma = gamma.ok_or_else(|| {
+                    Failed::because(FailedError::ParametersError, "gamma not set")
+                })?;
+                let v_diff = x_i.sub(x_j);
+                Ok((-gamma * v_diff.mul(&v_diff).sum()).exp())
+            }
+            Kernels::Polynomial {
+                degree,
+                gamma,
+                coef0,
+            } => {
+                let degree = degree.ok_or_else(|| {
+                    Failed::because(FailedError::ParametersError, "degree not set")
+                })?;
+                let gamma = gamma.ok_or_else(|| {
+                    Failed::because(FailedError::ParametersError, "gamma not set")
+                })?;
+                let coef0 = coef0.ok_or_else(|| {
+                    Failed::because(FailedError::ParametersError, "coef0 not set")
+                })?;
+                let dot = x_i.dot(x_j);
+                Ok((gamma * dot + coef0).powf(degree))
+            }
+            Kernels::Sigmoid { gamma, coef0 } => {
+                let gamma = gamma.ok_or_else(|| {
+                    Failed::because(FailedError::ParametersError, "gamma not set")
+                })?;
+                let coef0 = coef0.ok_or_else(|| {
+                    Failed::because(FailedError::ParametersError, "coef0 not set")
+                })?;
+                let dot = x_i.dot(x_j);
+                Ok((gamma * dot + coef0).tanh())
+            }
         }
-        let v_diff = x_i.sub(x_j);
-        Ok((-self.gamma.unwrap() * v_diff.mul(&v_diff).sum()).exp())
-    }
-}
-
-#[cfg_attr(all(feature = "serde", not(target_arch = "wasm32")), typetag::serde)]
-impl Kernel for PolynomialKernel {
-    fn apply(&self, x_i: &Vec<f64>, x_j: &Vec<f64>) -> Result<f64, Failed> {
-        if self.gamma.is_none() || self.coef0.is_none() || self.degree.is_none() {
-            return Err(Failed::because(
-                FailedError::ParametersError, "gamma, coef0, degree should be set, 
-                                                        use {Kernel}::default().with_{parameter}(..)")
-            );
-        }
-        let dot = x_i.dot(x_j);
-        Ok((self.gamma.unwrap() * dot + self.coef0.unwrap()).powf(self.degree.unwrap()))
-    }
-}
-
-#[cfg_attr(all(feature = "serde", not(target_arch = "wasm32")), typetag::serde)]
-impl Kernel for SigmoidKernel {
-    fn apply(&self, x_i: &Vec<f64>, x_j: &Vec<f64>) -> Result<f64, Failed> {
-        if self.gamma.is_none() || self.coef0.is_none() {
-            return Err(Failed::because(
-                FailedError::ParametersError, "gamma, coef0, degree should be set, 
-                                                        use {Kernel}::default().with_{parameter}(..)")
-            );
-        }
-        let dot = x_i.dot(x_j);
-        Ok(self.gamma.unwrap() * dot + self.coef0.unwrap().tanh())
     }
 }
 
@@ -247,6 +335,18 @@ mod tests {
     use super::*;
     use crate::svm::Kernels;
 
+    #[test]
+    fn rbf_kernel() {
+        let v1 = vec![1., 2., 3.];
+        let v2 = vec![4., 5., 6.];
+        let result = Kernels::rbf()
+            .with_gamma(0.055)
+            .apply(&v1, &v2)
+            .unwrap()
+            .abs();
+        assert!((0.2265f64 - result) < 1e-4);
+    }
+
     #[cfg_attr(
         all(target_arch = "wasm32", not(target_os = "wasi")),
         wasm_bindgen_test::wasm_bindgen_test
@@ -264,7 +364,7 @@ mod tests {
         wasm_bindgen_test::wasm_bindgen_test
     )]
     #[test]
-    fn rbf_kernel() {
+    fn test_rbf_kernel() {
         let v1 = vec![1., 2., 3.];
         let v2 = vec![4., 5., 6.];
 
@@ -287,7 +387,10 @@ mod tests {
         let v2 = vec![4., 5., 6.];
 
         let result = Kernels::polynomial()
-            .with_params(3.0, 0.5, 1.0)
+            .with_gamma(0.5)
+            .with_degree(3.0)
+            .with_coef0(1.0)
+            //.with_params(3.0, 0.5, 1.0)
             .apply(&v1, &v2)
             .unwrap()
             .abs();
@@ -305,7 +408,8 @@ mod tests {
         let v2 = vec![4., 5., 6.];
 
         let result = Kernels::sigmoid()
-            .with_params(0.01, 0.1)
+            .with_gamma(0.01)
+            .with_coef0(0.1)
             .apply(&v1, &v2)
             .unwrap()
             .abs();

src/svm/search/mod.rs

@@ -1,3 +1,5 @@
+//! SVC and Grid Search
+
 /// SVC search parameters
 pub mod svc_params;
 /// SVC search parameters

src/svm/search/svr_params.rs

@@ -1,112 +1,293 @@
-// /// SVR grid search parameters
-// #[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
-// #[derive(Debug, Clone)]
-// pub struct SVRSearchParameters<T: Number + RealNumber, M: Matrix<T>, K: Kernel<T, M::RowVector>> {
-//     /// Epsilon in the epsilon-SVR model.
-//     pub eps: Vec<T>,
-//     /// Regularization parameter.
-//     pub c: Vec<T>,
-//     /// Tolerance for stopping eps.
-//     pub tol: Vec<T>,
-//     /// The kernel function.
-//     pub kernel: Vec<K>,
-//     /// Unused parameter.
-//     m: PhantomData<M>,
-// }
+//! # SVR Grid Search Parameters
+//!
+//! This module provides utilities for defining and iterating over grid search parameter spaces
+//! for Support Vector Regression (SVR) models in [smartcore](https://github.com/smartcorelib/smartcore).
+//!
+//! The main struct, [`SVRSearchParameters`], allows users to specify multiple values for each
+//! SVR hyperparameter (epsilon, regularization parameter C, tolerance, and kernel function).
+//! The provided iterator yields all possible combinations (the Cartesian product) of these parameters,
+//! enabling exhaustive grid search for hyperparameter tuning.
+//!
+//!
+//! ## Example
+//! ```
+//! use smartcore::svm::Kernels;
+//! use smartcore::svm::search::svr_params::SVRSearchParameters;
+//! use smartcore::linalg::basic::matrix::DenseMatrix;
+//!
+//! let params = SVRSearchParameters::<f64, DenseMatrix<f64>> {
+//!     eps: vec![0.1, 0.2],
+//!     c: vec![1.0, 10.0],
+//!     tol: vec![1e-3],
+//!     kernel: vec![Kernels::linear(), Kernels::rbf().with_gamma(0.5)],
+//!     m: std::marker::PhantomData,
+//! };
+//!
+//! // for param_set in params.into_iter() {
+//!     // Use param_set (of type svr::SVRParameters) to fit and evaluate your SVR model.
+//! // }
+//! ```
+//!
+//!
+//! ## Note
+//! This module is intended for use with smartcore version 0.4 or later. The API is not compatible with older versions[1].
+#[cfg(feature = "serde")]
+use serde::{Deserialize, Serialize};
 
