gbdt分类算法

integr/my_gbdt.py

@@ -26,14 +26,14 @@ def fit(Xtrain, Ytrain):
 
     gx = Ytrain
 
-    for i in range(20):
+    for i in range(50):
         # 求残差
         gx = gx - fx0
         print("第", i, '轮 残差', gx[:10])
         clf = tree.DecisionTreeRegressor(criterion="mse", max_features=1, max_depth=4)
         clf.fit(Xtrain, gx)
 
-        fx0 = clf.predict(Xtrain)
+        fx0 = clf.predict(Xtrain)*0.7
         print("第", i, '轮 结果', fx0[:10])
         fx.append(fx0)
 

integr/my_gbdt_classic.py

@@ -0,0 +1,70 @@
+#-*- coding:utf-8 -*-
+import numpy as np
+from sklearn.ensemble import GradientBoostingClassifier,GradientBoostingRegressor
+from sklearn.model_selection import train_test_split
+from sklearn.datasets import load_wine,load_boston,load_breast_cancer
+from sklearn import tree
+
+
+def read_data():
+    boston = load_breast_cancer()
+    Xtrain, Xtest, Ytrain, Ytest = train_test_split(boston.data, boston.target, test_size=0.3)
+    for i in range(len(Ytrain)):
+        if Ytrain[i] == 0:
+            Ytrain[i] = -1
+    for i in range(len(Ytest)):
+        if Ytest[i] == 0:
+            Ytest[i] = -1
+    return Xtrain, Xtest, Ytrain, Ytest
+
+def init(Ytrain):
+    positive = sum(Ytrain == 1)
+    negative = Ytrain.shape[0] - positive
+    p = np.log2(positive/negative)
+    return np.ones(Ytrain.shape[0])*p
+
+def fit(Xtrain, Ytrain):
+    print("init", Ytrain[:10])
+    fx = []
+
+
+    fx0 = init(Ytrain)
+    fx.append(fx0)
+
+    print("0", fx0[:10])
+
+    gx = fx0
+
+    for i in range(50):
+        # 求伪残差
+        hx_0 = []
+        for j in range(Ytrain.shape[0]):
+            p = Ytrain[j] / (np.exp2(Ytrain[j]*gx[j]) + 1)
+            hx_0.append(p)
+
+        print("第", i, '轮 残差', gx[:10])
+        clf = tree.DecisionTreeRegressor(criterion="mse", max_features=1, max_depth=4)
+        clf.fit(Xtrain, np.array(hx_0))
+
+        fx_i = clf.predict(Xtrain)*0.7
+        print("第", i, '轮 结果', fx_i[:10])
+        fx.append(fx_i)
+        gx = gx + fx_i
+
+
+    gx = np.zeros(Ytrain.shape[0])
+    for i in range(len(fx)):
+        gx = gx + fx[i]
+
+    print(gx[:10])
+
+    gx = np.sign(gx)
+    p = sum(gx==Ytrain)/Ytrain.shape[0]
+
+    print(p)
+    return fx
+
+
+if __name__ == '__main__':
+    Xtrain, Xtest, Ytrain, Ytest = read_data()
+    fx = fit(Xtrain, Ytrain)