PCA impl

machine_learning/pca.py

@@ -0,0 +1,205 @@
+# Author: https://github.com/satori1995
+"""
+Principal Component Analysis (PCA) is primarily used for data dimensionality reduction.
+Its working principle involves mapping high-dimensional data to a lower-dimensional
+space through linear transformation while preserving the main information in the data.
+
+For example, if a sample has n features and n is very large, the training time
+will be severely impacted.
+However, in real-world situations, not all of these n features play a decisive role
+in determining the sample classification - some features have negligible impact.
+In such cases, PCA can be used for dimensionality reduction.
+
+Note: At this point, you might think that PCA simply filters features by
+selecting the most important ones. But that's not the case.
+What PCA actually does is create new features (principal components).
+These principal components are linear combinations of the original features,
+are orthogonal to each other (uncorrelated),
+and are arranged in descending order according to the maximum variance
+they capture in the data.
+
+Based on practical requirements, we only need the first k principal components
+to achieve the desired effect.
+As for the exact value of k, we can set a variance explanation ratio
+threshold (such as 95%) and select the minimum number of principal components that
+can explain this proportion of variance.
+In any case, k must be less than n.
+
+- The first few principal components usually capture most of the variability
+  in the data.
+  In many real-world datasets, the first 10~20% of principal components can explain
+  more than 90% of the data variance.
+- The later principal components often capture only small amounts of data variability,
+  with some primarily reflecting noise rather than useful information.
+
+Therefore, PCA is more powerful than simple feature selection because it can create
+new features(principal components) that are more effective than the original features
+while preserving the data structure.
+Each principal component is a weighted combination of all original features,
+just with different weights.
+This enables PCA to preserve the main patterns and structures of the original data
+while reducing dimensionality.
+
+Now you should understand what PCA is. Its characteristics are as follows:
+- Comprehensive information utilization: Each principal component integrates
+  information from all original features, just with different weights assigned.
+  This means that even if a feature's contribution is small, the useful information
+  it contains can still be preserved in the principal components.
+- Uncorrelatedness: Through orthogonal transformation, principal components
+  are uncorrelated with each other, eliminating redundant information and
+  multicollinearity problems that may exist between original features.
+- Noise filtering: Low-variance principal components often represent noise.
+  Discarding these components can improve the signal-to-noise ratio.
+- Data structure preservation: Although the dimensionality is reduced,
+  the main data structures and patterns are still preserved by
+  the first few high-variance principal components.
+
+In subsequent calculations and modeling,
+we use these principal components (the newly created features).
+Since the first k principal components already contain sufficient information
+from the original features (where k<n), we achieve dimensionality reduction.
+The benefits include faster training speed and lower memory consumption.
+
+So principal components are new features created by linearly
+combining the original features.
+These new features synthesize the information from the original features
+but represent the data in a more effective and streamlined way.
+Moreover, each principal component is orthogonal to the others - they extract
+information from the original features in mutually independent directions.
+The benefit of this approach is that it avoids information redundancy, with each
+principal component providing information that other principal components
+do not contain.
+
+These principal components are then arranged in descending order according to
+the amount of data variance they capture:
+- The first principal component is the new feature with the maximum variance.
+- The second principal component is the new feature with maximum variance in
+  the direction orthogonal to the first principal component.
+- And so on...
+
+For example, if a sample has 100 features, 100 principal components will be created.
+However, in practical scenarios, the first 10 principal components may already retain
+95% of the information from the original features.
+If this level of accuracy is acceptable, then we can select only the first
+10 principal components.
+This reduces the data from 100 dimensions to 10 dimensions, resulting in a
+significant speed improvement during training.
+"""
+
+import doctest
+
+import numpy as np
+
+
+class PCA:
+    def __init__(self, n_components=None):
+        """
+        Parameters
+        ----------
+        n_components : int or float, default=None
+            if n >= 1, Number of components to keep.
+
+            if n < 1, select the number of components such that the amount of
+            variance that needs to be explained is greater than the percentage
+            specified by n_components.
