本文围绕IEEE-CIS欺诈检测赛题展开，目标是识别欺诈交易。介绍了训练集和测试集数据情况，含交易和身份数据字段。阐述了关键策略，如构建用户唯一标识、聚合特征等，还涉及特征选择、编码、验证策略及模型训练，最终线上评分为0.959221，旨在学习特征构建。
☞☞☞AI 智能聊天, 问答助手, AI 智能搜索, 免费无限量使用 DeepSeek R1 模型☜☜☜
![【金融风控系列】_[2]_欺诈识别 - php中文网](https://img.php.cn/upload/article/202507/22/2025072211055170562.jpg)
该赛题的主要目标是识别出每笔交易是否是欺诈的。
其中训练集样本约59万（欺诈占3.5%），测试集样本约50万。
数据主要分为2类，交易数据transaction和identity数据。
本文主要是对与参考文献的收集整理
| Field | Description |
|---|---|
| TransactionDT | 来自给定参考日期时间的时间增量（不是实际时间戳） |
| TransactionAMT | 以美元为单位的交易支付金额 |
| ProductCD：产品代码，每笔交易的产品 | |
| card1 - card6 | 支付卡信息，如卡类型、卡类别、发卡行、国家等 |
| addr | 地址 |
| dist | 距离 |
| P_ 和 (R__) emaildomain | 购买者和收件人的电子邮件域 |
| C1-C14 | 计数，如发现有多少地址与支付卡关联等，实 |
| D1-D15 | timedelta，例如上次交易之间的天数等 |
| M1-M9 | 匹配，如卡上的姓名和地址等 |
| Vxxx | Vesta 设计了丰富的功能，包括排名、计数和其他实体关系 |
该表中的变量是身份信息——与交易相关的网络连接信息（IP、ISP、代理等）和数字签名（UA/浏览器/操作系统/版本等）。
它们由 Vesta 的欺诈保护系统和数字安全合作伙伴收集。
（字段名称被屏蔽，不提供成对字典用于隐私保护和合同协议）
[1] https://zhuanlan.zhihu.com/p/85947569
[2] https://www.kaggle.com/c/ieee-fraud-detection/discussion/111284
[3] https://www.kaggle.com/c/ieee-fraud-detection/discussion/111308
[4] https://www.kaggle.com/c/ieee-fraud-detection/discussion/101203
标记的逻辑是将卡上报告的退款定义为欺诈交易 (isFraud=1)，并将其后的用户帐户、电子邮件地址或账单地址直接关联到这些属性的交易也定义为欺诈。如果以上均未在120天内出现，则我们定义该笔定义为合法交易（isFraud=0）。
你可能认为 120 天后，一张卡片就变成了isFraud=0。我们很少在训练数据中看到这一点。（也许欺诈性信用卡会被终止使用）。训练数据集有 73838 个客户（信用卡）有2 个或更多交易。其中，71575 (96.9%) 始终为isFraud=0，2134 (2.9%) 始终为isFraud=1。只有129（0.2％）具有的混合物isFraud=0和isFraud=1。
从中，我们可以获得在业务中欺诈的逻辑，一个用户有过欺诈经历，那么他下次欺诈的概率还是非常高的，我们需要关注到这一点。
原始数据中未包含唯一UID，因此需要对客户进行唯一标识，识别客户的关键是三列card1，addr1和D1
D1 列是“自客户（信用卡）开始以来的天数”
card1 列是“银行卡的前多少位”
addr1 列是“用户地址代码”
确定了用户的唯一标识之后，我们并不能直接把它当作一个特征直接加入到模型中去，因为通过分析发现，测试集中有68.2%的用户是新用户，并不在训练集中。我们需要间接的使用`UID`，用`UID`构造一些聚合特征。
一个叫做“时间一致性”的有趣技巧是在训练数据集的第一个月使用单个特征（或一小组特征）训练单个模型，并预测isFraud最后一个月的训练数据集。这会评估特征本身是否随时间保持一致。95% 是，但我们发现 5% 的列不符合我们的模型。他们的训练 AUC 约为 0.60，验证 AUC 为 0.40。
主要使用以下五种特征编码方式
频率编码 ：统计该值出现的个数
def encode_FE(df1, df2, cols): for col in cols:
df = pd.concat([df1[col], df2[col]])
vc = df.value_counts(dropna=True, normalize=True).to_dict()
vc[-1] = -1
nm = col + "FE"
df1[nm] = df1[col].map(vc)
df1[nm] = df1[nm].astype("float32")
df2[nm] = df2[col].map(vc)
df2[nm] = df2[nm].astype("float32") print(col)标签编码 ：将原数据映射称为一组顺序数字，类似ONE-HOT,不过 pd.factorize 映射为[1]，[2]，[3]。 pd.get_dummies() 映射为 [1,0,0],[0,1,0],[0,0,1]
def encode_LE(col, train=X_train, test=X_test, verbose=True):