-// /// SVR grid search iterator
-// pub struct SVRSearchParametersIterator<T: Number + RealNumber, M: Matrix<T>, K: Kernel<T, M::RowVector>> {
-//     svr_search_parameters: SVRSearchParameters<T, M, K>,
-//     current_eps: usize,
-//     current_c: usize,
-//     current_tol: usize,
-//     current_kernel: usize,
-// }
+use crate::linalg::basic::arrays::Array2;
+use crate::numbers::basenum::Number;
+use crate::numbers::floatnum::FloatNumber;
+use crate::numbers::realnum::RealNumber;
+use crate::svm::{svr, Kernels};
+use std::marker::PhantomData;
 
-// impl<T: Number + RealNumber, M: Matrix<T>, K: Kernel<T, M::RowVector>> IntoIterator
-//     for SVRSearchParameters<T, M, K>
-// {
-//     type Item = SVRParameters<T, M, K>;
-//     type IntoIter = SVRSearchParametersIterator<T, M, K>;
+/// ## SVR grid search parameters
+/// A struct representing a grid of hyperparameters for SVR grid search in smartcore.
+///
+/// Each field is a vector of possible values for the corresponding SVR hyperparameter.
+/// The [`IntoIterator`] implementation yields every possible combination of these parameters
+/// as an `svr::SVRParameters` struct, suitable for use in model selection routines.
+///
+/// # Type Parameters
+/// - `T`: Numeric type for parameters (e.g., `f64`)
+/// - `M`: Matrix type implementing [`Array2<T>`]
+///
+/// # Fields
+/// - `eps`: Vector of epsilon values for the epsilon-insensitive loss in SVR.
+/// - `c`: Vector of regularization parameters (C) for SVR.
+/// - `tol`: Vector of tolerance values for the stopping criterion.
+/// - `kernel`: Vector of kernel function variants (see [`Kernels`]).
+/// - `m`: Phantom data for the matrix type parameter.
+///
+/// # Example
+/// ```
+/// use smartcore::svm::Kernels;
+/// use smartcore::svm::search::svr_params::SVRSearchParameters;
+/// use smartcore::linalg::basic::matrix::DenseMatrix;
+///
+/// let params = SVRSearchParameters::<f64, DenseMatrix<f64>> {
+///     eps: vec![0.1, 0.2],
+///     c: vec![1.0, 10.0],
+///     tol: vec![1e-3],
+///     kernel: vec![Kernels::linear(), Kernels::rbf().with_gamma(0.5)],
+///     m: std::marker::PhantomData,
+/// };
+/// ```
+#[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
+#[derive(Debug, Clone)]
+pub struct SVRSearchParameters<T: Number + RealNumber, M: Array2<T>> {
+    /// Epsilon in the epsilon-SVR model.
+    pub eps: Vec<T>,
+    /// Regularization parameter.
+    pub c: Vec<T>,
+    /// Tolerance for stopping eps.
+    pub tol: Vec<T>,
+    /// The kernel function.
+    pub kernel: Vec<Kernels>,
+    /// Unused parameter.
+    pub m: PhantomData<M>,
+}
 
-//     fn into_iter(self) -> Self::IntoIter {
-//         SVRSearchParametersIterator {
-//             svr_search_parameters: self,
-//             current_eps: 0,
-//             current_c: 0,
-//             current_tol: 0,
-//             current_kernel: 0,
-//         }
-//     }
-// }
+/// SVR grid search iterator
+pub struct SVRSearchParametersIterator<T: Number + RealNumber, M: Array2<T>> {
+    svr_search_parameters: SVRSearchParameters<T, M>,
+    current_eps: usize,
+    current_c: usize,
+    current_tol: usize,
+    current_kernel: usize,
+}
 
-// impl<T: Number + RealNumber, M: Matrix<T>, K: Kernel<T, M::RowVector>> Iterator
-//     for SVRSearchParametersIterator<T, M, K>
-// {
-//     type Item = SVRParameters<T, M, K>;
+impl<T: Number + FloatNumber + RealNumber, M: Array2<T>> IntoIterator
+    for SVRSearchParameters<T, M>
+{
+    type Item = svr::SVRParameters<T>;
+    type IntoIter = SVRSearchParametersIterator<T, M>;
 
-//     fn next(&mut self) -> Option<Self::Item> {
-//         if self.current_eps == self.svr_search_parameters.eps.len()
-//             && self.current_c == self.svr_search_parameters.c.len()
-//             && self.current_tol == self.svr_search_parameters.tol.len()
-//             && self.current_kernel == self.svr_search_parameters.kernel.len()
-//         {
-//             return None;
-//         }
+    fn into_iter(self) -> Self::IntoIter {
+        SVRSearchParametersIterator {
+            svr_search_parameters: self,
+            current_eps: 0,
+            current_c: 0,
+            current_tol: 0,
+            current_kernel: 0,
+        }
+    }
+}
 
-//         let next = SVRParameters::<T, M, K> {
-//             eps: self.svr_search_parameters.eps[self.current_eps],
-//             c: self.svr_search_parameters.c[self.current_c],
-//             tol: self.svr_search_parameters.tol[self.current_tol],
-//             kernel: self.svr_search_parameters.kernel[self.current_kernel].clone(),
-//             m: PhantomData,
-//         };
+impl<T: Number + FloatNumber + RealNumber, M: Array2<T>> Iterator
+    for SVRSearchParametersIterator<T, M>
+{
+    type Item = svr::SVRParameters<T>;
 
