合并XGBoost与Surprise模型时出现NoneType无drop属性报错如何解决

错误原因 #

pandas的DataFrame.drop()方法设置inplace=True参数时，会直接修改原DataFrame对象，返回值为None。你将该返回值赋值给new_df变量后，后续调用new_df.drop()本质是对None对象调用方法，自然触发AttributeError: 'NoneType' object has no attribute 'drop'报错。
另外还有一个隐藏问题：model_1如果直接修改传入的原始df，会导致后续model_2需要用到的studentId、testId列被提前删除，也会引发报错。

修正后的model_1代码 #

def model_1(df, test_indices=None):
    # 先做数据副本，避免修改原始传入的df
    df_copy = df.copy()
    cols_to_drop = ['testId', 'studentId']
    # 去掉inplace=True，直接接收drop后的新DataFrame
    new_df = df_copy.drop(cols_to_drop, axis=1)
    X = new_df.drop('result', axis=1) 
    y = new_df['result']
    # 如果传入了指定的测试集索引，就用指定的，保证和model2测试集对齐
    if test_indices is not None:
        X_train, X_test = X.drop(test_indices), X.loc[test_indices]
        y_train, y_test = y.drop(test_indices), y.loc[test_indices]
    else:
        X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, random_state=5)
    model = XGBClassifier() 
    model.fit(X_train, y_train) 
    # 输出正类预测概率，和协同过滤的0-1分数值域对齐
    y_pred_proba = model.predict_proba(X_test)[:, 1]
    # 把测试集的主键和预测结果一起返回，方便后续对齐
    res_df = df_copy.loc[X_test.index, ['studentId', 'testId', 'result']].copy()
    res_df['xgb_predictions'] = y_pred_proba
    return res_df

融合前需要解决两个核心问题：

• 两个模型必须使用同一份训练集、同一份测试集，避免预测的样本不匹配
• 两个模型的预测结果值域要统一，都转为0-1之间的连续概率/分数，再加权融合，最后设置阈值（比如0.5）得到最终分类标签

融合实现代码 #

def model_2(df, test_indices=None):
    df_copy = df.copy()
    reader = Reader(rating_scale=(0, 1))
    # 同样按指定索引切分数据集，保证和xgb的测试集一致
    if test_indices is not None:
        train_df = df_copy.drop(test_indices)
        test_df = df_copy.loc[test_indices]
        data = Dataset.load_from_df(train_df[['studentId', 'testId', 'result']], reader)
        trainset = data.build_full_trainset()
        # 构造surprise格式的测试集
        testset = list(test_df[['studentId', 'testId', 'result']].itertuples(index=False, name=None))
    else:
        data = Dataset.load_from_df(df_copy[['studentId', 'testId', 'result']], reader)
        trainset, testset = train_test_split(data, test_size=0.25, random_state=5)
    algo = KNNWithMeans()
    algo.fit(trainset)
    test = algo.test(testset)
    test = pd.DataFrame(test)
    test.drop("details", inplace=True, axis=1)
    test.columns = ['studentId', 'testId', 'actual', 'cf_predictions']
    return test

def merged_models(df, xgb_weight=0.5, cf_weight=0.5, threshold=0.5):
    # 先全局切分训练集和测试集，保证两个模型用同一份数据
    total_indices = df.index
    train_indices, test_indices = train_test_split(total_indices, test_size=0.25, random_state=5)
    # 分别获取两个模型的预测结果
    xgb_res = model_1(df, test_indices)
    cf_res = model_2(df, test_indices)
    # 按主键对齐两个预测结果
    merge_df = pd.merge(xgb_res, cf_res, on=['studentId', 'testId'], how='inner')
    # 加权融合
    merge_df['final_score'] = xgb_weight * merge_df['xgb_predictions'] + cf_weight * merge_df['cf_predictions']
    # 按阈值得到最终分类标签
    merge_df['final_pred'] = (merge_df['final_score'] >= threshold).astype(int)
    return merge_df

得到融合后的预测结果表merge_df后，用真实标签result（和actual是同一个值，任选其一即可）和final_pred/final_score计算分类指标即可：

• 分类硬指标：导入sklearn.metrics下的accuracy_score、precision_score、recall_score、f1_score，分别传入真实标签和final_pred计算
• 排序/概率指标：导入sklearn.metrics下的roc_auc_score，传入真实标签和final_score计算

如果需要调整最优权重，可以把权重作为超参数，在验证集上做网格搜索，选择指标最优的权重组合即可。

内容的提问来源于stack exchange，提问作者futuredataengineer