推荐算法_CIKM-2019-AnalytiCup 冠军源码解读_2

最近在为机器学习结合推荐算法的优化方法和数据来源想办法。抱着学习的态度继续解读19-AnalytiCup的冠军源码。

第一部分itemcf解读的连接： https://www.cnblogs.com/missouter/p/12701875.html

第二、三部分主要是特征提取和排序。在这篇博客中将作展开。

1、generate_static_features.ipynb 标题简洁明了 提取静态特征

import pandas as pd
import numpy as np

def reduce_mem_usage(df):
    """ iterate through all the columns of a dataframe and modify the data type
        to reduce memory usage.        
    """
    start_mem = df.memory_usage().sum() 
    print('Memory usage of dataframe is {:.2f} MB'.format(start_mem))
    
    for col in df.columns:
        col_type = df[col].dtype
        
        if col_type != object:
            c_min = df[col].min()
            c_max = df[col].max()
            if str(col_type)[:3] == 'int':
                if c_min > np.iinfo(np.int8).min and c_max < np.iinfo(np.int8).max:
                    df[col] = df[col].astype(np.int8)
                elif c_min > np.iinfo(np.int16).min and c_max < np.iinfo(np.int16).max:
                    df[col] = df[col].astype(np.int16)
                elif c_min > np.iinfo(np.int32).min and c_max < np.iinfo(np.int32).max:
                    df[col] = df[col].astype(np.int32)
                elif c_min > np.iinfo(np.int64).min and c_max < np.iinfo(np.int64).max:
                    df[col] = df[col].astype(np.int64)  
            else:
                if c_min > np.finfo(np.float16).min and c_max < np.finfo(np.float16).max:
                    df[col] = df[col].astype(np.float16)
                elif c_min > np.finfo(np.float32).min and c_max < np.finfo(np.float32).max:
                    df[col] = df[col].astype(np.float32)
                else:
                    df[col] = df[col].astype(np.float64)
        else:
            df[col] = df[col].astype('category')

    end_mem = df.memory_usage().sum() 
    print('Memory usage after optimization is: {:.2f} MB'.format(end_mem))
    print('Decreased by {:.1f}%'.format(100 * (start_mem - end_mem) / start_mem))
    
    return df

def load_data(path):
    user = reduce_mem_usage(pd.read_csv(path + 'user.csv',header=None))
    item = reduce_mem_usage(pd.read_csv(path + 'item.csv',header=None))
    data = pd.read_csv(path + 'user_behavior.csv',header=None)

    data.columns = ['userID','itemID','behavior','timestamp']
    data['day'] = data['timestamp'] // 86400
    data['hour'] = data['timestamp'] // 3600 % 24
    
    ## 生成behavior的onehot
    for i in ['pv','fav','cart','buy']:
        data[i] = 0
        data.loc[data['behavior'] == i, i] = 1

    ## 生成behavior的加权
    
    data['day_hour'] = data['day'] + data['hour'] / float(24)
    data.loc[data['behavior']=='pv','behavior'] = 1
    data.loc[data['behavior']=='fav','behavior'] = 2
    data.loc[data['behavior']=='cart','behavior'] = 3
    data.loc[data['behavior']=='buy','behavior'] = 1
    max_day = max(data['day'])
    min_day = min(data['day'])
    data['behavior'] = (1 - (max_day-data['day_hour']+2)/(max_day-min_day+2)) * data['behavior'] 

    item.columns = ['itemID','category','shop','brand']
    user.columns = ['userID','sex','age','ability']
    
    data = reduce_mem_usage(data)

    data = pd.merge(left=data, right=item, on='itemID',how='left')
    data = pd.merge(left=data, right=user, on='userID',how='left')

    return user, item, data
    

读取数据内存优化这块已经是老生常谈。loaddata()函数顺便完成了对各类行为权重的转换，值得一提的是购买权重被分配为1.而浏览、收藏等行为则被分配为1、2、3；目的是为了不向顾客推荐已购买过的商品。

