gbm.py

# init

# L - error function
def gbm_L(y, z):
	return (y-z)**2

# derivative of L function without coefficient `2`
def dbm_L_der(y, z):
	# ignore coefficient 2
	return y - z

# number of trees
J = 50
# tree coefficient
coef = 0.9
base_algorithms_list = []
coefficients_list = []

# predict function
def gbm_predict(X):
    return [np.sum([coeff * algo.predict([x])[0] for algo, coeff in zip(base_algorithms_list, coefficients_list)]) for x in X]

# train $$a_0$$ as simple DecisionTreeRegressor
b = DecisionTreeRegressor(max_depth=5 ,random_state=42)
b.fit(X_train,y_train)
base_algorithms_list.append(b)
coefficients_list.append(coef)

for j in xrange(1, J):
  # compute shift $$s = a_(N-1)(x_i) - a_j(x_i)$$
	b = DecisionTreeRegressor(max_depth=5 ,random_state=42)
        b.fit(X_train,y_train - gbm_predict(X_train))
	base_algorithms_list.append(b)
	coefficients_list.append(coef / (1.0 + j))

# test trained GBT with RMSE metric
print np.sqrt(mean_squared_error(y_test, gbm_predict(X_test)))