【阿里天池-医学影像报告异常检测】4 机器学习模型调参

阳光穿透心脏的1/2处 2022-11-15 01:10 241阅读 0赞

# 引言 #

（1）先对idtdf提取特征的ngram大小和feature调参，最终ngram=(1,2)feature=500，最佳  
（2）对LogisticRegression、XGBClassifier、LGBMClassifier三个模型单独调参，本人仅仅对XGB的几个参数进行了调整，工作量太庞大，就没有所有参数调整对比分析。这里仅仅提出调参的例子，提供模型调参的思路学习  
（3）开源源码[https://github.com/823316627bandeng/TIANCHI-2021-AI-Compition][https_github.com_823316627bandeng_TIANCHI-2021-AI-Compition]

# 实现 #

（1）导入包

import os
    import numpy as np
    import pandas as pd
    from sklearn.decomposition import NMF, TruncatedSVD, PCA
    from sklearn.linear_model import LogisticRegression
    from sklearn.multiclass import OneVsRestClassifier
    from sklearn.preprocessing import StandardScaler
    from xgboost import XGBClassifier
    from utils import *
    from lightgbm import LGBMClassifier
    from sklearn.model_selection import train_test_split
    from sklearn.feature_extraction.text import TfidfVectorizer
    from sklearn.metrics import roc_auc_score
    #加载数据
    label= np.array(pd.read_csv('./data/label.csv'))
    train = pd.read_csv('./temp/train.csv',header = None,names=['id','text','label'])
    adjust_model()

（2）模型的调参

def adjust_model():
        Tdf = TfidfVectorizer(ngram_range=(1,2),max_features=500)
        tdf_data = Tdf.fit_transform(train['text'])
        X_train,X_test,y_train,y_test = train_test_split(tdf_data,label,test_size=0.3)
        paralist = []
        score_dict = { "list_n":[],"list_f":[],"loss":[]}
        # for n in paralist
        param_test1 = { 'estimator__max_depth':range(2,8,2)}
        ''' model = OneVsRestClassifier(XGBClassifier(eval_metric= 'mlogloss', max_depth = 11, min_child_weight =1, use_label_encoder=False, learning_rate =0.01, n_estimators=150, gamma=0, subsample=0.8, colsample_bytree=0.8, nthread=100, scale_pos_weight=1, seed=27, verbose=True )) '''
        ''' model = OneVsRestClassifier(LGBMClassifier(is_unbalance = True, metric = 'binary_logloss,auc', # max_depth = 6, num_leaves = 40, learning_rate = 0.1, feature_fraction = 0.7, min_child_samples=21, min_child_weight=0.001, bagging_fraction = 1, bagging_freq = 2, reg_alpha = 0.001, reg_lambda = 8, cat_smooth = 0, # num_iterations = 200 )) '''
        # model = OneVsRestClassifier(LGBMClassifier())
        model = OneVsRestClassifier(XGBClassifier(eval_metric= 'mlogloss',use_label_encoder=False,n_estimators=150))
        model.fit(X_train, y_train)
    
        predict = model.predict_proba(X_test)
        score = roc_auc_score(y_test,predict)
        print(score)

> XGB  
> \{‘estimator\_\_max\_depth’: 9, ‘estimator\_\_min\_child\_weight’: 1\}  
> \{‘estimator\_\_max\_depth’: 11\}0.9812110365828264  
> \{‘estimator\_\_n\_estimators’: 150\} 0.9834881407453535  
> 调参后：0.9726861215062805  
> LGB  
> \{‘estimator\_\_max\_depth’: 6\}最佳得分 0.9811430144134826

（3）idtdf提取特征调参

#list_ngram = [1,2,3,4]
    #list_feature = [100,200,300,400]
    def adjust_idtdf():
        list_ngram = [1,2,3,4,5]
        list_feature = [100,200,300,400,500]
        #分数记录字典
        score_dict = { "list_n":[],"list_f":[],"loss":[]}
        #创建方法进行验证
        def para_Tdf(data_x):
            for n in list_ngram:
                for fea in list_feature:
                    Tdf = TfidfVectorizer(ngram_range=(1,n),max_features=fea)
                    tdf_data = Tdf.fit_transform(data_x)
                    # tdf_data = tdf_data.toarray()
                    X_train,X_test,y_train,y_test = train_test_split(tdf_data,label,test_size=0.3)
                    model = OneVsRestClassifier(XGBClassifier(eval_metric= 'mlogloss',use_label_encoder=False,n_estimators=50))
                    model.fit(X_train, y_train)
                    predict = model.predict_proba(X_test)
                    loss = Mutilogloss(y_test,predict)
                    score_dict["list_n"].append(n)
                    score_dict['list_f'].append(fea)
                    score_dict['loss'].append(loss)
                    print("n={0},feature={1},loss={2}".format(n,fea,loss))
        #方法调用
        para_Tdf(train['text'])
        #以DataFrame形式显示分数
        print(score_dict)

> 最佳是n=2,features = 500  
> n=1,feature=100,loss=0.09694388171340544  
> n=1,feature=200,loss=0.07941648607131963  
> n=1,feature=300,loss=0.0780516995282797  
> n=1,feature=400,loss=0.07654529189186797  
> n=1,feature=500,loss=0.07875673493941672  
> n=2,feature=100,loss=0.10700796997032506  
> n=2,feature=200,loss=0.0872626769884241  
> n=2,feature=300,loss=0.08134605319231948  
> n=2,feature=400,loss=0.07927331816025636  
> n=2,feature=500,loss=0.07391725763363112  
> n=3,feature=100,loss=0.10642417486808319  
> n=3,feature=200,loss=0.0932806660865527  
> n=3,feature=300,loss=0.0821267581008504  
> n=3,feature=400,loss=0.08258777666414407  
> n=3,feature=500,loss=0.07525704598697901  
> n=4,feature=100,loss=0.10395870861632356  
> n=4,feature=200,loss=0.09252871191998951  
> n=4,feature=300,loss=0.08208295772650118  
> n=4,feature=400,loss=0.08249975725295985  
> n=4,feature=500,loss=0.07920155662551372  
> n=5,feature=100,loss=0.10649166642764825  
> n=5,feature=200,loss=0.09238465463657325  
> n=5,feature=300,loss=0.08104836900458223  
> n=5,feature=400,loss=0.07833574743241475  
> n=5,feature=500,loss=0.07796380784547806

[https_github.com_823316627bandeng_TIANCHI-2021-AI-Compition]: https://github.com/823316627bandeng/TIANCHI-2021-AI-Compition