Add ml algorithms

machine_learning/adaboost.py

@@ -0,0 +1,157 @@
+"""
+AdaBoost implementation for binary classification using decision stumps.
+
+Reference: https://en.wikipedia.org/wiki/AdaBoost
+
+>>> import numpy as np
+>>> features = np.array([[0, 0], [1, 1], [1, 0], [0, 1]])
+>>> labels = np.array([0, 1, 1, 0])
+>>> clf = AdaBoost(n_estimators=5)
+>>> clf.fit(features, labels)
+>>> clf.predict(np.array([[0, 0], [1, 1]]))
+array([0, 1])
+"""
+
+from typing import Any
+
+import numpy as np
+
+
+class AdaBoost:
+    def __init__(self, n_estimators: int = 50) -> None:
+        """
+        Initialize AdaBoost classifier.
+
+        Args:
+            n_estimators: Number of boosting rounds (weak learners).
+        """
+        self.n_estimators: int = n_estimators
+        self.alphas: list[float] = []  # Weights assigned to each weak learner
+        self.models: list[dict[str, Any]] = []  # Stores each decision stump
+
+    def fit(self, feature_matrix: np.ndarray, target: np.ndarray) -> None:
+        """
+        Train AdaBoost model using decision stumps.
+
+        Args:
+            feature_matrix: 2D array of shape (n_samples, n_features)
+            target: 1D array of binary labels (0 or 1)
+        """
+        n_samples, _ = feature_matrix.shape
+
+        # Initialize uniform sample weights
+        sample_weights = np.ones(n_samples) / n_samples
+
+        # Reset model state
+        self.models = []
+        self.alphas = []
+
+        # Convert labels to {-1, 1} for boosting
+        y_signed = np.where(target == 0, -1, 1)
+
+        for _ in range(self.n_estimators):
+            # Train a weighted decision stump
+            stump = self._build_stump(feature_matrix, y_signed, sample_weights)
+            pred = stump["pred"]
+            err = stump["error"]
+
+            # Compute alpha (learner weight) with numerical stability
+            alpha = 0.5 * np.log((1 - err) / (err + 1e-10))
+
+            # Update sample weights to focus on misclassified points
+            sample_weights *= np.exp(-alpha * y_signed * pred)
+            sample_weights /= np.sum(sample_weights)
+
+            # Store the stump and its weight
+            self.models.append(stump)
+            self.alphas.append(alpha)
+
+    def predict(self, feature_matrix: np.ndarray) -> np.ndarray:
+        """
+        Predict binary class labels for input samples.
+
+        Args:
+            feature_matrix: 2D array of shape (n_samples, n_features)
+
+        Returns:
+            1D array of predicted labels (0 or 1)
+        """
+        clf_preds = np.zeros(feature_matrix.shape[0])
+
+        # Aggregate predictions from all stumps
+        for alpha, stump in zip(self.alphas, self.models):
+            pred = self._stump_predict(
+                feature_matrix,
+                stump["feature"],
+                stump["threshold"],
+                stump["polarity"],
+            )
+            clf_preds += alpha * pred
+
+        # Final prediction: sign of weighted sum
+        return np.where(clf_preds >= 0, 1, 0)
+
+    def _build_stump(
+        self,
+        feature_matrix: np.ndarray,
+        target_signed: np.ndarray,
+        sample_weights: np.ndarray,
+    ) -> dict[str, Any]:
+        """
+        Build the best decision stump for current sample weights.
+
+        Returns:
+            Dictionary containing stump parameters and predictions.
+        """
+        _, n_features = feature_matrix.shape
+        min_error = float("inf")
+        best_stump: dict[str, Any] = {}
+
+        # Iterate over all features and thresholds
+        for feature in range(n_features):
+            thresholds = np.unique(feature_matrix[:, feature])
+            for threshold in thresholds:
+                for polarity in [1, -1]:
+                    pred = self._stump_predict(
+                        feature_matrix,
+                        feature,
+                        threshold,
+                        polarity,
+                    )
+                    error = np.sum(sample_weights * (pred != target_signed))
+
+                    # Keep stump with lowest weighted error
+                    if error < min_error:
+                        min_error = error
+                        best_stump = {
+                            "feature": feature,
+                            "threshold": threshold,
+                            "polarity": polarity,
+                            "error": error,
+                            "pred": pred.copy(),
+                        }
+
+        return best_stump
+
+    def _stump_predict(
+        self,
+        feature_matrix: np.ndarray,
+        feature: int,
+        threshold: float,
+        polarity: int,
+    ) -> np.ndarray:
+        """
+        Predict using a single decision stump.
+
+        Returns:
+            1D array of predictions in {-1, 1}
+        """
+        pred = np.ones(feature_matrix.shape[0])
+
+        # Apply polarity to threshold comparison
+        if polarity == 1:
+            pred[feature_matrix[:, feature] < threshold] = -1
+        else:
+            pred[feature_matrix[:, feature] > threshold] = -1
+
+        return pred