-//         if self.current_eps + 1 < self.svr_search_parameters.eps.len() {
-//             self.current_eps += 1;
-//         } else if self.current_c + 1 < self.svr_search_parameters.c.len() {
-//             self.current_eps = 0;
-//             self.current_c += 1;
-//         } else if self.current_tol + 1 < self.svr_search_parameters.tol.len() {
-//             self.current_eps = 0;
-//             self.current_c = 0;
-//             self.current_tol += 1;
-//         } else if self.current_kernel + 1 < self.svr_search_parameters.kernel.len() {
-//             self.current_eps = 0;
-//             self.current_c = 0;
-//             self.current_tol = 0;
-//             self.current_kernel += 1;
-//         } else {
-//             self.current_eps += 1;
-//             self.current_c += 1;
-//             self.current_tol += 1;
-//             self.current_kernel += 1;
-//         }
+    fn next(&mut self) -> Option<Self::Item> {
+        if self.current_eps == self.svr_search_parameters.eps.len()
+            && self.current_c == self.svr_search_parameters.c.len()
+            && self.current_tol == self.svr_search_parameters.tol.len()
+            && self.current_kernel == self.svr_search_parameters.kernel.len()
+        {
+            return None;
+        }
 
-//         Some(next)
-//     }
-// }
+        let next = svr::SVRParameters::<T> {
+            eps: self.svr_search_parameters.eps[self.current_eps],
+            c: self.svr_search_parameters.c[self.current_c],
+            tol: self.svr_search_parameters.tol[self.current_tol],
+            kernel: Some(self.svr_search_parameters.kernel[self.current_kernel].clone()),
+        };
 
-// impl<T: Number + RealNumber, M: Matrix<T>> Default for SVRSearchParameters<T, M, LinearKernel> {
-//     fn default() -> Self {
-//         let default_params: SVRParameters<T, M, LinearKernel> = SVRParameters::default();
+        if self.current_eps + 1 < self.svr_search_parameters.eps.len() {
+            self.current_eps += 1;
+        } else if self.current_c + 1 < self.svr_search_parameters.c.len() {
+            self.current_eps = 0;
+            self.current_c += 1;
+        } else if self.current_tol + 1 < self.svr_search_parameters.tol.len() {
+            self.current_eps = 0;
+            self.current_c = 0;
+            self.current_tol += 1;
+        } else if self.current_kernel + 1 < self.svr_search_parameters.kernel.len() {
+            self.current_eps = 0;
+            self.current_c = 0;
+            self.current_tol = 0;
+            self.current_kernel += 1;
+        } else {
+            self.current_eps += 1;
+            self.current_c += 1;
+            self.current_tol += 1;
+            self.current_kernel += 1;
+        }
 
-//         SVRSearchParameters {
-//             eps: vec![default_params.eps],
-//             c: vec![default_params.c],
-//             tol: vec![default_params.tol],
-//             kernel: vec![default_params.kernel],
-//             m: PhantomData,
-//         }
-//     }
-// }
+        Some(next)
+    }
+}
 
-// #[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
-// #[derive(Debug)]
-// #[cfg_attr(
-//     feature = "serde",
-//     serde(bound(
-//         serialize = "M::RowVector: Serialize, K: Serialize, T: Serialize",
-//         deserialize = "M::RowVector: Deserialize<'de>, K: Deserialize<'de>, T: Deserialize<'de>",
-//     ))
-// )]
+impl<T: Number + FloatNumber + RealNumber, M: Array2<T>> Default for SVRSearchParameters<T, M> {
+    fn default() -> Self {
+        let default_params: svr::SVRParameters<T> = svr::SVRParameters::default();
+
+        SVRSearchParameters {
+            eps: vec![default_params.eps],
+            c: vec![default_params.c],
+            tol: vec![default_params.tol],
+            kernel: vec![default_params.kernel.unwrap_or_else(Kernels::linear)],
+            m: PhantomData,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::linalg::basic::matrix::DenseMatrix;
+    use crate::svm::Kernels;
+
+    type T = f64;
+    type M = DenseMatrix<T>;
+
+    #[test]
+    fn test_default_parameters() {
+        let params = SVRSearchParameters::<T, M>::default();
+        assert_eq!(params.eps.len(), 1);
+        assert_eq!(params.c.len(), 1);
+        assert_eq!(params.tol.len(), 1);
+        assert_eq!(params.kernel.len(), 1);
+        // Check that the default kernel is linear
+        assert_eq!(params.kernel[0], Kernels::linear());
+    }
+
+    #[test]
+    fn test_single_grid_iteration() {
+        let params = SVRSearchParameters::<T, M> {
+            eps: vec![0.1],
+            c: vec![1.0],
+            tol: vec![1e-3],
+            kernel: vec![Kernels::rbf().with_gamma(0.5)],
+            m: PhantomData,
+        };
+        let mut iter = params.into_iter();
+        let param = iter.next().unwrap();
+        assert_eq!(param.eps, 0.1);
+        assert_eq!(param.c, 1.0);
+        assert_eq!(param.tol, 1e-3);
+        assert_eq!(param.kernel, Some(Kernels::rbf().with_gamma(0.5)));
+        assert!(iter.next().is_none());
+    }
+
+    #[test]
+    fn test_cartesian_grid_iteration() {
+        let params = SVRSearchParameters::<T, M> {
+            eps: vec![0.1, 0.2],
+            c: vec![1.0, 2.0],
+            tol: vec![1e-3],
+            kernel: vec![Kernels::linear(), Kernels::rbf().with_gamma(0.5)],
+            m: PhantomData,
+        };
+        let expected_count =
+            params.eps.len() * params.c.len() * params.tol.len() * params.kernel.len();
+        let results: Vec<_> = params.into_iter().collect();
+        assert_eq!(results.len(), expected_count);
+
+        // Check that all parameter combinations are present
+        let mut seen = vec![];
+        for p in &results {
+            seen.push((p.eps, p.c, p.tol, p.kernel.clone().unwrap()));
+        }
+        for &eps in &[0.1, 0.2] {
+            for &c in &[1.0, 2.0] {
+                for &tol in &[1e-3] {
+                    for kernel in &[Kernels::linear(), Kernels::rbf().with_gamma(0.5)] {
+                        assert!(seen.contains(&(eps, c, tol, kernel.clone())));
+                    }
+                }
+            }
+        }
+    }
+
+    #[test]
+    fn test_empty_grid() {
+        let params = SVRSearchParameters::<T, M> {
+            eps: vec![],
+            c: vec![],
+            tol: vec![],
+            kernel: vec![],
+            m: PhantomData,
+        };
+        let mut iter = params.into_iter();
+        assert!(iter.next().is_none());
+    }
+
+    #[test]
+    fn test_kernel_enum_variants() {
+        let lin = Kernels::linear();
+        let rbf = Kernels::rbf().with_gamma(0.2);
+        let poly = Kernels::polynomial()
+            .with_degree(2.0)
+            .with_gamma(1.0)
+            .with_coef0(0.5);
+        let sig = Kernels::sigmoid().with_gamma(0.3).with_coef0(0.1);
+
+        assert_eq!(lin, Kernels::Linear);
+        match rbf {
+            Kernels::RBF { gamma } => assert_eq!(gamma, Some(0.2)),
+            _ => panic!("Not RBF"),
+        }
+        match poly {
+            Kernels::Polynomial {
+                degree,
+                gamma,
+                coef0,
+            } => {
+                assert_eq!(degree, Some(2.0));
+                assert_eq!(gamma, Some(1.0));
+                assert_eq!(coef0, Some(0.5));
+            }
+            _ => panic!("Not Polynomial"),
+        }
+        match sig {
+            Kernels::Sigmoid { gamma, coef0 } => {
+                assert_eq!(gamma, Some(0.3));
+                assert_eq!(coef0, Some(0.1));
+            }
+            _ => panic!("Not Sigmoid"),
+        }
+    }
+}

