import numpy as np class Node: j = None theta = None p = None left = None right = None class DecisionTreeBase: def __init__(self, max_depth, feature_sample_rate, get_score): self.max_depth = max_depth self.feature_sample_rate = feature_sample_rate self.get_score = get_score def split_data(self, j, theta, X, idx): idx1, idx2 = list(), list() for i in idx: value = X[i][j] if value <= theta: idx1.append(i) else: idx2.append(i) return idx1, idx2 def get_random_features(self, n): shuffled = np.random.permutation(n) size = int(self.feature_sample_rate * n) selected = shuffled[:size] return selected def find_best_split(self, X, y, idx): m, n = X.shape best_score = float("inf") best_j = -1 best_theta = float("inf") best_idx1, best_idx2 = list(), list() selected_j = self.get_random_features(n) for j in selected_j: thetas = set([x[j] for x in X]) for theta in thetas: idx1, idx2 = self.split_data(j, theta, X, idx) if min(len(idx1), len(idx2)) == 0 : continue score1, score2 = self.get_score(y, idx1), self.get_score(y, idx2) w = 1.0 * len(idx1) / len(idx) score = w * score1 + (1-w) * score2 if score < best_score: best_score = score best_j = j best_theta = theta best_idx1 = idx1 best_idx2 = idx2 return best_j, best_theta, best_idx1, best_idx2, best_score def generate_tree(self, X, y, idx, d): r = Node() r.p = np.average(y[idx], axis=0) if d == 0 or len(idx)<2: return r current_score = self.get_score(y, idx) j, theta, idx1, idx2, score = self.find_best_split(X, y, idx) if score >= current_score: return r r.j = j r.theta = theta r.left = self.generate_tree(X, y, idx1, d-1) r.right = self.generate_tree(X, y, idx2, d-1) return r def fit(self, X, y): self.root = self.generate_tree(X, y, range(len(X)), self.max_depth) def get_prediction(self, r, x): if r.left == None and r.right == None: return r.p value = x[r.j] if value <= r.theta: return self.get_prediction(r.left, x) else: return self.get_prediction(r.right, x) def predict(self, X): y = list() for i in range(len(X)): y.append(self.get_prediction(self.root, X[i])) return np.array(y)
最新发布