lightgbm 自定义metric函数

lgb 自定义函数

模板：

# f(preds: array, train_data: Dataset) -> name: str, eval_result: float, is_higher_better: bool
def f(preds, train_data):
    '''
    :param preds:  array, 预测值
    :param train_data:  lgb Dataset, lgb的传入数据集
    :return:
    '''
    ###  返回 (评估指标名称, 评估计算值, 是否评估值越大模型性能越好)
    return meaticname: str, eval_result: float, is_higher_better: bool

示例：

二分类 accuracy：

def accuracy(preds, train_data):
    labels = train_data.get_label()
    preds = 1. / (1. + np.exp(-preds))
    return 'accuracy', np.mean(labels == (preds > 0.5)), True

多分类 f1_score：

# 多分类 f1_score  14分类
def MultiAuc_f1score(preds, train_data):
    y_label = train_data.get_label()
    preds = preds.reshape(14,-1).T  # 14 分类
    y_pred = np.argmax(preds, axis=-1)
    multi_f1_score = f1_score(y_pred, y_label, average='macro')

    return 'MultiAuc_f1score', multi_f1_score, True

测试案例：

import sklearn
from sklearn import datasets
from sklearn.model_selection import train_test_split
from sklearn.metrics import roc_auc_score,mean_squared_error
import numpy as np
from matplotlib import pyplot as plt
import lightgbm as lgb
import pickle
from scipy.stats import rankdata


boston_price = datasets.load_breast_cancer()

data = boston_price.data
target = boston_price.target

# 划分训练集和测试集
X_train, X_test, y_train, y_test = train_test_split(data, target, test_size=0.2, random_state=1024)
print("Train data length:", len(X_train))
print("Test data length:", len(X_test))

# 转换为Dataset数据格式
lgb_train = lgb.Dataset(X_train, y_train)
lgb_eval = lgb.Dataset(X_test, y_test, reference=lgb_train)

# 参数
params = {
    'boosting_type': 'gbdt',  # 设置提升类型
    'objective': 'binary',  # 目标函数
    'metric': {'auc'},  # 评估函数
    'num_leaves': 64,  # 叶子节点数
    'learning_rate': 0.05,  # 学习速率
    'nthread': 20,
    'verbose': -1  # <0 显示致命的, =0 显示错误 (警告), >0 显示信息
}


### custom metric test
print("custom metric test ")
# 自定义metric
def custom_auc(preds, train_data):
    labels = train_data.get_label()
    # preds = 1. / (1. + np.exp(-preds))
    return 'my_auc', roc_auc_score(labels, preds,), True

evals_result = {}  # to record eval results
gbm = lgb.train(params, lgb_train, num_boost_round=100, valid_sets=lgb_eval,early_stopping_rounds=5,evals_result=evals_result,
                feval=custom_auc)

# 模型预测
y_pred = gbm.predict(X_test, num_iteration=gbm.best_iteration)  # shape [B]

执行结果：

custom metric test
[1] valid_0’s auc: 0.965602 valid_0’s my_auc: 0.965602
Training until validation scores don’t improve for 5 rounds.
[2] valid_0’s auc: 0.966304 valid_0’s my_auc: 0.966304
[3] valid_0’s auc: 0.965953 valid_0’s my_auc: 0.967883
[4] valid_0’s auc: 0.96683 valid_0’s my_auc: 0.96683
[5] valid_0’s auc: 0.962092 valid_0’s my_auc: 0.962092
[6] valid_0’s auc: 0.964549 valid_0’s my_auc: 0.964549
[7] valid_0’s auc: 0.964022 valid_0’s my_auc: 0.964022
[8] valid_0’s auc: 0.962969 valid_0’s my_auc: 0.962969
Early stopping, best iteration is:
[3] valid_0’s auc: 0.965953 valid_0’s my_auc: 0.967883

自定义的auc和lightgbm自带的auc评估效果一致，处理epoch=3时两者对不上，可能是因为内部的细节。