src/svm/svc.rs

@@ -58,10 +58,11 @@
 //!            1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1];
 //!
 //! let knl = Kernels::linear();
-//! let params = &SVCParameters::default().with_c(200.0).with_kernel(knl);
-//! let svc = SVC::fit(&x, &y, params).unwrap();
+//! let parameters = &SVCParameters::default().with_c(200.0).with_kernel(knl);
+//! let svc = SVC::fit(&x, &y, parameters).unwrap();
 //!
 //! let y_hat = svc.predict(&x).unwrap();
+//!
 //! ```
 //!
 //! ## References:
@@ -84,12 +85,194 @@ use serde::{Deserialize, Serialize};
 
 use crate::api::{PredictorBorrow, SupervisedEstimatorBorrow};
 use crate::error::{Failed, FailedError};
-use crate::linalg::basic::arrays::{Array1, Array2, MutArray};
+use crate::linalg::basic::arrays::{Array, Array1, Array2, MutArray};
 use crate::numbers::basenum::Number;
 use crate::numbers::realnum::RealNumber;
 use crate::rand_custom::get_rng_impl;
 use crate::svm::Kernel;
 
+#[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
+#[derive(Debug)]
+/// Configuration for a multi-class Support Vector Machine (SVM) classifier.
+/// This struct holds the indices of the data points relevant to a specific binary
+/// classification problem within a multi-class context, and the two classes
+/// being discriminated.
+struct MultiClassConfig<TY: Number + Ord> {
+    /// The indices of the data points from the original dataset that belong to the two `classes`.
+    indices: Vec<usize>,
+    /// A tuple representing the two classes that this configuration is designed to distinguish.
+    classes: (TY, TY),
+}
+
+impl<'a, TX: Number + RealNumber, TY: Number + Ord, X: Array2<TX>, Y: Array1<TY>>
+    SupervisedEstimatorBorrow<'a, X, Y, SVCParameters<TX, TY, X, Y>>
+    for MultiClassSVC<'a, TX, TY, X, Y>
+{
+    /// Creates a new, empty `MultiClassSVC` instance.
+    fn new() -> Self {
+        Self {
+            classifiers: Option::None,
+        }
+    }
+
+    /// Fits the `MultiClassSVC` model to the provided data and parameters.
+    ///
+    /// This method delegates the fitting process to the inherent `MultiClassSVC::fit` method.
+    ///
+    /// # Arguments
+    /// * `x` - A reference to the input features (2D array).
+    /// * `y` - A reference to the target labels (1D array).
+    /// * `parameters` - A reference to the `SVCParameters` controlling the SVM training.
+    ///
+    /// # Returns
+    /// A `Result` indicating success (`Self`) or failure (`Failed`).
+    fn fit(
+        x: &'a X,
+        y: &'a Y,
+        parameters: &'a SVCParameters<TX, TY, X, Y>,
+    ) -> Result<Self, Failed> {
+        MultiClassSVC::fit(x, y, parameters)
+    }
+}
+
+impl<'a, TX: Number + RealNumber, TY: Number + Ord, X: Array2<TX>, Y: Array1<TY>>
+    PredictorBorrow<'a, X, TX> for MultiClassSVC<'a, TX, TY, X, Y>
+{
+    /// Predicts the class labels for new data points.
+    ///
+    /// This method delegates the prediction process to the inherent `MultiClassSVC::predict` method.
+    ///
+    /// # Arguments
+    /// * `x` - A reference to the input features (2D array) for which to make predictions.
+    ///
+    /// # Returns
+    /// A `Result` containing a `Vec` of predicted class labels (`TX`) or a `Failed` error.
+    fn predict(&self, x: &'a X) -> Result<Vec<TX>, Failed> {
+        Ok(self.predict(x).unwrap())
+    }
+}
+
+/// A multi-class Support Vector Machine (SVM) classifier.
+///
+/// This struct implements a multi-class SVM using the "one-vs-one" strategy,
+/// where a separate binary SVC classifier is trained for every pair of classes.
+///
+/// # Type Parameters
+/// * `'a` - Lifetime parameter for borrowed data.
+/// * `TX` - The numeric type of the input features (must implement `Number` and `RealNumber`).
+/// * `TY` - The numeric type of the target labels (must implement `Number` and `Ord`).
+/// * `X` - The type representing the 2D array of input features (e.g., a matrix).
+/// * `Y` - The type representing the 1D array of target labels (e.g., a vector).
+pub struct MultiClassSVC<
+    'a,
+    TX: Number + RealNumber,
+    TY: Number + Ord,
+    X: Array2<TX>,
+    Y: Array1<TY>,
+> {
+    /// An optional vector of binary `SVC` classifiers.
+    classifiers: Option<Vec<SVC<'a, TX, TY, X, Y>>>,
+}
+
+impl<'a, TX: Number + RealNumber, TY: Number + Ord, X: Array2<TX>, Y: Array1<TY>>
+    MultiClassSVC<'a, TX, TY, X, Y>
+{
+    /// Fits the `MultiClassSVC` model to the provided data using a one-vs-one strategy.
+    ///
+    /// This method identifies all unique classes in the target labels `y` and then
+    /// trains a binary `SVC` for every unique pair of classes. For each pair, it
+    /// extracts the relevant data points and their labels, and then trains a
+    /// specialized `SVC` for that binary classification task.
+    ///
+    /// # Arguments
+    /// * `x` - A reference to the input features (2D array).
+    /// * `y` - A reference to the target labels (1D array).
+    /// * `parameters` - A reference to the `SVCParameters` controlling the SVM training for each individual binary classifier.
+    ///  
+    ///
+    /// # Returns
+    /// A `Result` indicating success (`MultiClassSVC`) or failure (`Failed`).
+    pub fn fit(
+        x: &'a X,
+        y: &'a Y,
+        parameters: &'a SVCParameters<TX, TY, X, Y>,
+    ) -> Result<MultiClassSVC<'a, TX, TY, X, Y>, Failed> {
+        let unique_classes = y.unique();
+        let mut classifiers = Vec::new();
+        // Iterate through all unique pairs of classes (one-vs-one strategy)
+        for i in 0..unique_classes.len() {
+            for j in i..unique_classes.len() {
+                if i == j {
+                    continue;
+                }
+                let class0 = unique_classes[j];
+                let class1 = unique_classes[i];
+
+                let mut indices = Vec::new();
+                // Collect indices of data points belonging to the current pair of classes
+                for (index, v) in y.iterator(0).enumerate() {
+                    if *v == class0 || *v == class1 {
+                        indices.push(index)
+                    }
