[第13名] 我的解决方案：建模的第二阶段

大家好 😊
这次比赛已经结束，我想分享一些想法……
这次我分享我的具体解决方案。你可以看到下面的工作流程：

第一阶段

在第一阶段，我使用这种类型的代码来获取模型：

def lmodelv(X_train,y_train,X_val,y_val,lgb_params):
    d_train = lgb.Dataset(X_train, label=y_train)
    d_valid = lgb.Dataset(X_val, label=y_val)
    watchlist = [d_train, d_valid]
    model = lgb.train(lgb_params,
                      train_set=d_train,
                      valid_sets=watchlist,
                      verbose_eval=0,
                      early_stopping_rounds=EARLY_STOPPING_ROUNDS)
    return model

def training_lgbv(X_train,y_train,X_test,txt='lgb'):
    skf=StratifiedKFold(n_splits = NUM_FOLDS,shuffle = True,random_state = RANDOM_STATE)
    yv=np.zeros((len(X_train),4))
    yt=np.zeros((len(X_test),4))
    for fold,(idx_tr,idx_vl) in enumerate(skf.split(X_train,y_train)):
        X_tr,y_tr=pd.DataFrame(X_train).iloc[idx_tr],y_train.iloc[idx_tr]
        X_vl,y_vl=pd.DataFrame(X_train).iloc[idx_vl],y_train.iloc[idx_vl]
        model=lmodelv(X_tr,y_tr,X_vl,y_vl,lgb_tune)
        # 查看模式推断、评估和预测
        yv[idx_vl]=model.predict(X_vl)
        yt+=model.predict(X_test)/NUM_FOLDS #predict_proba
        print(f"Found Metric in {fold}:{log_loss(y_vl,yv[idx_vl])}")
    print(f'Results of the training: {log_loss(y_train,yv)}')
    np.save(path+f'preds/train_{txt}_{log_loss(y_train,yv)}.npy',yv)
    np.save(path+f'preds/test_{txt}_{log_loss(y_train,yv)}.npy',yt)
    return yv,yt

第一个函数用于定义模型，另一个用于训练。我这样做是因为你可以通过两种方式优化算法。（例如微调 https://www.kaggle.com/awwalmalhi/extreme-fine-tuning-lgbm-using-7-step-training）
类似地，我编写了用于