2019数据智能算法大赛赛后复盘

趁着刚开学这段时间参加了一个比赛《2019数据智能算法大赛》，最终结果：26/694

赛题：商家客户购买转化率预测

购买转化率是品牌商家在电商平台运营时最关注的指标之一，本次大赛中云积互动提供了品牌商家的历史订单数据，参赛选手通过人工智能技术构建预测模型，预估用户人群在规定时间内产生购买行为的概率。

数据格式：

具体说明如下： 1. 训练集:round2_diac2019_train.csv ，每行数据代表一个训练样本，各字段之间由逗号分隔，包含一次购买行为中的基础信息与上下文信息。 2. 测试集: round2_diac2019_test.csv ，每行数据代表一个测试样本，格式为“customer_id, result”，代表用户ID和预测结果。 3. 各类字段数据的描述如下：

评分标准：

评分标准

评分算法通过logarithmic loss（记为logloss）评估模型效果，logloss越小越好。

其中N表示测试集样本数量，yi表示测试集中第i个样本的真实标签，pi表示第 i个样本的预估转化率， δ为惩罚系数。

AB榜的划分方式和比例：

【1】评分采用AB榜形式。排行榜显示A榜成绩，竞赛结束后2小时切换成B榜单。B榜成绩以选定的两次提交或者默认的最后两次提交的最高分为准，最终比赛成绩以B榜单为准。