df_comb = pd.concat([train[col], test[col]], axis=0)
df_comb, _ = pd.factorize(df_comb) nm = col
if df_comb.max() > 32000:
train[nm] = df_comb[0: len(train)].astype("float32")
test[nm] = df_comb[len(train):].astype("float32") else:
train[nm] = df_comb[0: len(train)].astype("float16")
test[nm] = df_comb[len(train):].astype("float16")
del df_comb
gc.collect() if verbose: print(col)统计特征：主要使用 pd.groupby对变量进行分组，再使用agg计算分组的统计特征
def encode_AG(main_columns, uids, aggregations=["mean"], df_train=X_train, df_test=X_test, fillna=True, usena=False): for main_column in main_columns:
for col in uids:
for agg_type in aggregations:
new_column = main_column + "_" + col + "_" + agg_type
temp_df = pd.concat([df_train[[col, main_column]], df_test[[col, main_column]]]) if usena:
temp_df.loc[temp_df[main_column] == -1, main_column] = np.nan
#求每个uid下，该col的均值或标准差
temp_df = temp_df.groupby([col])[main_column].agg([agg_type]).reset_index().rename(
columns={agg_type: new_column})
#将uid设成index
temp_df.index = list(temp_df[col])
temp_df = temp_df[new_column].to_dict()
#temp_df是一个映射字典
df_train[new_column] = df_train[col].map(temp_df).astype("float32")
df_test[new_column] = df_test[col].map(temp_df).astype("float32") if fillna:
df_train[new_column].fillna(-1, inplace=True)
df_test[new_column].fillna(-1, inplace=True) print(new_column)交叉特征：对两列的特征重新组合成为新特征，再进行标签编码
def encode_CB(col1, col2, df1=X_train, df2=X_test):
nm = col1 + '_' + col2
df1[nm] = df1[col1].astype(str) + '_' + df1[col2].astype(str)
df2[nm] = df2[col1].astype(str) + '_' + df2[col2].astype(str)
encode_LE(nm, verbose=False) print(nm, ', ', end='')唯一值特征：分组后返回目标属性的唯一值个数
def encode_AG2(main_columns, uids, train_df=X_train, test_df=X_test):
for main_column in main_columns:
for col in uids:
comb = pd.concat([train_df[[col] + [main_column]], test_df[[col] + [main_column]]], axis=0)
mp = comb.groupby(col)[main_column].agg(['nunique'])['nunique'].to_dict()
train_df[col + '_' + main_column + '_ct'] = train_df[col].map(mp).astype('float32')
test_df[col + '_' + main_column + '_ct'] = test_df[col].map(mp).astype('float32')
print(col + '_' + main_column + '_ct, ', end='')# 解压数据集 仅第一次运行时运行!unzip -q -o data/data104475/IEEE_CIS_Fraud_Detection.zip -d /home/aistudio/data
unzip: cannot find or open data/data104475/IEEE_CIS_Fraud_Detection.zip, data/data104475/IEEE_CIS_Fraud_Detection.zip.zip or data/data104475/IEEE_CIS_Fraud_Detection.zip.ZIP.
