[amazonaccess 1]logistic.py 特征提取

---恢复内容开始---

本文件对应logistic.py

amazonaccess介绍：

根据入职员工的定位(员工角色代码、角色所属家族代码等特征)判断员工是否有访问某资源的权限

logistic.py(python)的关键：

1.通过组合组合几个特征来获取新的特征

例如：组合MGR_ID ROLE_FAMILY得到新特征 hash((85475,290919))=1071656665

2.greedy feature selection

i.  首先从候选特征中选择1个在训练集上表现最好的特征，将其加入好特征goodfeatures中，并将该特征从中候选特征中排除

ii. 从候选特征中选择一个特征与goodfeatures中特征一起，选取在训练数据集中表现最好的特征，加入goodfeatures中，并将该特征从中候选特征中排除

iii.继续选取，直到在训练集上的表现不再增加为止

3.One Hot Encoding

例如：对数据离散数据 [23 33 33 44]进行编码

i. 首先relable，转换为 [0 1 1 2]

ii.对0进行编码 0 0 1   对应 23

对1进行编码 0 1 0   对应 33

对2进行编码 1 0 0   对应 44

这样在最后使用线性模型的时候，离散数据的每个标签都会对应一个权重

代码流程：

1.读取数据，去除ROLE_CODE属性

learner = 'log'
print "Reading dataset..."
train_data = pd.read_csv('train.csv')
test_data = pd.read_csv('test.csv')
submit=learner + str(SEED) + '.csv'
#去除ROLE_CODE特征，因为train和test数据需要同时做变换，所以合到一块
all_data = np.vstack((train_data.ix[:,1:-1], test_data.ix[:,1:-1]))
num_train = np.shape(train_data)[0]

2.对数据进行relable

# Transform data
print "Transforming data..."
# Relabel the variable values to smallest possible so that I can use bincount
# on them later.
relabler = preprocessing.LabelEncoder()
for col in range(len(all_data[0,:])):
    relabler.fit(all_data[:, col])
    all_data[:, col] = relabler.transform(all_data[:, col])

3.组合特征生成新特征，这里分别组合了2个特征和3个特征，分别生成(28-2)和(56-12)个新特征，并与原特征合并

在组合特征时，排除了(ROLE_FAMILY,ROLE_FAMILY_DESC)和(ROLE_ROLLUP_1,ROLE_ROLLUP_2)组合

因为特征中很多标签对应的数据只有1条或2条，将这些数据合并到个标签中

组合特征的函数

def group_data(data, degree=3, hash=hash):
    """
    numpy.array -> numpy.array
 
    Groups all columns of data into all combinations of triples
    """
    new_data = []
    m,n = data.shape
    for indicies in combinations(range(n), degree):
        #去除ROLE_TITLE和ROLE_FAMILY组合
    if 5 in indicies and 7 in indicies:
        print "feature Xd"
        #去除ROLE_ROLLUP_1和ROLE_ROLLUP_2组合
    elif 2 in indicies and 3 in indicies:
        print "feature Xd"
    else:
            new_data.append([hash(tuple(v)) for v in data[:,indicies]])
    return array(new_data).T

合并数据只有1条或两条的标签

dp = group_data(all_data, degree=2)
for col in range(len(dp[0,:])):
    relabler.fit(dp[:, col])
    dp[:, col] = relabler.transform(dp[:, col])
    uniques = len(set(dp[:,col]))
    maximum = max(dp[:,col])
    print col
    if maximum < 65534:
        count_map = np.bincount((dp[:, col]).astype('uint16'))
        for n,i in enumerate(dp[:, col]):
            #只有1条数据的标签，合并
            if count_map[i] <= 1:
                dp[n, col] = uniques
            #只有2条数据的标签，合并
            elif count_map[i] == 2:
                dp[n, col] = uniques+1
    else:
        for n,i in enumerate(dp[:, col]):
            if (dp[:, col] == i).sum() <= 1:
                dp[n, col] = uniques
            elif (dp[:, col] == i).sum() == 2:
                dp[n, col] = uniques+1
    print uniques # unique values
    uniques = len(set(dp[:,col]))
    print uniques
    relabler.fit(dp[:, col])
    dp[:, col] = relabler.transform(dp[:, col])

