Add t-SNE algorithm for dimensionality reduction (#13432)

DIRECTORY.md

@@ -623,6 +623,7 @@
   * [Sequential Minimum Optimization](machine_learning/sequential_minimum_optimization.py)
   * [Similarity Search](machine_learning/similarity_search.py)
   * [Support Vector Machines](machine_learning/support_vector_machines.py)
+  * [t-SNE](machine_learning/tsne.py)
   * [Word Frequency Functions](machine_learning/word_frequency_functions.py)
   * [Xgboost Classifier](machine_learning/xgboost_classifier.py)
   * [Xgboost Regressor](machine_learning/xgboost_regressor.py)

machine_learning/tsne.py

@@ -0,0 +1,160 @@
+"""
+t-Distributed Stochastic Neighbor Embedding (t-SNE)
+---------------------------------------------------
+
+t-SNE is a nonlinear dimensionality reduction algorithm for visualizing
+high-dimensional data in a low-dimensional space (2D or 3D).
+
+It computes pairwise similarities in both spaces and minimizes the
+Kullback-Leibler divergence using gradient descent.
+
+References:
+- van der Maaten, L. & Hinton, G. (2008), JMLR.
+- https://lvdmaaten.github.io/tsne/
+"""
+
+import doctest
+
+import numpy as np
+from sklearn.datasets import load_iris
+
+
+def collect_dataset() -> tuple[np.ndarray, np.ndarray]:
+    """
+    Load Iris dataset and return features and labels.
+
+    Returns:
+        Tuple[np.ndarray, np.ndarray]: feature matrix and target labels
+
+    Example:
+    >>> x, y = collect_dataset()
+    >>> x.shape
+    (150, 4)
+    >>> y.shape
+    (150,)
+    """
+    data = load_iris()
+    return np.array(data.data), np.array(data.target)
+
+
+def compute_pairwise_affinities(data_x: np.ndarray, sigma: float = 1.0) -> np.ndarray:
+    """
+    Compute high-dimensional affinities (P matrix) using Gaussian kernel.
+
+    Args:
+        data_x: Input data of shape (n_samples, n_features)
+        sigma: Gaussian kernel bandwidth
+
+    Returns:
+        np.ndarray: Symmetrized probability matrix
+
+    Example:
+    >>> import numpy as np
+    >>> x = np.array([[0.0, 0.0], [1.0, 0.0]])
+    >>> p = compute_pairwise_affinities(x)
+    >>> float(round(p[0, 1], 3))
+    0.25
+    """
+    n_samples = data_x.shape[0]
+    sum_x = np.sum(np.square(data_x), axis=1)
+    d = np.add(np.add(-2 * np.dot(data_x, data_x.T), sum_x).T, sum_x)
+    p = np.exp(-d / (2 * sigma**2))
+    np.fill_diagonal(p, 0)
+    p /= np.sum(p)
+    return (p + p.T) / (2 * n_samples)
+
+
+def compute_low_dim_affinities(
+    y: np.ndarray,
+) -> tuple[np.ndarray, np.ndarray]:
+    """
+    Compute low-dimensional affinities (Q matrix) using Student-t distribution.
+
+    Args:
+        y: Low-dimensional embeddings of shape (n_samples, n_components)
+
+    Returns:
+        Tuple[np.ndarray, np.ndarray]: Q probability matrix and numerator array
+    """
+    sum_y = np.sum(np.square(y), axis=1)
+    num = 1 / (1 + np.add(np.add(-2 * np.dot(y, y.T), sum_y).T, sum_y))
+    np.fill_diagonal(num, 0)
+    q = num / np.sum(num)
+    return q, num
+
+
+def apply_tsne(
+    data_x: np.ndarray,
+    n_components: int = 2,
+    learning_rate: float = 200.0,
+    n_iter: int = 500,
+) -> np.ndarray:
+    """
+    Apply t-SNE for dimensionality reduction.
+
+    Args:
+        data_x: Original dataset (features)
+        n_components: Target dimension (2D or 3D)
+        learning_rate: Step size for gradient descent
+        n_iter: Number of iterations
+
+    Returns:
+        np.ndarray: Low-dimensional embedding of the data
+
+    Example:
+    >>> x, _ = collect_dataset()
+    >>> y_emb = apply_tsne(x, n_components=2, n_iter=50)
+    >>> y_emb.shape
+    (150, 2)
+    """
+    if n_components < 1 or n_iter < 1:
+        raise ValueError("n_components and n_iter must be >= 1")
+
+    n_samples = data_x.shape[0]
+    rng = np.random.default_rng()
+    y = rng.standard_normal((n_samples, n_components)) * 1e-4
+
+    p = compute_pairwise_affinities(data_x)
+    p = np.maximum(p, 1e-12)
+
+    y_inc = np.zeros_like(y)
+    momentum = 0.5
+
+    for i in range(n_iter):
+        q, num = compute_low_dim_affinities(y)
+        q = np.maximum(q, 1e-12)
+
+        pq = p - q
+        d_y = 4 * (
+            np.dot((pq * num), y)
+            - np.multiply(np.sum(pq * num, axis=1)[:, np.newaxis], y)
+        )
+
+        y_inc = momentum * y_inc - learning_rate * d_y
+        y += y_inc
+
+        if i == int(n_iter / 4):
+            momentum = 0.8
+
+    return y
+
+
+def main() -> None:
+    """
+    Run t-SNE on Iris dataset and display the first 5 embeddings.
+    """
+    data_x, _ = collect_dataset()
+    y_emb = apply_tsne(data_x, n_components=2, n_iter=300)
+
+    print("t-SNE embedding (first 5 points):")
+    print(y_emb[:5])
+
+    # Optional visualization (commented out)
+    # import matplotlib.pyplot as plt
+    # plt.scatter(y_emb[:, 0], y_emb[:, 1], c=_labels, cmap="viridis")
+    # plt.show()
+
+
+if __name__ == "__main__":
+    doctest.testmod()
+    main()