+
+            if n is None, keep min(n_samples, n_features) components.
+        """
+        self.n_components = n_components
+        self.components_ = None
+        self.explained_variance_ = None
+        self.explained_variance_ratio_ = None
+
+    def fit(self, x: np.ndarray):
+        # Zero-mean
+        x = x - np.mean(x, axis=0)
+        # calculate eigenvalues and eigenvectors
+        # the eigenvectors are the corresponding principal components
+        # the eigenvalues are the amount of variance that the corresponding principal
+        # components can explain
+        eigenvalues, eigenvectors = np.linalg.eig(x @ x.T)
+        # order by eigenvalues
+        sorted_indices = np.argsort(eigenvalues)[::-1]
+        eigenvalues = eigenvalues[sorted_indices]
+        eigenvectors = eigenvectors[:, sorted_indices]
+
+        # Matrix SVD, solve U, Σ, V
+        u = eigenvectors
+        singular_values = np.sqrt(np.where(eigenvalues > 0, eigenvalues, 0))
+        sigma = np.diag(singular_values)
+        sigma_inv = np.linalg.pinv(sigma)
+        v = x.T @ u @ sigma_inv
+
+        component_array = np.real(v.T[np.argsort(singular_values)[::-1]])
+        explained_variance_array = np.real(
+            np.sort(singular_values**2 / (len(x) - 1))[::-1]
+        )
+        explained_variance_ratio_array = explained_variance_array / np.sum(
+            explained_variance_array
+        )
+
+        if self.n_components is None:
+            n_components = min(x.shape)
+        elif self.n_components >= 1:
+            n_components = self.n_components
+        elif self.n_components < 1:
+            current_explained_variance_ratio = 0
+            i = 0
+            for i in range(len(explained_variance_ratio_array)):
+                current_explained_variance_ratio += explained_variance_ratio_array[i]
+                if current_explained_variance_ratio >= self.n_components:
+                    break
+            n_components = i + 1
+        else:
+            raise ValueError("n_components must be a number or None")
+
+        self.components_ = component_array[:n_components]
+        self.explained_variance_ = explained_variance_array[:n_components]
+        self.explained_variance_ratio_ = explained_variance_ratio_array[:n_components]
+
+    def transform(self, x: np.ndarray) -> np.ndarray:
+        # Project the centered data onto the selected principal components
+        return (x - np.mean(x, axis=0)) @ self.components_.T
+
+
+def main():
+    import time
+
+    from sklearn.datasets import load_digits
+    from sklearn.model_selection import train_test_split
+    from sklearn.neighbors import KNeighborsClassifier
+
+    data = load_digits()
+    x = data.data
+    y = data.target
+    x_train, x_test, y_train, y_test = train_test_split(x, y, random_state=520)
+    print(x_train.shape, y_train.shape)  # (1347, 64) (1347,)
+
+    knn_clf = KNeighborsClassifier()
+    # Use all features
+    knn_clf.fit(x_train, y_train)
+    start = time.perf_counter()
+    print("score:", knn_clf.score(x_test, y_test))  # score: 0.9822222222222222
+    end = time.perf_counter()
+    print("costs:", end - start)  # costs: 0.13106690000131493
+
+    # decomposition
+    pca = PCA(n_components=0.95)
+    pca.fit(x_train)
+    x_train_reduction = pca.transform(x_train)
+    x_test_reduction = pca.transform(x_test)
+    print(x_train_reduction.shape)  # (1347, 28)
+    print("n_features:", x_train_reduction.shape[1])  # n_features: 28
+    knn_clf = KNeighborsClassifier()
+    knn_clf.fit(x_train_reduction, y_train)
+    start = time.perf_counter()
+    print(
+        "score:", knn_clf.score(x_test_reduction, y_test)
+    )  # score: 0.9888888888888889
+    end = time.perf_counter()
+    print("costs:", end - start)  # costs: 0.010363900000811554
+
+
+if __name__ == "__main__":
+    doctest.testmod()
+    main()