【2】此题目的AB榜数据采用同时段数据是随机划分，A榜为随机抽样测试集50%数据，B榜为其余50%的数据。

我们的思路：

根据划分时间段，将数据划分为训练集和测试集，根据给的信息进行groupby和聚合操作，利用Xgboost算法，最终结果未进复赛：

先放上我们的代码：

  1 '''
  2 数据预处理
  3 '''
  4 import pandas as pd
  5 import numpy as np
  6 from xgboost import plot_importance
  7 import xgboost as xgb
  8 from matplotlib import pyplot as plt
  9 from sklearn.model_selection import train_test_split
 10 import datetime
 11 import warnings
 12 warnings.filterwarnings('ignore')
 13 # 读取原始数据     
 14 
 15 train = pd.read_csv('./round1_diac2019_train.csv',low_memory=False)
 16 
 17 all_customer = pd.DataFrame(train[['customer_id']]).drop_duplicates(['customer_id']).dropna() 
 18 
 19 train['order_pay_time'] = pd.to_datetime(train['order_pay_time'])
 20 train['order_pay_date'] = train['order_pay_time'].dt.date  #取日期  生成一个新的列 order_pay_date
 21 off_train = train[(train['order_pay_date'].astype(str)<='2013-07-03')] #用于预测
 22 off_test = train[(train.order_pay_date.astype(str)>'2013-07-03')&(train.order_pay_date !='null')] #用作上半年的标签
 23 online_train = train
 24 
 25 
 26 
 27 
 28 
 29 def feature_and_label(train_data,test_data,islabel):
 30     '''
 31     提取特征，构建特征集和标签
 32     '''
 33     # t1 代表用户购买同种货物的次数
 34     t1 = train_data[['customer_id','goods_id','order_pay_date','order_detail_goods_num']]
 35     t1 = t1[t1.order_pay_date !='null']
 36     t1 = t1.groupby(['customer_id','goods_id'])['order_detail_goods_num'].agg('sum').reset_index()
 37     t1 = t1.groupby('customer_id')['goods_id','order_detail_goods_num'].agg('max')
 38     t1.rename(columns={'goods_id':'customer_max_goods'},inplace=True)
 39     # t2 代表用户下单数
 40     t2 = train_data[['customer_id','order_id','order_pay_date']]
 41     t2 = t2[t2.order_pay_date !='null']
 42     t2['user_order_num'] = 1
 43     t2 = t2.groupby('customer_id').agg('sum').reset_index()
 44     # t3 代表商品价格
 45     t3 = train_data[['customer_id','goods_price','order_pay_date']]
 46     t3 = t3[t3.order_pay_date !='null']
 47     t3 = t3.groupby('customer_id')['goods_price'].agg({'price_mean':'mean','price_max':'max','price_min':'min'})
 48     # t4 代表用户的订单购买时间
 49     t4 = train_data[['customer_id','order_pay_date']]
 50     t4 = t4[t4.order_pay_date != 'null']
 51     t4 = t4.groupby(['customer_id'],as_index=False)['order_pay_date'].agg({'order_pay_date_first':'min','order_pay_date_last':'max'})
 52     t4['long_time'] = pd.to_datetime(t4['order_pay_date_last']) - pd.to_datetime(t4['order_pay_date_first'])
 53     t4['long_time'] = t4['long_time'].dt.days + 1  
 54     # t5 代表用户是否喜欢参与评价
 55     t5 = train_data[['customer_id','is_customer_rate']]
 56     t5 = t5.replace(np.nan,0)
 57     t5 = t5.groupby('customer_id')['is_customer_rate'].agg('sum').reset_index()
 58     # t6 代表父订单商品购买数量
 59     t6 = train_data[['customer_id','order_total_num','order_pay_date']]
 60     t6 = t6[t6.order_pay_date != 'null']
 61     t6 = t6.groupby('customer_id')['order_total_num'].agg({'order_num_mean':'mean'})
 62     # t7 代表父订单商品购买数量
 63     t7 = train_data[['customer_id','order_total_payment']]
 64     t7 = t7.groupby('customer_id')['order_total_payment'].agg({'order_pay_mean':'mean'})
 65     # t8 代表父订单优惠金额
 66     t8 = train_data[['customer_id','order_total_discount']]
 67     t8 = t8.groupby('customer_id')['order_total_discount'].agg({'order_discount_mean':'mean'})
 68     # t9 代表子订单商品购买数量
 69     t9 = train_data[['customer_id','order_detail_amount']]
 70     t9 = t9.groupby('customer_id')['order_detail_amount'].agg({'order_amount_mean':'mean'})
 71     
 72     # t10 代表子订单应付总金额
 73     t10 = train_data[['customer_id','order_detail_payment']]
 74     t10 = t10.groupby('customer_id')['order_detail_payment'].agg({'detail_pay_mean':'mean'})
 75     # t11 代表是否支持会员折扣
 76     t11 = train_data[['customer_id','goods_has_discount']]
 77     t11 = t11.groupby('customer_id')['goods_has_discount'].agg({'goods_discount_mean':'mean'})
 78     # t12 代表
 79     t12 = train_data[['customer_id','customer_gender','member_status']].drop_duplicates(['customer_id'])
 80     t12 = t12.fillna(0)
 81     # t13
 82     t13 = train_data[['customer_id','order_count']]
