用于回归的油烟变化问题

Question

我正在估计数据不足和不平衡的机票销售情况。为了解决这个问题，我使用了smogn包中的smoter（用于回归的smote）。但是每次我运行模型时，我都会对目标有不同的预测。我认为吸烟者每次都会产生不同的输出数据。有什么办法可以解决这个随机状态？

请指导我在这里可以做什么，下面是代码段。

import smogn

def solution(df, p_bar: bool = 1, params: dict = model_params):
    # sort
    try:
        df = df.sort_values(["transition_date", "event_date"], ascending=False)
    except Exception as e:
        print("e")

    # bootstraping hyperparams
    n_samples = 140
    n_range = 40

    # tqdm
    if p_bar == 1:
        rg = tqdm(range(len(df)))
    else:
        rg = range(len(df))

    pred_list = []
    time.sleep(0.5)
    try:
        for i in rg:
            time.sleep(0.1)
            test_tour_id = df.iloc[i]['tour_id']
            df_without_test_tour = df[df['tour_id'] != test_tour_id].reset_index(drop=True)
            dt = smogn.smoter(
                data=df_without_test_tour,
                y='total_sales',
                k=3,
                samp_method='extreme',

                rel_thres=0.8,
                rel_method='auto',
                rel_xtrm_type='high',
                rel_coef=2.25
            )

            test_data = df.iloc[[i]].drop(e_col, axis=1)
            test_label = [df.iloc[i]['total_sales']]
            test_prid_id = df.iloc[i ]['promotion_id']

            train_data = dt.drop(e_col, axis=1)
            train_label = dt['total_sales']
            pred_tmp = []
            for j in range(n_range):
                x_train = resample(train_data, n_samples=n_samples, random_state=j)
                y_train = resample(train_label, n_samples=n_samples, random_state=j)

                model = xgb.get_model(x_train, y_train, params)
                pred = model.predict(test_data)

                pred_tmp.append(pred)

            pred = np.mean(pred_tmp, axis=0)
            mape_pred = abs(test_label - pred) * 100 / pred
            mape_real = abs(test_label - pred) * 100 / test_label

            pred_list.append([test_prid_id, pred[0], mape_pred[0], mape_real[0]])
    except Exception as ex:
        print(ex)
        tqdm._instances.clear() if p_bar == 1 else None

    pred = pd.DataFrame(pred_list, columns=['promotion_id', 'pred', 'mape_pred', 'mape_real'])
    return pd.merge(pred, df[e_col], how='left')

Answer 1

事实证明，Smote Regress在选择最近邻居时有一些随机性：

在此处查看其代码中的代码行：here

尽管我假设您正在使用Nick Kunz's Repository中的python版本，但我建议您使用here中的R one。如果您将Python用于机器学习项目，请考虑使用rpy2 python模块以便在R和Python之间进行通信