logicjake / 2020-yizhifu-credit-risk-user-identification-top2

第二届翼支付杯大数据建模大赛-信用风险用户识别Top2

🔭 I’m currently working on and learning:

• Network Embedding
• Recommender System

🏆 Competition

• 科大讯飞移动广告反欺诈算法挑战赛 Top2
• 图灵联邦视频点击预测大赛 Top3 Code
• KDD Cup 2020 Challenges for Modern E-Commerce Platform: Debiasing Top13 Code
• 2020腾讯广告算法大赛 Top12 Code
• 第二届翼支付杯大数据建模大赛-信用风险用户识别 Solo Top2 Code
• 零基础入门推荐系统 - 新闻推荐 Top2 Code
• Kaggle Riiid! Answer Correctness Prediction 前2% Code
• 一点资讯技术编程大赛CTR赛道 Top1 Code

🔧 Project

• MLCompetitionHub: 机器学习竞赛信息聚合(Machine learning competition information aggregation) Code Homepage
• WebMonitor: 实时监控网页变化并发送通知 (Real-time web page monitoring and change notifications) Code Homepage

2020-yizhifu-credit-risk-user-identification-top2's Issues

请问可以提供数据集吗？

您好，现比赛结束无法下载数据，请问能提供feature文件中的相应数据集吗？谢谢！

feature中的一些函数好像并没有用到

例如select——feature、w2v_emb、tfidf_emb等函数好像没有再使用过，请问作者这是什么原因？
`def select_feature(df, select_feature, ycol, p):
X = df[select_feature]
X.fillna(0, inplace=True)
Y = df[ycol]

selectChi2 = SelectPercentile(chi2, percentile=p).fit(X, Y)
selectF_classif = SelectPercentile(f_classif, percentile=p).fit(X, Y)

chi2_selected = selectChi2.get_support()
print('Chi2 selected {} features.'.format(chi2_selected.sum()))
f_classif_selected = selectF_classif.get_support()
print('F_classif selected {} features.'.format(f_classif_selected.sum()))
selected = chi2_selected & f_classif_selected
print('Chi2 & F_classif selected {} features'.format(selected.sum()))
selected_features = [f for f, s in zip(select_feature, selected) if s]

del_features = list(set(select_feature) - set(selected_features))
del_features.sort()
return del_features`

关于特征选择的一些问题

作者你好，自己是个新手想问一些问题，在做聚合的时候，我们有sum、mean、max、min等聚合函数可以选择，我注意到你对amount的一些聚合有时候只使用了sum,比如下面这段代码

for col in ['platform', 'type1', 'type2', 'hour']:
    df_temp = df_trans.pivot_table(
        index='user', columns=col, values='amount', aggfunc='sum').reset_index()
    df_temp.columns = [c if c == 'user' else 'trans_{}_{}_amount_sum'.format(
        col, c) for c in df_temp.columns]
    df_feature = df_feature.merge(df_temp, how='left')

for col in ['type1', 'hour']:
    df_temp = df_trans.pivot_table(
        index='user', columns=col, values='amount', aggfunc='mean').reset_index()
    df_temp.columns = [c if c == 'user' else 'trans_{}_{}_amount_mean'.format(
        col, c) for c in df_temp.columns]
    df_feature = df_feature.merge(df_temp, how='left')

for col in ['type1']:
    df_temp = df_trans.pivot_table(
        index='user', columns=col, values='amount', aggfunc='max').reset_index()
    df_temp.columns = [c if c == 'user' else 'trans_{}_{}_amount_max'.format(
        col, c) for c in df_temp.columns]
    df_feature = df_feature.merge(df_temp, how='left')

for col in ['type1']:
    df_temp = df_trans.pivot_table(
        index='user', columns=col, values='amount', aggfunc='min').reset_index()
    df_temp.columns = [c if c == 'user' else 'trans_{}_{}_amount_min'.format(
        col, c) for c in df_temp.columns]
    df_feature = df_feature.merge(df_temp, how='left')

我想问一下，这些聚合函数的选择有什么规律嘛？是要根据产生的数据再进行判断筛选吗？
感谢作者的开源

EDA的代码

比较想看到是怎么分析过程的，EDA有相关的代码吗