主函数部分：

path = '../ECommAI_EUIR_round2_train_20190816/'

user, item, data = load_data(path = path)

for count_feature in ['itemID', 'shop', 'category','brand']:
    data[['behavior', count_feature]].groupby(count_feature, as_index=False).agg(
        {'behavior':'count'}).rename(columns={'behavior':count_feature + '_count'}).to_csv(str(count_feature)+'_count.csv', index=False)

for count_feature in ['itemID', 'shop', 'category','brand']:
    data[['behavior', count_feature]].groupby(count_feature, as_index=False).agg(
        {'behavior':'sum'}).rename(columns={'behavior':count_feature + '_sum'}).to_csv(str(count_feature)+'_sum.csv', index=False)

确定路径后，对item、shop、category与brand的特征进行提取。使用groupby().agg()分别提取用户行为权重的次数与累加和（agg参数'count'与'sum'）。生成文件分别储存于csv文件中。

temp = data[['behavior','category']].groupby('category', as_index=False).agg({'behavior': ['median','std','skew']})
temp.columns = ['category','category_median','category_std','category_skew']

temp.to_csv('category_higher.csv',index=False)

temp = data[['behavior','itemID']].groupby('itemID', as_index=False).agg({'behavior': ['median','std','skew']})
temp.columns = ['itemID','itemID_median','itemID_std','itemID_skew']

temp.to_csv('itemID_higher.csv',index=False)

上述代码使用groupby().agg()提取每个单独category、单独id的行为中值、标准差与偏斜。

data['age'] = data['age'] // 10
train = data[data['day'] < 15]

for count_feature in ['sex','ability','age']:
    data[['behavior','itemID',count_feature]].groupby(['itemID', count_feature], as_index=False).agg(
        {'behavior': 'count'}).rename(columns={'behavior':'user_to_'
                                               + count_feature + '_count'}).to_csv('item_to_' + str(count_feature)+'_count_online.csv', index=False)

这段以每个用户的基本数据（性别、对推荐系统的影响力、年龄）为基准，对其对应的行为次数进行特征提取。

itemcount = pd.read_csv('itemID_count.csv')

temp = pd.merge(left=item, right=itemcount, how='left', on='itemID')

item_rank = []
for eachcat in temp.groupby('category'):
    each_df = eachcat[1].sort_values('itemID_count', ascending=False).reset_index(drop=True)
    each_df['rank'] = each_df.index + 1
    lenth = each_df.shape[0]
    each_df['rank_percent'] = (each_df.index + 1) / lenth
    item_rank.append(each_df[['itemID','rank','rank_percent']])

使用merge对item与item的行为次数进行拼接。使用groupby按照商品类别进行分类。每个类别内商品按照商品的行为次数进行排序，算出商品的类内排名与排名百分比，

item_rank = pd.concat(item_rank, sort=False)

item_rank.to_csv('item_rank.csv',index=False)

将生成的类内排序使用concat()去除多余标签，写入文件。

def unique_count(x):
    return len(set(x))

cat1 = item.groupby('category',as_index=False).agg({'itemID': unique_count}).rename(columns={'itemID':'itemnum_undercat'})

cat2 = item.groupby('category',as_index=False).agg({'brand': unique_count}).rename(columns={'brand':'brandnum_undercat'})

cat3 = item.groupby('category',as_index=False).agg({'shop': unique_count}).rename(columns={'shop':'shopnum_undercat'})

pd.concat([cat1, cat2[['brandnum_undercat']], cat3[['shopnum_undercat']]], axis=1).to_csv('category_lower.csv',index=False)

这里先定义一个统计集合内元素数量的函数，应用在agg()中作为参数，用groupby以类别进行分类，统计每个类别中商品、品牌与商家的数量，写入csv文件。

2、generate_dynamic_feature.ipynb  提取动态特征

import pandas as pd
import numpy as np

def reduce_mem_usage(df):
    """ iterate through all the columns of a dataframe and modify the data type
        to reduce memory usage.        
    """
    start_mem = df.memory_usage().sum() 
    print('Memory usage of dataframe is {:.2f} MB'.format(start_mem))
    