+                }
+                let classes = (class0, class1);
+                let multiclass_config = MultiClassConfig { classes, indices };
+                // Fit a binary SVC for the current pair of classes
+                let svc = SVC::multiclass_fit(x, y, parameters, multiclass_config).unwrap();
+                classifiers.push(svc);
+            }
+        }
+        Ok(Self {
+            classifiers: Some(classifiers),
+        })
+    }
+
+    /// Predicts the class labels for new data points using the trained multi-class SVM.
+    ///
+    /// This method uses a "voting" scheme (majority vote) among all the binary
+    /// classifiers to determine the final prediction for each data point.
+    ///
+    /// # Arguments
+    /// * `x` - A reference to the input features (2D array) for which to make predictions.
+    ///
+    /// # Returns
+    /// A `Result` containing a `Vec` of predicted class labels (`TX`) or a `Failed` error.
+    ///
+    pub fn predict(&self, x: &X) -> Result<Vec<TX>, Failed> {
+        // Initialize a HashMap for each data point to store votes for each class
+        let mut polls = vec![HashMap::new(); x.shape().0];
+        // Retrieve the trained binary classifiers
+        let classifiers = self.classifiers.as_ref().unwrap();
+
+        // Iterate through each binary classifier
+        for i in 0..classifiers.len() {
+            let svc = classifiers.get(i).unwrap();
+            let predictions = svc.predict(x).unwrap(); // call SVC::predict for each binary classifier
+
+            // For each prediction from the current binary classifier
+            for (j, prediction) in predictions.iter().enumerate() {
+                let prediction = prediction.to_i32().unwrap();
+                let poll = polls.get_mut(j).unwrap(); // Get the poll for the current data point
+                                                      // Increment the vote for the predicted class
+                if let Some(count) = poll.get_mut(&prediction) {
+                    *count += 1
+                } else {
+                    poll.insert(prediction, 1);
+                }
+            }
+        }
+
+        // Determine the final prediction for each data point based on majority vote
+        Ok(polls
+            .iter()
+            .map(|v| {
+                // Find the class with the maximum votes for each data point
+                TX::from(*v.iter().max_by_key(|(_, class)| *class).unwrap().0).unwrap()
+            })
+            .collect())
+    }
+}
+
 #[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
 #[derive(Debug)]
 /// SVC Parameters
@@ -123,7 +306,7 @@ pub struct SVCParameters<TX: Number + RealNumber, TY: Number + Ord, X: Array2<TX
 )]
 /// Support Vector Classifier
 pub struct SVC<'a, TX: Number + RealNumber, TY: Number + Ord, X: Array2<TX>, Y: Array1<TY>> {
-    classes: Option<Vec<TY>>,
+    classes: Option<(TY, TY)>,
     instances: Option<Vec<Vec<TX>>>,
     #[cfg_attr(feature = "serde", serde(skip))]
     parameters: Option<&'a SVCParameters<TX, TY, X, Y>>,
@@ -152,7 +335,9 @@ struct Cache<TX: Number + RealNumber, TY: Number + Ord, X: Array2<TX>, Y: Array1
 struct Optimizer<'a, TX: Number + RealNumber, TY: Number + Ord, X: Array2<TX>, Y: Array1<TY>> {
     x: &'a X,
     y: &'a Y,
+    indices: Option<Vec<usize>>,
     parameters: &'a SVCParameters<TX, TY, X, Y>,
+    classes: &'a (TY, TY),
     svmin: usize,
     svmax: usize,
     gmin: TX,
@@ -180,12 +365,12 @@ impl<TX: Number + RealNumber, TY: Number + Ord, X: Array2<TX>, Y: Array1<TY>>
         self.tol = tol;
         self
     }
+
     /// The kernel function.
     pub fn with_kernel<K: Kernel + 'static>(mut self, kernel: K) -> Self {
         self.kernel = Some(Box::new(kernel));
         self
     }
-
     /// Seed for the pseudo random number generator.
     pub fn with_seed(mut self, seed: Option<u64>) -> Self {
         self.seed = seed;
@@ -241,17 +426,98 @@ impl<'a, TX: Number + RealNumber, TY: Number + Ord, X: Array2<TX>, Y: Array1<TY>
 impl<'a, TX: Number + RealNumber, TY: Number + Ord, X: Array2<TX> + 'a, Y: Array1<TY> + 'a>
     SVC<'a, TX, TY, X, Y>
 {
-    /// Fits SVC to your data.
-    /// * `x` - _NxM_ matrix with _N_ observations and _M_ features in each observation.
-    /// * `y` - class labels
-    /// * `parameters` - optional parameters, use `Default::default()` to set parameters to default values.
+    /// Fits a binary Support Vector Classifier (SVC) to the provided data.
+    ///
+    /// This is the primary `fit` method for a standalone binary SVC. It expects
+    /// the target labels `y` to contain exactly two unique classes. If more or
+    /// fewer than two classes are found, it returns an error. It then extracts
+    /// these two classes and proceeds to optimize and fit the SVC model.
+    ///
+    /// # Arguments
+    /// * `x` - A reference to the input features (2D array) of the training data.
+    /// * `y` - A reference to the target labels (1D array) of the training data. `y` must contain exactly two unique class labels.
+    /// * `parameters` - A reference to the `SVCParameters` controlling the training process.
+    ///
+    /// # Returns
+    /// A `Result` which is:
+    /// - `Ok(SVC<'a, TX, TY, X, Y>)`: A new, fitted binary SVC instance.
+    /// - `Err(Failed)`: If the number of unique classes in `y` is not exactly two, or if the underlying optimization fails.
     pub fn fit(
         x: &'a X,
         y: &'a Y,
         parameters: &'a SVCParameters<TX, TY, X, Y>,
     ) -> Result<SVC<'a, TX, TY, X, Y>, Failed> {
-        let (n, _) = x.shape();
+        let classes = y.unique();
+        // Validate that there are exactly two unique classes in the target labels.
+        if classes.len() != 2 {
+            return Err(Failed::fit(&format!(
+                "Incorrect number of classes: {}. A binary SVC requires exactly two classes.",
+                classes.len()
+            )));
+        }
+        let classes = (classes[0], classes[1]);
+        let svc = Self::optimize_and_fit(x, y, parameters, classes, None);
+        svc
+    }
 