# 安装依赖包!pip install xgboost
import numpy as np # linear algebraimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)import os, gcfrom sklearn.model_selection import GroupKFoldfrom sklearn.metrics import roc_auc_scoreimport xgboost as xgbimport datetime
path_train_transaction = "./data/raw_data/train_transaction.csv"path_train_id = "./data/raw_data/train_identity.csv"path_test_transaction = "./data/raw_data/test_transaction.csv"path_test_id = "./data/raw_data/test_identity.csv"path_sample_submission = './data/raw_data/sample_submission.csv'path_submission = 'sub_xgb_95.csv'
BUILD95 = FalseBUILD96 = True# cols with stringsstr_type = ['ProductCD', 'card4', 'card6', 'P_emaildomain', 'R_emaildomain', 'M1', 'M2', 'M3', 'M4', 'M5', 'M6', 'M7', 'M8', 'M9', 'id_12', 'id_15', 'id_16', 'id_23', 'id_27', 'id_28', 'id_29', 'id_30', 'id_31', 'id_33', 'id_34', 'id_35', 'id_36', 'id_37', 'id_38', 'DeviceType', 'DeviceInfo']# fisrt 53 columnscols = ['TransactionID', 'TransactionDT', 'TransactionAmt', 'ProductCD', 'card1', 'card2', 'card3', 'card4', 'card5', 'card6', 'addr1', 'addr2', 'dist1', 'dist2', 'P_emaildomain', 'R_emaildomain', 'C1', 'C2', 'C3', 'C4', 'C5', 'C6', 'C7', 'C8', 'C9', 'C10', 'C11', 'C12', 'C13', 'C14', 'D1', 'D2', 'D3', 'D4', 'D5', 'D6', 'D7', 'D8', 'D9', 'D10', 'D11', 'D12', 'D13', 'D14', 'D15', 'M1', 'M2', 'M3', 'M4', 'M5', 'M6', 'M7', 'M8', 'M9']# V COLUMNS TO LOAD DECIDED BY CORRELATION EDA# https://www.kaggle.com/cdeotte/eda-for-columns-v-and-idv = [1, 3, 4, 6, 8, 11]
v += [13, 14, 17, 20, 23, 26, 27, 30]
v += [36, 37, 40, 41, 44, 47, 48]
v += [54, 56, 59, 62, 65, 67, 68, 70]
v += [76, 78, 80, 82, 86, 88, 89, 91]# v += [96, 98, 99, 104] #relates to groups, no NANv += [107, 108, 111, 115, 117, 120, 121, 123] # maybe group, no NANv += [124, 127, 129, 130, 136] # relates to groups, no NAN# LOTS OF NAN BELOWv += [138, 139, 142, 147, 156, 162] # b1v += [165, 160, 166] # b1v += [178, 176, 173, 182] # b2v += [187, 203, 205, 207, 215] # b2v += [169, 171, 175, 180, 185, 188, 198, 210, 209] # b2v += [218, 223, 224, 226, 228, 229, 235] # b3v += [240, 258, 257, 253, 252, 260, 261] # b3v += [264, 266, 267, 274, 277] # b3v += [220, 221, 234, 238, 250, 271] # b3v += [294, 284, 285, 286, 291, 297] # relates to grous, no NANv += [303, 305, 307, 309, 310, 320] # relates to groups, no NANv += [281, 283, 289, 296, 301, 314] # relates to groups, no NAN# v += [332, 325, 335, 338] # b4 lots NANcols += ['V' + str(x) for x in v]
dtypes = {}for c in cols + ['id_0' + str(x) for x in range(1, 10)] + ['id_' + str(x) for x in range(10, 34)]:
dtypes[c] = 'float32'for c in str_type:
dtypes[c] = 'category'# load data and mergeprint("load data...")
