模型训练常用的一些diy functions

1.1 打印特征重要性

print(get_feature_importance_pair(gbm))
save_feature_importance_pair(feature_list, gbm)
或
importances = pd.Series(gbm.feature_importance('gain'), index=feature_list).sort_values(ascending=False)

1.2 给训练样本增加权重

from woe.eval import eval_segment_metrics

train['weight'] = 1
train.loc[(train['d7_open'] == 1) & (train['new_user'] == 1), 'weight'] =  (len(train) - sum(train['d7_open'])) / sum(train['d7_open'])
train.loc[(train['d7_open'] == 0) & (train['new_user'] == 1), 'weight'] = (len(train) - sum(train['d7_open'])) / sum(train['d7_open']) * 2
train.loc[(train['d7_open'] == 0) & (train['new_user'] == 0), 'weight'] = 1
weights = train['weight']

trainset = lgb.Dataset(data=train.loc[:, feature_list], label=train.loc[:, 'd7_open'], weight=weights)
eval_segment_metrics(target=test.loc[:, 'd7_open'].values,
    segment_cnt=50,
    predict_proba=gbm.predict(test.loc[:, feature_list]),
    out_path='../../reports/d7open.xlsx')

1.3 计算woe

def calc_woe(counter):
    counter = pd.DataFrame(counter).T
    counter.index.name = 'app'
    counter['count-d7_open'] = counter['count'] - counter['d7_open']
    counter['woe'] = np.log((counter['d7_open'] + 1e-4) /
        (counter['count-d7_open'] + 1e-4)) -     np.log(total_mean/(1-total_mean))
    counter['abs(woe)'] = np.abs(counter['woe'])
    counter['r7r'] = counter['d7_open'] / counter['count']
    return counter

1.4 决策树可视化

# 画出决策树
def plot_tree(gbm_model):
    ax = lgb.plot_tree(gbm_model, tree_index=0, figsize=(100, 40),
                   show_info=['split_gain', 'internal_value', 'internal_count', 'internal_weight', 'leaf_count', 'leaf_weight', 'data_percentage'])
    #plt.show()
    plt.savefig("lgb.png")
    graph = lgb.create_tree_digraph(gbm_model, tree_index=0, name='Treepdf')
    graph.render(view=True)

    # 或者使用pydotplus来画树
    '''
    import pydotplus
    from IPython.display import Image
    dot_data = lgb.export_graphviz(gbm, out_file=None,
                            feature_names=feature_list,
                            class_names=iris.target_names,
                            filled=True, rounded=True, special_characters=True)
    graph = pydotplus.graph_from_dot_data(dot_data)
    Image(graph.create_png())
    graph.write_png(r'pang.png')
    '''

1.5 门槛统计

 def model_stat(train, test, df_opt, dump_report=False):
     df_opt = df[df['score'] >= THRESHOLD]
     # r7r_opt = df_opt['d7_open'].sum() / df_opt['new_user'].sum()
     # cvr_opt = df_opt['new_user'].sum() / len(df_opt)
     # r7r = df['d3_open'].sum() / df['is_new_user'].sum()
     # cvr = df['is_new_user'].sum() / len(df)
     # print('r7r_before_opt = %.4f' % r7r)
     # print('r7r_lift = %.1f%%' % ((r7r_opt / r7r - 1) * 100))
     # print('cvr_before_opt = %.4f' % cvr)
     # print('cvr_after_opt = %.4f' % cvr_opt)
     # print('cvr_lift = %.1f%%' % ((cvr_opt / cvr - 1) * 100))
   
     res = pd.DataFrame()
     res['样本量'] = [len(train), len(test), len(df_opt), len(df_opt) / len(test), test['is_new_user'].sum(),
                   df_opt['is_new_user'].sum()]
     res['正样本量'] = [train['d3_open'].sum(), test['d3_open'].sum(), df_opt['d3_open'].sum(),
                    df_opt['d3_open'].sum() / test['d3_open'].sum(), None, None]
     res.index = ['训练数据', '测试数据', '测试门槛召回数据', '占比', '原始打开用户', '优化打开用户']
     if dump_report:
         res.to_excel('../../reports/model_stat.xlsx', index=True)
     return res

1.6 画PR曲线

def plot_pr_curve(y_test, y_score):
    import matplotlib.pyplot as plt
    from sklearn.metrics import precision_recall_curve
    precision, recall, _ = precision_recall_curve(y_test, y_score)
    plt.step(recall, precision, color='b', alpha=0.2, where='post')
    plt.fill_between(recall, precision, step='post', alpha=0.2, color='b')

    plt.xlabel('Recall')
    plt.ylabel('Precision')
    plt.ylim([0.0, 0.0175])
    plt.xlim([0.0, 1.0])
    plt.savefig('../../reports/figures/PR-curve.png')

1.7 测试集训练集划分

    print(data.groupby('pt')['uid'].count())
    tr = data[data['pt'] < '2021-01-03']
    te = data[data['pt'] == '2021-01-03']
    tr.to_csv('../data/train.data_huo_v4', sep='\t', index=False)
    te.to_csv('../data/test.data_huo_v4', sep='\t', index=False)

1.8 grp_id

def assign_grp_idx(df, segment_cnt):
    total_sample_cnt = len(df)
    grp_idx = 1
    start_idx = 0
    proba_descend_idx = np.argsort(y_pred_prob)[::-1]
    segment_sample_cnt = int(len(y_pred_prob) / segment_cnt)
    while start_idx < total_sample_cnt:
        segment_idx_list = proba_descend_idx[start_idx: start_idx + segment_sample_cnt]
        segment_sample_cnt = len(segment_idx_list)
        df.loc[segment_idx_list, 'grp'] = grp_idx
        grp_idx += 1
        start_idx += segment_sample_cnt

if __name__ == "__main__":
    segment_cnt = 50
    df['grp'] = -1
    assign_grp_idx(df, segment_cnt)
    print(df.groupby('grp').apply(lambda x: np.sum(x[LABEL]) / np.sum(x['is_recall'])))
    print(df.groupby('grp').apply(lambda x: np.sum(x['is_recall'] / len(x))))
    print(df.groupby('grp').apply(lambda x: np.sum(x[LABEL] / len(x))))