+    /// Fits a binary Support Vector Classifier (SVC) specifically for multi-class scenarios.
+    ///
+    /// This function is intended to be called by a multi-class strategy (e.g., one-vs-one)
+    /// to train individual binary SVCs. It takes a `MultiClassConfig` which specifies
+    /// the two classes this SVC should discriminate and the subset of data indices
+    /// relevant to these classes. It then delegates the actual optimization and fitting
+    /// to `optimize_and_fit`.
+    ///
+    /// # Arguments
+    /// * `x` - A reference to the input features (2D array) of the training data.
+    /// * `y` - A reference to the target labels (1D array) of the training data.
+    /// * `parameters` - A reference to the `SVCParameters` controlling the training process (e.g., kernel, C-value, tolerance).
+    /// * `multiclass_config` - A `MultiClassConfig` struct containing:
+    ///     - `classes`: A tuple `(class0, class1)` specifying the two classes this SVC should distinguish.
+    ///     - `indices`: A `Vec<usize>` containing the indices of the data points in `x` and `y that belong to either `class0` or `class1`.`
+    ///
+    /// # Returns
+    /// A `Result` which is:
+    /// - `Ok(SVC<'a, TX, TY, X, Y>)`: A new, fitted binary SVC instance.
+    /// - `Err(Failed)`: If the fitting process encounters an error (e.g., invalid parameters).
+    fn multiclass_fit(
+        x: &'a X,
+        y: &'a Y,
+        parameters: &'a SVCParameters<TX, TY, X, Y>,
+        multiclass_config: MultiClassConfig<TY>,
+    ) -> Result<SVC<'a, TX, TY, X, Y>, Failed> {
+        let classes = multiclass_config.classes;
+        let indices = multiclass_config.indices;
+        let svc = Self::optimize_and_fit(x, y, parameters, classes, Some(indices));
+        svc
+    }
+
+    /// Internal function to optimize and fit the Support Vector Classifier.
+    ///
+    /// This is the core logic for training a binary SVC. It performs several checks
+    /// (e.g., kernel presence, data shape consistency) and then initializes an
+    /// `Optimizer` to find the support vectors, weights (`w`), and bias (`b`).
+    ///
+    /// # Arguments
+    /// * `x` - A reference to the input features (2D array) of the training data.
+    /// * `y` - A reference to the target labels (1D array) of the training data.
+    /// * `parameters` - A reference to the `SVCParameters` defining the SVM model's configuration.
+    /// * `classes` - A tuple `(class0, class1)` representing the two distinct class labels that the SVC will learn to separate.
+    /// * `indices` - An `Option<Vec<usize>>`. If `Some`, it contains the specific indices of data points from `x` and `y` that should be used for training this binary classifier. If `None`, all data points in `x` and `y` are considered.
+    /// # Returns
+    /// A `Result` which is:
+    /// - `Ok(SVC<'a, TX, TY, X, Y>)`: A new `SVC` instance populated with the learned model components (support vectors, weights, bias).
+    /// - `Err(Failed)`: If any of the validation checks fail (e.g., missing kernel, mismatched data shapes), or if the optimization process fails.
+    fn optimize_and_fit(
+        x: &'a X,
+        y: &'a Y,
+        parameters: &'a SVCParameters<TX, TY, X, Y>,
+        classes: (TY, TY),
+        indices: Option<Vec<usize>>,
+    ) -> Result<SVC<'a, TX, TY, X, Y>, Failed> {
+        let (n_samples, _) = x.shape();
+
+        // Validate that a kernel has been defined in the parameters.
         if parameters.kernel.is_none() {
             return Err(Failed::because(
                 FailedError::ParametersError,
@@ -259,55 +525,39 @@ impl<'a, TX: Number + RealNumber, TY: Number + Ord, X: Array2<TX> + 'a, Y: Array
             ));
         }
 
-        if n != y.shape() {
+        // Validate that the number of samples in X matches the number of labels in Y.
+        if n_samples != y.shape() {
             return Err(Failed::fit(
-                "Number of rows of X doesn\'t match number of rows of Y",
+                "Number of rows of X doesn't match number of rows of Y",
             ));
         }
 
-        let classes = y.unique();
-
-        if classes.len() != 2 {
-            return Err(Failed::fit(&format!(
-                "Incorrect number of classes: {}",
-                classes.len()
-            )));
-        }
-
-        // Make sure class labels are either 1 or -1
-        for e in y.iterator(0) {
-            let y_v = e.to_i32().unwrap();
-            if y_v != -1 && y_v != 1 {
-                return Err(Failed::because(
-                    FailedError::ParametersError,
-                    "Class labels must be 1 or -1",
-                ));
-            }
-        }
-
-        let optimizer: Optimizer<'_, TX, TY, X, Y> = Optimizer::new(x, y, parameters);
+        let optimizer: Optimizer<'_, TX, TY, X, Y> =
+            Optimizer::new(x, y, indices, parameters, &classes);
 
+        // Perform the optimization to find the support vectors, weight vector, and bias.
+        // This is where the core SVM algorithm (e.g., SMO) would run.
         let (support_vectors, weight, b) = optimizer.optimize();
 
+        // Construct and return the fitted SVC model.
         Ok(SVC::<'a> {
-            classes: Some(classes),
-            instances: Some(support_vectors),
-            parameters: Some(parameters),
-            w: Some(weight),
-            b: Some(b),
-            phantomdata: PhantomData,
+            classes: Some(classes), // Store the two classes the SVC was trained on.
+            instances: Some(support_vectors), // Store the data points that are support vectors.
+            parameters: Some(parameters), // Reference to the parameters used for fitting.
+            w: Some(weight),        // The learned weight vector (for linear kernels).
+            b: Some(b),             // The learned bias term.
+            phantomdata: PhantomData, // Placeholder for type parameters not directly stored.
         })
     }
-
     /// Predicts estimated class labels from `x`
     /// * `x` - _KxM_ data where _K_ is number of observations and _M_ is number of features.
     pub fn predict(&self, x: &'a X) -> Result<Vec<TX>, Failed> {
         let mut y_hat: Vec<TX> = self.decision_function(x)?;
 
