iT邦幫忙

2018 iT 邦幫忙鐵人賽
DAY 22
0

提醒: code在這裡

資料轉換(Transformation)

這個部分其實跟資料整合很像,不過不一定要透過多個欄位的協作已產生新的欄位。舉例來說,我們現在有每個人的名字的資料,這份資料如果要讓電腦讀得懂,其實有一定的難度,以下試圖從這個欄位找出兩個Feature,第一個比較簡單: 名字的長度,第二個是稱謂的分類。

# title
title_mapping= {
    'Ms':"Miss",
    'Mlle':"Miss",
    'Miss':"Miss",
    'Mrs':"Mrs",
    'Mme':"Mrs",
    'MrsMartin(ElizabethL':"Mrs",
    'Mr':"Mr"
    }

title_cat = {"Mr": 1, "Miss": 2, "Mrs": 3, "Master": 4, "Rare": 5}

def applyfun(row):
    row['Name_Length'] = len(row['Name'])  ## 名字的長度
    call = re.search(r'\,.+\.', row['Name']).group(0).replace(",", "").replace(".", "").replace(" ", "")
    call_cat = title_cat.get(title_mapping.get(call, "Rare"))
    row['Title'] = call_cat  ## 稱謂
    return row
df = df.apply(applyfun, axis=1)

資料切片(Discretization)

之所以要做資料切片的原因是有時候有時候連續變數電腦來說不是特別友善,雖然現在機器學習中套件大都已經有很強的自動切片功能,但資料量如果比較大,效能是一個考量,另外透過資料切片新增Feature,嘗試看看有時候還真的可以增加預測的精準度。至於資料切片的切片方法,比較常見的有兩個方式,以下舉「票價」(Fare)這個欄位示範。
https://ithelp.ithome.com.tw/upload/images/20171222/20107576jZfTS8dGTH.png

  1. 每一個等分傭有相同的資料量
ranges = pd.cut(df['Fare'], 5)  ## 回傳與原本索引對應的分類資料回來,如果覺得range太醜不好處理,可以把label設為False
df_Fare_discretization = pd.DataFrame(list(Counter(ranges).values()), columns=['count'], index= list(Counter(ranges).keys()))
df_Fare_discretization
range count
(204.932, 307.398] 17
(-0.512, 102.466] 838
(409.863, 512.329] 3
(102.466, 204.932] 33
  1. 每一個等分的長度相同
ranges = pd.qcut(df['Fare'], 5)
df_Fare_discretization = pd.DataFrame(list(Counter(ranges).values()), columns=['count'], index= list(Counter(ranges).keys()))
df_Fare_discretization
range count
(7.854, 10.5] 184
(39.688, 512.329] 176
(10.5, 21.679] 172
(-0.001, 7.854] 179
(21.679, 39.688] 180
  1. 其他切分方法
    其實學術上面還有其他比較科學的方法進行切分,不過對於最終預測的影響度並沒有想像中的大,因此這邊就不再多做介紹,有興趣的可以google關鍵字entropy。

統整

最後,我們把整份Titanic的資料前處理完整說明一次。

def preprocess(df):
    ## 處理None值
#      avg_age = df['Age'].mean() #作法一: 取平均值
#     avg_age = df['Age'].median() #作法二: 取中位數
#     作法三: 用相同的分布產生亂數塞入
    std = df['Age'].std()
    mean = df['Age'].mean()
    size = len(df[pd.isnull(df['Age'])])
    age_null_random_list = np.random.randint(mean - std, mean + std, size=size)
    df.loc[pd.isnull(df['Age']), 'Age'] = age_null_random_list
    
    # fare discretization
    def farecat(f):
        if f <= 16:
            cat = 0
        elif f > 17 and f <= 32:
            cat = 1
        elif f > 32 and f <= 48:
            cat = 2
        elif f < 48 and f <= 64:
            cat = 3
        elif f < 64 and f <= 80:
            cat = 4
        elif f < 80 and f <= 96:
            cat = 5
        else:
            cat = 6
        return cat