X_train = pd.read_csv(path_train_transaction, index_col="TransactionID", dtype=dtypes, usecols=cols + ["isFraud"])
train_id = pd.read_csv(path_train_id, index_col="TransactionID", dtype=dtypes)
X_train = X_train.merge(train_id, how="left", left_index=True, right_index=True)
X_test = pd.read_csv(path_test_transaction, index_col="TransactionID", dtype=dtypes, usecols=cols)
test_id = pd.read_csv(path_test_id, index_col="TransactionID", dtype=dtypes)
X_test = X_test.merge(test_id, how="left", left_index=True, right_index=True)# targety_train = X_train["isFraud"]del train_id, test_id, X_train["isFraud"]print("X_train shape:{}, X_test shape:{}".format(X_train.shape, X_test.shape))load data... X_train shape:(590540, 213), X_test shape:(506691, 213)
# transform D feature "time delta" as "time point"for i in range(1, 16): if i in [1, 2, 3, 5, 9]: continue
X_train["D" + str(i)] = X_train["D" + str(i)] - X_train["TransactionDT"] / np.float32(60 * 60 * 24)
X_test["D" + str(i)] = X_test["D" + str(i)] - X_test["TransactionDT"] / np.float32(60 * 60 * 24)# encoding function# frequency encodedef encode_FE(df1, df2, cols):
for col in cols:
df = pd.concat([df1[col], df2[col]])
vc = df.value_counts(dropna=True, normalize=True).to_dict()
vc[-1] = -1
nm = col + "FE"
df1[nm] = df1[col].map(vc)
df1[nm] = df1[nm].astype("float32")
df2[nm] = df2[col].map(vc)
df2[nm] = df2[nm].astype("float32") print(col)# label encodedef encode_LE(col, train=X_train, test=X_test, verbose=True):
df_comb = pd.concat([train[col], test[col]], axis=0)
df_comb, _ = pd.factorize(df_comb)
nm = col if df_comb.max() > 32000:
train[nm] = df_comb[0: len(train)].astype("float32")
test[nm] = df_comb[len(train):].astype("float32") else:
train[nm] = df_comb[0: len(train)].astype("float16")
test[nm] = df_comb[len(train):].astype("float16") del df_comb
gc.collect() if verbose: print(col)def encode_AG(main_columns, uids, aggregations=["mean"], df_train=X_train, df_test=X_test, fillna=True, usena=False):
for main_column in main_columns: for col in uids: for agg_type in aggregations:
new_column = main_column + "_" + col + "_" + agg_type
temp_df = pd.concat([df_train[[col, main_column]], df_test[[col, main_column]]]) if usena:
temp_df.loc[temp_df[main_column] == -1, main_column] = np.nan #求每个uid下，该col的均值或标准差
temp_df = temp_df.groupby([col])[main_column].agg([agg_type]).reset_index().rename(
columns={agg_type: new_column}) #将uid设成index
temp_df.index = list(temp_df[col])
temp_df = temp_df[new_column].to_dict() #temp_df是一个映射字典
df_train[new_column] = df_train[col].map(temp_df).astype("float32")
df_test[new_column] = df_test[col].map(temp_df).astype("float32") if fillna:
df_train[new_column].fillna(-1, inplace=True)
df_test[new_column].fillna(-1, inplace=True) print(new_column)# COMBINE FEATURES交叉特征def encode_CB(col1, col2, df1=X_train, df2=X_test):
nm = col1 + '_' + col2
df1[nm] = df1[col1].astype(str) + '_' + df1[col2].astype(str)
df2[nm] = df2[col1].astype(str) + '_' + df2[col2].astype(str)
encode_LE(nm, verbose=False) print(nm, ', ', end='')# GROUP AGGREGATION NUNIQUEdef encode_AG2(main_columns, uids, train_df=X_train, test_df=X_test):
for main_column in main_columns: for col in uids:
comb = pd.concat([train_df[[col] + [main_column]], test_df[[col] + [main_column]]], axis=0)
mp = comb.groupby(col)[main_column].agg(['nunique'])['nunique'].to_dict()
train_df[col + '_' + main_column + '_ct'] = train_df[col].map(mp).astype('float32')
test_df[col + '_' + main_column + '_ct'] = test_df[col].map(mp).astype('float32') print(col + '_' + main_column + '_ct, ', end='')print("encode cols...")# TRANSACTION AMT CENTSX_train['cents'] = (X_train['TransactionAmt'] - np.floor(X_train['TransactionAmt'])).astype('float32')
X_test['cents'] = (X_test['TransactionAmt'] - np.floor(X_test['TransactionAmt'])).astype('float32')print('cents, ', end='')encode cols... cents,
# FREQUENCY ENCODE: ADDR1, CARD1, CARD2, CARD3, P_EMAILDOMAINencode_FE(X_train, X_test, ['addr1', 'card1', 'card2', 'card3', 'P_emaildomain'])# COMBINE COLUMNS CARD1+ADDR1, CARD1+ADDR1+P_EMAILDOMAINencode_CB('card1', 'addr1')
encode_CB('card1_addr1', 'P_emaildomain')# FREQUENCY ENOCDEencode_FE(X_train, X_test, ['card1_addr1', 'card1_addr1_P_emaildomain'])# GROUP AGGREGATEencode_AG(['TransactionAmt', 'D9', 'D11'], ['card1', 'card1_addr1', 'card1_addr1_P_emaildomain'], ['mean', 'std'],
usena=False)for col in str_type:
encode_LE(col, X_train, X_test)"""
Feature Selection - Time Consistency
We added 28 new feature above. We have already removed 219 V Columns from correlation analysis done here.