         for i in 0..y_hat.len() {
-            let cls_idx = match *y_hat.get(i).unwrap() > TX::zero() {
-                false => TX::from(self.classes.as_ref().unwrap()[0]).unwrap(),
-                true => TX::from(self.classes.as_ref().unwrap()[1]).unwrap(),
+            let cls_idx = match *y_hat.get(i) > TX::zero() {
+                false => TX::from(self.classes.as_ref().unwrap().0).unwrap(),
+                true => TX::from(self.classes.as_ref().unwrap().1).unwrap(),
             };
 
             y_hat.set(i, cls_idx);
@@ -445,14 +695,18 @@ impl<'a, TX: Number + RealNumber, TY: Number + Ord, X: Array2<TX>, Y: Array1<TY>
     fn new(
         x: &'a X,
         y: &'a Y,
+        indices: Option<Vec<usize>>,
         parameters: &'a SVCParameters<TX, TY, X, Y>,
+        classes: &'a (TY, TY),
     ) -> Optimizer<'a, TX, TY, X, Y> {
         let (n, _) = x.shape();
 
         Optimizer {
             x,
             y,
+            indices,
             parameters,
+            classes,
             svmin: 0,
             svmax: 0,
             gmin: <TX as Bounded>::max_value(),
@@ -478,7 +732,12 @@ impl<'a, TX: Number + RealNumber, TY: Number + Ord, X: Array2<TX>, Y: Array1<TY>
             for i in self.permutate(n) {
                 x.clear();
                 x.extend(self.x.get_row(i).iterator(0).take(n).copied());
-                self.process(i, &x, *self.y.get(i), &mut cache);
+                let y = if *self.y.get(i) == self.classes.1 {
+                    1
+                } else {
+                    -1
+                } as f64;
+                self.process(i, &x, y, &mut cache);
                 loop {
                     self.reprocess(tol, &mut cache);
                     self.find_min_max_gradient();
@@ -514,14 +773,16 @@ impl<'a, TX: Number + RealNumber, TY: Number + Ord, X: Array2<TX>, Y: Array1<TY>
         for i in self.permutate(n) {
             x.clear();
             x.extend(self.x.get_row(i).iterator(0).take(n).copied());
-            if *self.y.get(i) == TY::one() && cp < few {
-                if self.process(i, &x, *self.y.get(i), cache) {
+            let y = if *self.y.get(i) == self.classes.1 {
+                1
+            } else {
+                -1
+            } as f64;
+            if y == 1.0 && cp < few {
+                if self.process(i, &x, y, cache) {
                     cp += 1;
                 }
-            } else if *self.y.get(i) == TY::from(-1).unwrap()
-                && cn < few
-                && self.process(i, &x, *self.y.get(i), cache)
-            {
+            } else if y == -1.0 && cn < few && self.process(i, &x, y, cache) {
                 cn += 1;
             }
 
@@ -531,14 +792,14 @@ impl<'a, TX: Number + RealNumber, TY: Number + Ord, X: Array2<TX>, Y: Array1<TY>
         }
     }
 
-    fn process(&mut self, i: usize, x: &[TX], y: TY, cache: &mut Cache<TX, TY, X, Y>) -> bool {
+    fn process(&mut self, i: usize, x: &[TX], y: f64, cache: &mut Cache<TX, TY, X, Y>) -> bool {
         for j in 0..self.sv.len() {
             if self.sv[j].index == i {
                 return true;
             }
         }
 
-        let mut g: f64 = y.to_f64().unwrap();
+        let mut g = y;
 
         let mut cache_values: Vec<((usize, usize), TX)> = Vec::new();
 
@@ -559,8 +820,8 @@ impl<'a, TX: Number + RealNumber, TY: Number + Ord, X: Array2<TX>, Y: Array1<TY>
         self.find_min_max_gradient();
 
         if self.gmin < self.gmax
-            && ((y > TY::zero() && g < self.gmin.to_f64().unwrap())
-                || (y < TY::zero() && g > self.gmax.to_f64().unwrap()))
+            && ((y > 0.0 && g < self.gmin.to_f64().unwrap())
+                || (y < 0.0 && g > self.gmax.to_f64().unwrap()))
         {
             return false;
         }
@@ -590,7 +851,7 @@ impl<'a, TX: Number + RealNumber, TY: Number + Ord, X: Array2<TX>, Y: Array1<TY>
             ),
         );
 
-        if y > TY::zero() {
+        if y > 0.0 {
             self.smo(None, Some(0), TX::zero(), cache);
         } else {
             self.smo(Some(0), None, TX::zero(), cache);
@@ -647,7 +908,6 @@ impl<'a, TX: Number + RealNumber, TY: Number + Ord, X: Array2<TX>, Y: Array1<TY>
         let gmin = self.gmin;
 
         let mut idxs_to_drop: HashSet<usize> = HashSet::new();
-
         self.sv.retain(|v| {
             if v.alpha == 0f64
                 && ((TX::from(v.grad).unwrap() >= gmax && TX::zero() >= TX::from(v.cmax).unwrap())
@@ -666,7 +926,11 @@ impl<'a, TX: Number + RealNumber, TY: Number + Ord, X: Array2<TX>, Y: Array1<TY>
 
     fn permutate(&self, n: usize) -> Vec<usize> {
         let mut rng = get_rng_impl(self.parameters.seed);
-        let mut range: Vec<usize> = (0..n).collect();
+        let mut range = if let Some(indices) = self.indices.clone() {
+            indices
+        } else {
+            (0..n).collect::<Vec<usize>>()
+        };
         range.shuffle(&mut rng);
         range
     }
@@ -965,12 +1229,12 @@ mod tests {
         ];
 
         let knl = Kernels::linear();
-        let params = SVCParameters::default()
+        let parameters = SVCParameters::default()
             .with_c(200.0)
             .with_kernel(knl)
             .with_seed(Some(100));
 
-        let y_hat = SVC::fit(&x, &y, &params)
+        let y_hat = SVC::fit(&x, &y, &parameters)
             .and_then(|lr| lr.predict(&x))
             .unwrap();
         let acc = accuracy(&y, &(y_hat.iter().map(|e| e.to_i32().unwrap()).collect()));
@@ -1070,6 +1334,56 @@ mod tests {
         assert!(acc >= 0.9, "accuracy ({acc}) is not larger or equal to 0.9");
     }
 