    for col in df.columns:
        col_type = df[col].dtype
        
        if col_type != object:
            c_min = df[col].min()
            c_max = df[col].max()
            if str(col_type)[:3] == 'int':
                if c_min > np.iinfo(np.int8).min and c_max < np.iinfo(np.int8).max:
                    df[col] = df[col].astype(np.int8)
                elif c_min > np.iinfo(np.int16).min and c_max < np.iinfo(np.int16).max:
                    df[col] = df[col].astype(np.int16)
                elif c_min > np.iinfo(np.int32).min and c_max < np.iinfo(np.int32).max:
                    df[col] = df[col].astype(np.int32)
                elif c_min > np.iinfo(np.int64).min and c_max < np.iinfo(np.int64).max:
                    df[col] = df[col].astype(np.int64)  
            else:
                if c_min > np.finfo(np.float16).min and c_max < np.finfo(np.float16).max:
                    df[col] = df[col].astype(np.float16)
                elif c_min > np.finfo(np.float32).min and c_max < np.finfo(np.float32).max:
                    df[col] = df[col].astype(np.float32)
                else:
                    df[col] = df[col].astype(np.float64)
        else:
            df[col] = df[col].astype('category')

    end_mem = df.memory_usage().sum() 
    print('Memory usage after optimization is: {:.2f} MB'.format(end_mem))
    print('Decreased by {:.1f}%'.format(100 * (start_mem - end_mem) / start_mem))
    
    return df

def load_data(path):
    user = reduce_mem_usage(pd.read_csv(path + 'user.csv',header=None))
    item = reduce_mem_usage(pd.read_csv(path + 'item.csv',header=None))
    data = pd.read_csv(path + 'user_behavior.csv',header=None)

    data.columns = ['userID','itemID','behavior','timestamp']
    data['day'] = data['timestamp'] // 86400
    data['hour'] = data['timestamp'] // 3600 % 24
    
    ## 生成behavior的onehot
    for i in ['pv','fav','cart','buy']:
        data[i] = 0
        data.loc[data['behavior'] == i, i] = 1

    ## 生成behavior的加权
    
    data['day_hour'] = data['day'] + data['hour'] / float(24)
    data.loc[data['behavior']=='pv','behavior'] = 1
    data.loc[data['behavior']=='fav','behavior'] = 2
    data.loc[data['behavior']=='cart','behavior'] = 3
    data.loc[data['behavior']=='buy','behavior'] = 1
    max_day = max(data['day'])
    min_day = min(data['day'])
    data['behavior'] = (1 - (max_day-data['day_hour']+2)/(max_day-min_day+2)) * data['behavior'] 

    item.columns = ['itemID','category','shop','brand']
    user.columns = ['userID','sex','age','ability']
    
    data = reduce_mem_usage(data)

    data = pd.merge(left=data, right=item, on='itemID',how='left')
    data = pd.merge(left=data, right=user, on='userID',how='left')

    return user, item, data
    

与静态特征提取一样。

主函数部分：

#path = '..\\data\\'
path = '../ECommAI_EUIR_round2_train_20190816/'
user, item, data = load_data(path = path)

train = data[data['day'] < 15]

online_features = []
for count_feature in ['category','shop','brand']:
    train[['behavior','userID',count_feature]].groupby(['userID', count_feature], as_index=False).agg(
        {'behavior': 'count'}).rename(columns={'behavior':'user_to_'
                                               + count_feature + '_count'}).to_csv('user_to_' + str(count_feature)+'_count.csv', index=False)
for count_feature in ['category','shop','brand']:
    train[['behavior','userID',count_feature]].groupby(['userID', count_feature], as_index=False).agg(
        {'behavior': 'sum'}).rename(columns={'behavior':'user_to_' 
                                             + count_feature + '_sum'}).to_csv('user_to_' + str(count_feature)+'_sum.csv', index=False)

for count_feature in ['category','shop','brand']:
    for behavior_type in ['pv','fav','cart','buy']:
        train[[behavior_type,'userID',count_feature]].groupby(['userID', count_feature], as_index=False).agg(
            {behavior_type: 'sum'}).rename(columns={behavior_type:'user_to_'
                                                   + count_feature + '_count_' + behavior_type}).to_csv('user_to_' + str(count_feature) + '_count_' + behavior_type + '.csv', index=False)