So we currently have 242 features now. We will now check each of our 242 for "time consistency".
We will build 242 models. Each model will be trained on the first month of the training data and will only use one feature.
We will then predict the last month of the training data. We want both training AUC and validation AUC to be above AUC = 0.5.
It turns out that 19 features fail this test so we will remove them.
Additionally we will remove 7 D columns that are mostly NAN. More techniques for feature selection are listed here
"""cols = list(X_train.columns)
cols.remove('TransactionDT')for c in ['D6', 'D7', 'D8', 'D9', 'D12', 'D13', 'D14']:
cols.remove(c)# FAILED TIME CONSISTENCY TESTfor c in ['C3', 'M5', 'id_08', 'id_33']:
cols.remove(c)for c in ['card4', 'id_07', 'id_14', 'id_21', 'id_30', 'id_32', 'id_34']:
cols.remove(c)for c in ['id_' + str(x) for x in range(22, 28)]:
cols.remove(c)print('NOW USING THE FOLLOWING', len(cols), 'FEATURES.')# CHRIS - TRAIN 75% PREDICT 25%idxT = X_train.index[:3 * len(X_train) // 4]
idxV = X_train.index[3 * len(X_train) // 4:]print(X_train.info())# X_train = X_train.convert_objects(convert_numeric=True)# X_test = X_test.convert_objects(convert_numeric=True)for col in str_type: print(col)
X_train[col] = X_train[col].astype(int)
X_test[col] = X_test[col].astype(int)print("after transform:")print(X_train.info())# fillnafor col in cols:
X_train[col].fillna(-1, inplace=True)
X_test[col].fillna(-1, inplace=True)START_DATE = datetime.datetime.strptime('2017-11-30', '%Y-%m-%d')
X_train['DT_M'] = X_train['TransactionDT'].apply(lambda x: (START_DATE + datetime.timedelta(seconds=x)))
X_train['DT_M'] = (X_train['DT_M'].dt.year - 2017) * 12 + X_train['DT_M'].dt.month
X_test['DT_M'] = X_test['TransactionDT'].apply(lambda x: (START_DATE + datetime.timedelta(seconds=x)))
X_test['DT_M'] = (X_test['DT_M'].dt.year - 2017) * 12 + X_test['DT_M'].dt.monthprint("training...")if BUILD95:
oof = np.zeros(len(X_train))
preds = np.zeros(len(X_test))
skf = GroupKFold(n_splits=6) for i, (idxT, idxV) in enumerate(skf.split(X_train, y_train, groups=X_train['DT_M'])):
month = X_train.iloc[idxV]['DT_M'].iloc[0] print('Fold', i, 'withholding month', month) print(' rows of train =', len(idxT), 'rows of holdout =', len(idxV))
clf = xgb.XGBClassifier(
n_estimators=5000,
max_depth=12,
learning_rate=0.02,
subsample=0.8,
colsample_bytree=0.4,
missing=-1,
eval_metric='auc', # USE CPU
# nthread=4,
# tree_method='hist'
# USE GPU
tree_method='gpu_hist'
)
h = clf.fit(X_train[cols].iloc[idxT], y_train.iloc[idxT],
eval_set=[(X_train[cols].iloc[idxV], y_train.iloc[idxV])],
verbose=100, early_stopping_rounds=200)
oof[idxV] += clf.predict_proba(X_train[cols].iloc[idxV])[:, 1]
preds += clf.predict_proba(X_test[cols])[:, 1] / skf.n_splits del h, clf
x = gc.collect() print('#' * 20) print('XGB95 OOF CV=', roc_auc_score(y_train, oof))if BUILD95:
sample_submission = pd.read_csv(path_sample_submission)
sample_submission.isFraud = preds
sample_submission.to_csv(path_submission, index=False)
X_train['day'] = X_train.TransactionDT / (24 * 60 * 60)
X_train['uid'] = X_train.card1_addr1.astype(str) + '_' + np.floor(X_train.day - X_train.D1).astype(str)
