“达观杯”文本智能处理挑战赛-任务五

一、LightGBM模型

1.简介

顾名思义，lightGBM包含两个关键点：light即轻量级，GBM 梯度提升机。

LightGBM 是一个梯度 boosting 框架，使用基于学习算法的决策树。它可以说是分布式的，高效的，有以下优势：

更快的训练效率

低内存使用

更高的准确率

支持并行化学习

可处理大规模数据

2. 特点

概括来说，lightGBM主要有以下特点：
基于Histogram的决策树算法
带深度限制的Leaf-wise的叶子生长策略
直方图做差加速
直接支持类别特征(Categorical Feature)
Cache命中率优化
基于直方图的稀疏特征优化
多线程优化
与常用的机器学习算法进行比较：速度飞起

二、LightGBM实践

from sklearn.model_selection import GridSearchCV
import pickle
import pandas as pd
from sklearn.externals import joblib
from sklearn.metrics import roc_auc_score,f1_score,accuracy_score
from sklearn.model_selection import GridSearchCV
from sklearn.model_selection import train_test_split
import lightgbm as LGB
import numpy as np

feat_path="E:/Jupyter/anava/DaGuanBei_Analysis/feat/"
model_path="E:/Jupyter/anava/DaGuanBei_Analysis/model/"
result_path="E:/Jupyter/anava/DaGuanBei_Analysis/result/"

#自定义验证集的评价函数
def f1_score_vali(preds, data_vali):
    labels = data_vali.get_label()
    preds = np.argmax(preds.reshape(19, -1), axis=0)
    score_vali = f1_score(y_true=labels, y_pred=preds, average='macro')
    return 'f1_score', score_vali, True

#读取特征
data_fp=open(feat_path + "data_w_tfidf.pkl", 'rb')
x_train, y_train, x_test = pickle.load(data_fp)
xTrain, xVal, yTrain, yVal = train_test_split(x_train, y_train, test_size=0.30, random_state=531)

d_train = LGB.Dataset(data=xTrain, label=yTrain)
d_vali = LGB.Dataset(data=xVal, label=yVal)

#构建模型
params = {
    'boosting': 'gbdt',
    'application': 'multiclassova',
    'num_class': 19,
    'learning_rate': 0.1,
    'num_leaves': 31,
    'max_depth': -1,
    'lambda_l1': 0,
    'lambda_l2': 0.5,
    'bagging_fraction': 1.0,
}
y_prob = bst.predict(xVal)
predict = np.argmax(y_prob .reshape(19, -1), axis=0)
print(len(predict ),len(yVal))
accuracy=accuracy_score(yVal,predict)
f1score=f1_score(yVal,predict ,average='micro')

print('Accuracy :%2f%%'% (accuracy*100.0))
print('F1_score :%f1'% f1score)

结果得分如下：