    # avg_fare
    fares_notnull = df[pd.notnull(df['Fare'])]['Fare']
    avg_fare = fares_notnull.median()

    # ticket
    ticket_cat = {}
    for ticket in df['Ticket']:
        if ticket.isdigit():
            ticket_cat[ticket] = 1
        elif ticket.startswith('A'):
            ticket_cat[ticket] = 2
        elif ticket.startswith('C'):
            ticket_cat[ticket] = 3
        elif ticket.startswith('F'):
            ticket_cat[ticket] = 4
        elif ticket.startswith('P'):
            ticket_cat[ticket] = 5
        elif ticket.startswith('SOTON'):
            ticket_cat[ticket] = 6
        elif ticket.startswith('STON'):
            ticket_cat[ticket] = 7
        elif ticket.startswith('S'):
            ticket_cat[ticket] = 8
        elif ticket.startswith('W'):
            ticket_cat[ticket] = 9
        else:
            ticket_cat[ticket] = 0
            
    ticket_cat1 = {}
    for num, name in enumerate(list(set([item.split()[0].replace(".", "").replace("/", "") for item in df['Ticket'] if not item.isdigit()]))):
        ticket_cat1[name] = num
    
    # cabin
    cabin_cat = {}
    for cabin in df['Cabin']:
        if pd.isnull(cabin):
            cabin_cat[cabin] = 0
        elif cabin.startswith('A'):
            cabin_cat[cabin] = 1
        elif cabin.startswith('B'):
            cabin_cat[cabin] = 2
        elif cabin.startswith('C'):
            cabin_cat[cabin] = 3
        elif cabin.startswith('D'):
            cabin_cat[cabin] = 4
        elif cabin.startswith('E'):
            cabin_cat[cabin] = 5
        else:
            cabin_cat[cabin] = 0
    
    # embarked
    embarked_cat = {}
    for embarked in df['Embarked']:
        if pd.isnull(embarked):
            embarked_cat[embarked] = 0
        elif embarked.startswith('S'):
            embarked_cat[embarked] = 0
        elif embarked.startswith('Q'):
            embarked_cat[embarked] = 1
        elif embarked.startswith('C'):
            embarked_cat[embarked] = 2
            
    # title
    title_mapping= {
        'Ms':"Miss",
        'Mlle':"Miss",
        'Miss':"Miss",
        'Mrs':"Mrs",
        'Mme':"Mrs",
        'MrsMartin(ElizabethL':"Mrs",
        'Mr':"Mr"
        }
    
    title_cat = {"Mr": 1, "Miss": 2, "Mrs": 3, "Master": 4, "Rare": 5}
    
    # Name_with specail chars
    def withspecailchar(name):
        for char in name:
            if char in string.punctuation:
                return 1
        return 0
    
    def applyfun(row):
        # Normalization
        row['Fare'] = row['Fare'] if pd.notnull(row['Fare']) else avg_fare
        row['Fare_log10'] = math.log(row['Fare'], 10) if  row['Fare'] != 0 else 0
                
        # Transformation
        row['Cabin'] = cabin_cat.get(row['Cabin'])  ## 整理艙位的類別
        row['Ticket'] = ticket_cat.get(row['Ticket'])  ## 整理票券類別
        row['Embarked'] = embarked_cat.get(row['Embarked'])  ## 整出發地類別
        row['Sex'] = 1 if row['Sex'] == 'male' else 0  ## 性別
        row['Name_Length'] = len(row['Name'])  ## Name => Name_Length
        row['Name_With_Special_Char'] = withspecailchar(row['Name'].replace(',', "").replace('.', ""))  ## Name => Name_With_Special_Char
        call = re.search(r'\,.+\.', row['Name']).group(0).replace(",", "").replace(".", "").replace(" ", "")
        call_cat = title_cat.get(title_mapping.get(call, "Rare"))
        row['Title'] = call_cat  ## Name => Title

        # Discretization
        row['Fare_Cat'] = farecat(row['Fare'])
        
        # Integration
        row['Family_Size'] = row['SibSp'] + row['Parch']
        row['Is_Alone']= 1 if row['Family_Size'] == 1 else 0
        row['Is_Mother']= 1 if row['Parch'] > 1 and row['Age']>20 else 0

        return row
    
    df = df.apply(applyfun, axis=1)
    df.drop(['Name', 'Fare'], axis=1, inplace=True)
    return df

df = pd.read_csv('train.csv')
df = preprocess(df)

上一篇
資料前處理
下一篇
機器學習系列概述
系列文
玩轉資料與機器學習-以自然語言處理為例31
圖片
  直播研討會
圖片
{{ item.channelVendor }} {{ item.webinarstarted }} |
{{ formatDate(item.duration) }}
直播中

尚未有邦友留言

立即登入留言