X_test['day'] = X_test.TransactionDT / (24 * 60 * 60)
X_test['uid'] = X_test.card1_addr1.astype(str) + '_' + np.floor(X_test.day - X_test.D1).astype(str)# FREQUENCY ENCODE UIDencode_FE(X_train, X_test, ['uid'])# AGGREGATEencode_AG(['TransactionAmt', 'D4', 'D9', 'D10', 'D15'], ['uid'], ['mean', 'std'], fillna=True, usena=True)# AGGREGATEencode_AG(['C' + str(x) for x in range(1, 15) if x != 3], ['uid'], ['mean'], X_train, X_test, fillna=True, usena=True)# AGGREGATEencode_AG(['M' + str(x) for x in range(1, 10)], ['uid'], ['mean'], fillna=True, usena=True)# AGGREGATEencode_AG2(['P_emaildomain', 'dist1', 'DT_M', 'id_02', 'cents'], ['uid'], train_df=X_train, test_df=X_test)# AGGREGATEencode_AG(['C14'], ['uid'], ['std'], X_train, X_test, fillna=True, usena=True)# AGGREGATEencode_AG2(['C13', 'V314'], ['uid'], train_df=X_train, test_df=X_test)# AGGREATEencode_AG2(['V127', 'V136', 'V309', 'V307', 'V320'], ['uid'], train_df=X_train, test_df=X_test)# NEW FEATUREX_train['outsider15'] = (np.abs(X_train.D1 - X_train.D15) > 3).astype('int8')
X_test['outsider15'] = (np.abs(X_test.D1 - X_test.D15) > 3).astype('int8')print('outsider15')
cols = list(X_train.columns)
cols.remove('TransactionDT')for c in ['D6', 'D7', 'D8', 'D9', 'D12', 'D13', 'D14']: if c in cols:
cols.remove(c)for c in ['oof', 'DT_M', 'day', 'uid']: if c in cols:
cols.remove(c)# FAILED TIME CONSISTENCY TESTfor c in ['C3', 'M5', 'id_08', 'id_33']: if c in cols:
cols.remove(c)for c in ['card4', 'id_07', 'id_14', 'id_21', 'id_30', 'id_32', 'id_34']: if c in cols:
cols.remove(c)for c in ['id_' + str(x) for x in range(22, 28)]: if c in cols:
cols.remove(c)print('NOW USING THE FOLLOWING', len(cols), 'FEATURES.')print(np.array(cols))if BUILD96:
oof = np.zeros(len(X_train))
preds = np.zeros(len(X_test))
skf = GroupKFold(n_splits=6) for i, (idxT, idxV) in enumerate(skf.split(X_train, y_train, groups=X_train['DT_M'])):
month = X_train.iloc[idxV]['DT_M'].iloc[0] print('Fold', i, 'withholding month', month) print(' rows of train =', len(idxT), 'rows of holdout =', len(idxV))
clf = xgb.XGBClassifier(
n_estimators=5000,
max_depth=12,
learning_rate=0.02,
subsample=0.8,
colsample_bytree=0.4,
missing=-1,
eval_metric='auc', # USE CPU
# nthread=4,
# tree_method='hist'
# USE GPU
tree_method='gpu_hist'
)
h = clf.fit(X_train[cols].iloc[idxT], y_train.iloc[idxT],
eval_set=[(X_train[cols].iloc[idxV], y_train.iloc[idxV])],
verbose=100, early_stopping_rounds=200)
oof[idxV] += clf.predict_proba(X_train[cols].iloc[idxV])[:, 1]
preds += clf.predict_proba(X_test[cols])[:, 1] / skf.n_splits del h, clf
x = gc.collect() print('#' * 20) print('XGB96 OOF CV=', roc_auc_score(y_train, oof))if BUILD96:
sample_submission = pd.read_csv(path_sample_submission)
sample_submission.isFraud = preds
sample_submission.to_csv(path_submission, index=False)本项目主要对IEEE-CIS Fraud Detection相关资料进行了收集汇总，目的是学习特征的构建。
| 数据集 | IEEE-CIS Fraud Detection |
|---|---|
| 线上评分 | 0.959221 |
以上就是【金融风控系列】_[2]_欺诈识别的详细内容，更多请关注php中文网其它相关文章！
每个人都需要一台速度更快、更稳定的 PC。随着时间的推移，垃圾文件、旧注册表数据和不必要的后台进程会占用资源并降低性能。幸运的是，许多工具可以让 Windows 保持平稳运行。
广告Copyright 2014-2025 https://www.php.cn/ All Rights Reserved | php.cn | 湘ICP备2023035733号
641
收藏