+    #[cfg_attr(
+        all(target_arch = "wasm32", not(target_os = "wasi")),
+        wasm_bindgen_test::wasm_bindgen_test
+    )]
+    #[test]
+    fn svc_multiclass_fit_predict() {
+        let x = DenseMatrix::from_2d_array(&[
+            &[5.1, 3.5, 1.4, 0.2],
+            &[4.9, 3.0, 1.4, 0.2],
+            &[4.7, 3.2, 1.3, 0.2],
+            &[4.6, 3.1, 1.5, 0.2],
+            &[5.0, 3.6, 1.4, 0.2],
+            &[5.4, 3.9, 1.7, 0.4],
+            &[4.6, 3.4, 1.4, 0.3],
+            &[5.0, 3.4, 1.5, 0.2],
+            &[4.4, 2.9, 1.4, 0.2],
+            &[4.9, 3.1, 1.5, 0.1],
+            &[7.0, 3.2, 4.7, 1.4],
+            &[6.4, 3.2, 4.5, 1.5],
+            &[6.9, 3.1, 4.9, 1.5],
+            &[5.5, 2.3, 4.0, 1.3],
+            &[6.5, 2.8, 4.6, 1.5],
+            &[5.7, 2.8, 4.5, 1.3],
+            &[6.3, 3.3, 4.7, 1.6],
+            &[4.9, 2.4, 3.3, 1.0],
+            &[6.6, 2.9, 4.6, 1.3],
+            &[5.2, 2.7, 3.9, 1.4],
+        ])
+        .unwrap();
+
+        let y: Vec<i32> = vec![0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2];
+
+        let knl = Kernels::linear();
+        let parameters = SVCParameters::default()
+            .with_c(200.0)
+            .with_kernel(knl)
+            .with_seed(Some(100));
+
+        let y_hat = MultiClassSVC::fit(&x, &y, &parameters)
+            .and_then(|lr| lr.predict(&x))
+            .unwrap();
+
+        let acc = accuracy(&y, &(y_hat.iter().map(|e| e.to_i32().unwrap()).collect()));
+
+        assert!(
+            acc >= 0.9,
+            "Multiclass accuracy ({acc}) is not larger or equal to 0.9"
+        );
+    }
+
     #[cfg_attr(
         all(target_arch = "wasm32", not(target_os = "wasi")),
         wasm_bindgen_test::wasm_bindgen_test
@@ -1106,8 +1420,8 @@ mod tests {
         ];
 
         let knl = Kernels::linear();
-        let params = SVCParameters::default().with_kernel(knl);
-        let svc = SVC::fit(&x, &y, &params).unwrap();
+        let parameters = SVCParameters::default().with_kernel(knl);
+        let svc = SVC::fit(&x, &y, &parameters).unwrap();
 
         // serialization
         let deserialized_svc: SVC<'_, f64, i32, _, _> =

src/svm/svr.rs

@@ -51,9 +51,9 @@
 //!
 //! let knl = Kernels::linear();
 //! let params = &SVRParameters::default().with_eps(2.0).with_c(10.0).with_kernel(knl);
-//! // let svr = SVR::fit(&x, &y, params).unwrap();
+//! let svr = SVR::fit(&x, &y, params).unwrap();
 //!
-//! // let y_hat = svr.predict(&x).unwrap();
+//! let y_hat = svr.predict(&x).unwrap();
 //! ```
 //!
 //! ## References:
@@ -80,11 +80,12 @@ use crate::error::{Failed, FailedError};
 use crate::linalg::basic::arrays::{Array1, Array2, MutArray};
 use crate::numbers::basenum::Number;
 use crate::numbers::floatnum::FloatNumber;
-use crate::svm::Kernel;
 
+use crate::svm::{Kernel, Kernels};
+
+/// SVR Parameters
 #[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
 #[derive(Debug)]
-/// SVR Parameters
 pub struct SVRParameters<T: Number + FloatNumber + PartialOrd> {
     /// Epsilon in the epsilon-SVR model.
     pub eps: T,
@@ -97,7 +98,7 @@ pub struct SVRParameters<T: Number + FloatNumber + PartialOrd> {
         all(feature = "serde", target_arch = "wasm32"),
         serde(skip_serializing, skip_deserializing)
     )]
-    pub kernel: Option<Box<dyn Kernel>>,
+    pub kernel: Option<Kernels>,
 }
 
 #[cfg_attr(feature = "serde", derive(Serialize, Deserialize))]
@@ -160,8 +161,8 @@ impl<T: Number + FloatNumber + PartialOrd> SVRParameters<T> {
         self
     }
     /// The kernel function.
-    pub fn with_kernel<K: Kernel + 'static>(mut self, kernel: K) -> Self {
-        self.kernel = Some(Box::new(kernel));
+    pub fn with_kernel(mut self, kernel: Kernels) -> Self {
+        self.kernel = Some(kernel);
         self
     }
 }
@@ -597,25 +598,25 @@ mod tests {
     use super::*;
     use crate::linalg::basic::matrix::DenseMatrix;
     use crate::metrics::mean_squared_error;
+    use crate::svm::search::svr_params::SVRSearchParameters;
     use crate::svm::Kernels;
 
-    // #[test]
-    // fn search_parameters() {
-    //     let parameters: SVRSearchParameters<f64, DenseMatrix<f64>, LinearKernel> =
-    //         SVRSearchParameters {
-    //             eps: vec![0., 1.],
-    //             kernel: vec![LinearKernel {}],
-    //             ..Default::default()
-    //         };
-    //     let mut iter = parameters.into_iter();
-    //     let next = iter.next().unwrap();
-    //     assert_eq!(next.eps, 0.);
-    //     assert_eq!(next.kernel, LinearKernel {});
-    //     let next = iter.next().unwrap();
-    //     assert_eq!(next.eps, 1.);
-    //     assert_eq!(next.kernel, LinearKernel {});
-    //     assert!(iter.next().is_none());
-    // }
+    #[test]
+    fn search_parameters() {
+        let parameters: SVRSearchParameters<f64, DenseMatrix<f64>> = SVRSearchParameters {
+            eps: vec![0., 1.],
+            kernel: vec![Kernels::linear()],
+            ..Default::default()
+        };
+        let mut iter = parameters.into_iter();
+        let next = iter.next().unwrap();
+        assert_eq!(next.eps, 0.);
+        // assert_eq!(next.kernel, LinearKernel {});
+        // let next = iter.next().unwrap();
+        // assert_eq!(next.eps, 1.);
+        // assert_eq!(next.kernel, LinearKernel {});
+        // assert!(iter.next().is_none());
+    }
 
     #[cfg_attr(
         all(target_arch = "wasm32", not(target_os = "wasi")),
@@ -648,7 +649,7 @@ mod tests {
             114.2, 115.7, 116.9,
         ];
 
-        let knl = Kernels::linear();
+        let knl: Kernels = Kernels::linear();
         let y_hat = SVR::fit(
             &x,
             &y,