将过去十五天的用户数据进行特征提取。同第一个文件一样的特征提取方式，只不过第二步提取的主体是用户。分别对用户与其产生行为的类别、商家与品牌进行次数、行为加权的特征提取。再对用户的四种行为类型与类别、商家与品牌进行累加和（次数？但它agg参数使用了sum）提取。最后写入csv文件。

yestday = data[data['day'] == 14]

for count_feature in ['category','shop','brand']:
    yestday[['behavior','userID',count_feature]].groupby(['userID', count_feature], as_index=False).agg(
        {'behavior': 'count'}).rename(columns={'behavior':'user_to_'
                                               + count_feature + '_count_yestday'}).to_csv('user_to_' + str(count_feature)+'_count_yestday.csv', index=False)

for count_feature in ['category','shop','brand']:
    for behavior_type in ['pv','fav','cart','buy']:
        yestday[[behavior_type,'userID',count_feature]].groupby(['userID', count_feature], as_index=False).agg(
            {behavior_type: 'sum'}).rename(columns={behavior_type:'user_to_'
                                                   + count_feature + '_count_' + behavior_type+'_yestday'}).to_csv('user_to_' + str(count_feature) + '_count_' + behavior_type + '_yestday.csv', index=False)

单独对昨天的用户数据进行提取，针对行为次数与类别写入csv文件。

a5days = data[(data['day'] > 15 - 5) & (data['day'] < 15)]

for count_feature in ['category','shop','brand']:
    a5days[['behavior','userID',count_feature]].groupby(['userID', count_feature], as_index=False).agg(
        {'behavior': 'count'}).rename(columns={'behavior':'user_to_'
                                               + count_feature + '_count_5days'}).to_csv('user_to_' + str(count_feature)+'_count_5days.csv', index=False)

for count_feature in ['category','shop','brand']:
    for behavior_type in ['pv','fav','cart','buy']:
        a5days[[behavior_type,'userID',count_feature]].groupby(['userID', count_feature], as_index=False).agg(
            {behavior_type: 'sum'}).rename(columns={behavior_type:'user_to_'
                                                   + count_feature + '_count_' + behavior_type+'_5days'}).to_csv('user_to_' + str(count_feature) + '_count_' + behavior_type + '_5days.csv', index=False)

针对近五天的用户数据进行提取，针对行为次数与类别写入csv文件。

start_timestamp  = max(data[data['day'] < 15]['timestamp'])

time_features = []
test = data[data['day'] < 15]
for time_feature in ['shop', 'category','brand']:
    time_features.append(test[['last_time','userID',time_feature,'day']].groupby(['userID',time_feature], as_index=False).agg({'last_time': 'min', 'day':'max'}).rename(columns={'last_time': 'user_to_'+ time_feature + '_lasttime', 'day':'user_to_'+ time_feature + '_lastday'}))

for f in time_features:
    f.to_csv(str(f.columns[2])+'.csv', index=False)

for f in time_features:
    print(str(f.columns[2])+'.csv')

对每个用户访问商户、品牌与类别的最新时间进行提取，写入csv中。

for count_feature in ['sex','ability','age']:
    train[['behavior','itemID',count_feature]].groupby(['itemID', count_feature], as_index=False).agg(
        {'behavior': 'count'}).rename(columns={'behavior':'user_to_'+ count_feature + '_count'}).to_csv('item_to_' + str(count_feature)+'_count.csv', index=False)

最后以每个用户的基本数据（性别、对推荐系统的影响力、年龄）为基准，对其对应的行为次数进行特征提取，生成一个与第一步对应的线下特征文件。

3、generate_time_feature.ipynb 提取时间特征

def reduce_mem_usage(df):
    """ iterate through all the columns of a dataframe and modify the data type
        to reduce memory usage.        
    """
    start_mem = df.memory_usage().sum() 
    print('Memory usage of dataframe is {:.2f} MB'.format(start_mem))
    
