思路

基于上一个说的，一起买的策略，用代码实现一下，public LB从 0.0204到0.0215

• 取出最常被一起购买的商品
• 剩下的用补充

代码

import cudf

导入数据集

train = cudf.read_csv('../input/h-and-m-personalized-fashion-recommendations/transactions_train.csv')

#为了降低内存占用，id先转换为 int
train['customer_id'] = train['customer_id'].str[-16:].str.hex_to_int().astype('int64')
train['article_id'] = train.article_id.astype('int32')

# 按parquet存一份
train.t_dat = cudf.to_datetime(train.t_dat)
train = train[['t_dat','customer_id','article_id']]
train.to_parquet('train.pqt',index=False)
print( train.shape )
train.head()

(31788324, 3)

只要和最后一笔交易差距小于6天的交易记录

# 最后交易的日期
tmp = train.groupby('customer_id').t_dat.max().reset_index()
tmp.columns = ['customer_id','max_dat']
train = train.merge(tmp,on=['customer_id'],how='left')

# 这笔交易和最后交易的日期差了多少天
train['diff_dat'] = (train.max_dat - train.t_dat).dt.days

# 如果超过6天就不要了
train = train.loc[train['diff_dat']<=6]
print('Train shape:',train.shape)

Train shape: (5181535, 5)

计算出customer买了多少次该产品

# 计算出customer买了多少次该产品
tmp = train.groupby(['customer_id','article_id'])['t_dat'].agg('count').reset_index()
tmp.columns = ['customer_id','article_id','ct']

# 按照交易数量和交易时间倒序，这也是为了后面连成字符串的时候，交易数量越高，越近的article越靠前
train = train.merge(tmp,on=['customer_id','article_id'],how='left')
train = train.sort_values(['ct','t_dat'],ascending=False)
train = train.drop_duplicates(['customer_id','article_id'])
train = train.sort_values(['ct','t_dat'],ascending=False)
train.head()

pair_cudf.npy是一个map，把一个article_id映射为另一个最常一起购买的article_id

import pandas as pd, numpy as np
train = train.to_pandas()
pairs = np.load('../input/hmitempairs/pairs_cudf.npy',allow_pickle=True).item()
train['article_id2'] = train.article_id.map(pairs)

train2为针对每个交易的article，把它映射为最频繁的id

train2 = train[['customer_id','article_id2']].copy()
train2 = train2.loc[train2.article_id2.notnull()]
train2 = train2.drop_duplicates(['customer_id','article_id2'])
train2 = train2.rename({'article_id2':'article_id'},axis=1)

把train和train2结合到一起(包含过去买过的和最相关的)，去掉重复值（可能会有多条一样的）

train = train[['customer_id','article_id']]
train = pd.concat([train,train2],axis=0,ignore_index=True)
train.article_id = train.article_id.astype('int32')
train = train.drop_duplicates(['customer_id','article_id'])

把多个article id转换为字符串

train.article_id = ' 0' + train.article_id.astype('str')
preds = cudf.DataFrame( train.groupby('customer_id').article_id.sum().reset_index() )
preds.columns = ['customer_id','prediction']
preds.head()

取出上一周最火的12个article

train = cudf.read_parquet('train.pqt')
train.t_dat = cudf.to_datetime(train.t_dat)
train = train.loc[train.t_dat >= cudf.to_datetime('2020-09-16')]
top12 = ' 0' + ' 0'.join(train.article_id.value_counts().to_pandas().index.astype('str')[:12])
print("Last week's top 12 popular items:")
print( top12 )

Last week's top 12 popular items:
 0924243001 0924243002 0918522001 0923758001 0866731001 0909370001 0751471001 0915529003 0915529005 0448509014 0762846027 0714790020

写submission

sub = cudf.read_csv('../input/h-and-m-personalized-fashion-recommendations/sample_submission.csv')
sub = sub[['customer_id']]
sub['customer_id_2'] = sub['customer_id'].str[-16:].str.hex_to_int().astype('int64')

sub = sub.merge(preds.rename({'customer_id':'customer_id_2'},axis=1),\
    on='customer_id_2', how='left').fillna('')

del sub['customer_id_2']

把预测结果和top12加到一起

sub.prediction = sub.prediction + top12

只留12个，保存预测结果。

# 去掉左右空格，只取前131个: 12x10+11个空格=131
sub.prediction = sub.prediction.str.strip()
sub.prediction = sub.prediction.str[:131]

sub.to_csv(f'submission.csv',index=False)
sub.head()

最终public LB为0.0215