将新特征和原特征合并

# Collect the training features together
y = array(train_data.ACTION)
X = all_data[:num_train]
X_2 = dp[:num_train]
X_3 = dt[:num_train]
 
# Collect the testing features together
X_test = all_data[num_train:]
X_test_2 = dp[num_train:]
X_test_3 = dt[num_train:]
 
X_train_all = np.hstack((X, X_2, X_3))
X_test_all = np.hstack((X_test, X_test_2, X_test_3))

4.one hot encoding

def OneHotEncoder(data, keymap=None):
     """
     OneHotEncoder takes data matrix with categorical columns and
     converts it to a sparse binary matrix.
 
     Returns sparse binary matrix and keymap mapping categories to indicies.
     If a keymap is supplied on input it will be used instead of creating one
     and any categories appearing in the data that are not in the keymap are
     ignored
     """
     if keymap is None:
          keymap = []
          for col in data.T:
               uniques = set(list(col))
               keymap.append(dict((key, i) for i, key in enumerate(uniques)))
     total_pts = data.shape[0]
     outdat = []
     for i, col in enumerate(data.T):
          km = keymap[i]
          num_labels = len(km)
          spmat = sparse.lil_matrix((total_pts, num_labels))
          for j, val in enumerate(col):
               if val in km:
                    spmat[j, km[val]] = 1
          outdat.append(spmat)
     outdat = sparse.hstack(outdat).tocsr()
     return outdat, keymap
 
# Xts holds one hot encodings for each individual feature in memory
# speeding up feature selection
Xts = [OneHotEncoder(X_train_all[:,[i]])[0] for i in range(num_features)]

5.greedy feature selection

print "Performing greedy feature selection..."
score_hist = []
N = 10
good_features = set([])
# Greedy feature selection loop
while len(score_hist) < 2 or score_hist[-1][0] > score_hist[-2][0]:
    scores = []
    for f in range(len(Xts)):
        if f not in good_features:
            feats = list(good_features) + [f]
            Xt = sparse.hstack([Xts[j] for j in feats]).tocsr()
            score = cv_loop(Xt, y, model, N)
            scores.append((score, f))
            print "Feature: %i Mean AUC: %f" % (f, score)
    good_features.add(sorted(scores)[-1][1])
    score_hist.append(sorted(scores)[-1])
    print "Current features: %s" % sorted(list(good_features))
 
# Remove last added feature from good_features
good_features.remove(score_hist[-1][1])
good_features = sorted(list(good_features))
print "Selected features %s" % good_features
gf = open("feats" + submit, 'w')
print >>gf, good_features
gf.close()
print len(good_features), " features"

6.通过validation选取最优参数，logistic regression为regularization strength

print "Performing hyperparameter selection..."
# Hyperparameter selection loop
score_hist = []
Xt = sparse.hstack([Xts[j] for j in good_features]).tocsr()
if learner == 'NB':
    Cvals = [0.001, 0.003, 0.006, 0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.1]
else:
    Cvals = np.logspace(-4, 4, 15, base=2)  # for logistic
for C in Cvals:
    if learner == 'NB':
        model.alpha = C
    else:
        model.C = C
    score = cv_loop(Xt, y, model, N)
    score_hist.append((score,C))
    print "C: %f Mean AUC: %f" %(C, score)
bestC = sorted(score_hist)[-1][1]
print "Best C value: %f" % (bestC)

7.预测

print "Performing One Hot Encoding on entire dataset..."
Xt = np.vstack((X_train_all[:,good_features], X_test_all[:,good_features]))
Xt, keymap = OneHotEncoder(Xt)
X_train = Xt[:num_train]
X_test = Xt[num_train:]
 
if learner == 'NB':
    model.alpha = bestC
else:
    model.C = bestC
 
print "Training full model..."
print "Making prediction and saving results..."
model.fit(X_train, y)
preds = model.predict_proba(X_test)[:,1]
create_test_submission(submit, preds)
preds = model.predict_proba(X_train)[:,1]
create_test_submission('Train'+submit, preds)

---恢复内容结束---