    for col in df.columns:
        col_type = df[col].dtype
        
        if col_type != object:
            c_min = df[col].min()
            c_max = df[col].max()
            if str(col_type)[:3] == 'int':
                if c_min > np.iinfo(np.int8).min and c_max < np.iinfo(np.int8).max:
                    df[col] = df[col].astype(np.int8)
                elif c_min > np.iinfo(np.int16).min and c_max < np.iinfo(np.int16).max:
                    df[col] = df[col].astype(np.int16)
                elif c_min > np.iinfo(np.int32).min and c_max < np.iinfo(np.int32).max:
                    df[col] = df[col].astype(np.int32)
                elif c_min > np.iinfo(np.int64).min and c_max < np.iinfo(np.int64).max:
                    df[col] = df[col].astype(np.int64)  
            else:
                if c_min > np.finfo(np.float16).min and c_max < np.finfo(np.float16).max:
                    df[col] = df[col].astype(np.float16)
                elif c_min > np.finfo(np.float32).min and c_max < np.finfo(np.float32).max:
                    df[col] = df[col].astype(np.float32)
                else:
                    df[col] = df[col].astype(np.float64)
        else:
            df[col] = df[col].astype('category')

    end_mem = df.memory_usage().sum() 
    print('Memory usage after optimization is: {:.2f} MB'.format(end_mem))
    print('Decreased by {:.1f}%'.format(100 * (start_mem - end_mem) / start_mem))
    
    return df

def load_data(path):
    user = reduce_mem_usage(pd.read_csv(path + 'user.csv',header=None))
    item = reduce_mem_usage(pd.read_csv(path + 'item.csv',header=None))
    data = pd.read_csv(path + 'user_behavior.csv',header=None)

    data.columns = ['userID','itemID','behavior','timestamp']
    data['day'] = data['timestamp'] // 86400
    data['hour'] = data['timestamp'] // 3600 % 24
    
    ## 生成behavior的onehot
    for i in ['pv','fav','cart','buy']:
        data[i] = 0
        data.loc[data['behavior'] == i, i] = 1

    ## 生成behavior的加权
    
    data['day_hour'] = data['day'] + data['hour'] / float(24)
    data.loc[data['behavior']=='pv','behavior'] = 1
    data.loc[data['behavior']=='fav','behavior'] = 2
    data.loc[data['behavior']=='cart','behavior'] = 3
    data.loc[data['behavior']=='buy','behavior'] = 1
    max_day = max(data['day'])
    min_day = min(data['day'])
    data['behavior'] = (1 - (max_day-data['day_hour']+2)/(max_day-min_day+2)) * data['behavior'] 

    item.columns = ['itemID','category','shop','brand']
    user.columns = ['userID','sex','age','ability']
    
    data = reduce_mem_usage(data)

    data = pd.merge(left=data, right=item, on='itemID',how='left')
    data = pd.merge(left=data, right=user, on='userID',how='left')

    return user, item, data
    

一样的读取步骤。

path = '../ECommAI_EUIR_round2_train_20190816/'
user, item, data = load_data(path = path)

train = data[data['day'] < 15]

start_timestamp  = max(train['timestamp'])

train['last_time'] = start_timestamp - train['timestamp']

timefeatures = []

for time_feature in ['itemID', 'shop', 'category','brand']:
    name = time_feature + '_last_time_underline.csv'
    tf = train[['last_time', time_feature]].groupby(
        time_feature, as_index=False).agg({'last_time':'min'}).rename(columns={'last_time': time_feature + 'last_time'})
    tf[time_feature + 'last_time_hour_ed'] = tf[time_feature + 'last_time'] // 3600 % 24
    timefeatures.append((name, tf))

for f in timefeatures:
    f[1].to_csv(f[0], index=False)

这里作者演示了一种提取某个商品/店铺/类别/品牌 距离第15、16天的最后一次点击的方法。通过计算最大时间戳减去每个访问的时间戳得到last_time，通过groupby()分类，agg()提取最小的last_time列得到最后一次点击的商品。

至此，特征提取的源码分析就结束了。这部分的代码给我的感觉是groupby().agg()使用的非常熟练老道，特征工程的构建有很多值得学习的地方。

源码直接跑起来会出现一些意想不到的bug，我们非常感谢原作者薛传雨提供的帮助。