 83     t13 = t13.groupby('customer_id')['order_count'].agg({'order_count_mean':'mean'})
 84     # t14
 85     t14 = train_data[['customer_id','order_status']]
 86     t14 = t14.groupby('customer_id')['order_status'].agg({'order_status_mean':'mean'})
 87     
 88     # 特征组合
 89     feature_set = pd.merge(t1,t2,on='customer_id',how='left')
 90     feature_set = pd.merge(feature_set,t3,on='customer_id',how='left')
 91     feature_set = pd.merge(feature_set,t4,on='customer_id',how='left')
 92     feature_set = pd.merge(feature_set,t5,on='customer_id',how='left')
 93     feature_set = pd.merge(feature_set,t6,on='customer_id',how='left')
 94     feature_set = pd.merge(feature_set,t7,on='customer_id',how='left')
 95     feature_set = pd.merge(feature_set,t8,on='customer_id',how='left')
 96     feature_set = pd.merge(feature_set,t9,on='customer_id',how='left')
 97     feature_set = pd.merge(feature_set,t10,on='customer_id',how='left')
 98     feature_set = pd.merge(feature_set,t11,on='customer_id',how='left')
 99     feature_set = pd.merge(feature_set,t12,on='customer_id',how='left')
100     feature_set = pd.merge(feature_set,t13,on='customer_id',how='left')
101     feature_set = pd.merge(feature_set,t14,on='customer_id',how='left')
102     feature_set.drop_duplicates(['customer_id'])
103     del feature_set['order_pay_date_first']
104     # 构建标签
105     if islabel==False:
106         feature_set['order_pay_date_last'] = pd.to_datetime(test_data['order_pay_date'].min()) - pd.to_datetime(feature_set['order_pay_date_last'])
107         feature_set['order_pay_date_last'] = feature_set['order_pay_date_last'].dt.days + 1
108         feature_set['label'] = 0
109         feature_set.loc[feature_set['customer_id'].isin(list(test_data['customer_id'].unique())),'label'] = 1
110     else:
111         feature_set['order_pay_date_last'] = pd.to_datetime('2013-12-31') - pd.to_datetime(feature_set['order_pay_date_last'])
112         feature_set['order_pay_date_last'] = feature_set['order_pay_date_last'].dt.days + 1
113     
114     return feature_set
115 
116 '''
117 数据抽取
118 '''
119 train_cup = feature_and_label(off_train,off_test,False)
120 submit = feature_and_label(online_train,None,True)   
121 y = train_cup.pop('label')  
122 feature = [x for x in train_cup.columns if x not in ['customer_id']]
123 X = train_cup[feature]  
124 
125 '''
126 分割+算法实践
127 
128 '''
129 X_train, X_valid, y_train, y_valid = train_test_split(X, y, test_size=0.35, random_state=42,stratify=y)
130 
131 submit_df = submit[['customer_id']]
132 X_submit = submit[feature]      
133 
134 params ={'learning_rate': 0.1,
135           'max_depth': 6,                                    
136           'num_boost_round':3000,
137           'objective': 'binary:logistic', 
138           'random_state': 7,
139           'silent':0,
140           'eta':0.8,
141          'eval_metric':'auc'
142         }
143 dtrain = xgb.DMatrix(X_train, y_train)
144 dvalid = xgb.DMatrix(X_valid, y_valid)
145 dtest  = xgb.DMatrix(X_submit)
146 evallist = [(dtrain,"train"),(dvalid,"valid")]
147 bst = xgb.train(params, dtrain, evals=evallist, early_stopping_rounds=30)
148 y_submit = bst.predict(dtest, ntree_limit = bst.best_ntree_limit)
149 submit_df['result'] = y_submit
150 
151 submit_df
152 
153 # 写入最终数据
154 submit_df = submit_df.drop(0)
155 submit_df.to_csv('xgboost_simple.csv',index=False)
156 
157 
158 feature_score = bst.get_fscore()
159 feature_score = sorted(feature_score.items(), key=lambda x:x[1],reverse=True)
160 fs = []
161 for (key,value) in feature_score:
162     fs.append("{0},{1}
".format(key,value))
163     
164 with open('result/xgb_feature_score.txt','w') as f:
165     f.writelines("feature,score
")
166     f.writelines(fs)

核心Code

反思问题：

1、对数据竞赛的认识不够充分，未能充分与其他参赛选手沟通，参赛经验太少，都是自己在闭门造车，以后如果要参加的话，需要多看看大佬的分享，积累些经验；

2、找好队友很重要，本次队友之前并未接触过此类比赛，导致主要工作都是自己在做，未能有效发挥团队的优势，与队友未能有效沟通，导致队友的工作主要集中在调参；

3、通过这次比赛真正认识到了，打比赛是真的累，每天睁眼闭眼想的都是特征的构建，然后构建出来，发现效果反而下降，搞得人心累，特征这块真的有种炼丹的感觉，只能去试试，自己感觉没有一套所谓的逻辑在背后（也可能是自己认识太短浅）；

4、自己python功底还是不好，代码量不够，以后需要对深度学习框架进行学习，多敲代码，积累代码量。

最后：

有个好队友真的很重要，很重要很重要很重要，多沟通；

数据特征很重要，算法是其次，在做的过程中，先调好算法，写出Baseline，再根据其进行优化特征，会好很多，此次想用多个算法融合进行优化，但发现结果更差了，还是对这些算法框架的使用不是很熟悉。