[英]is there a faster way for Training IsolationForest on each row of pandas DataFrame?
I have a Pandas DataFrame which contain transaction count of 2000 terminals for 30 days(columns are day of month) DataFrame looks like this: 我有一个Pandas DataFrame,其中包含30天的2000个终端的交易计数(列是一个月中的一天)DataFrame看起来像这样:
trx.head()
TerminalID 8881 8882 8883 8884 8885 8886 ...
0 11546 0.0 0.0 0.0 0.0 0.0 0.0 ...
1 200002 0.0 0.0 0.0 0.0 0.0 0.0 ...
2 200512 1.0 0.0 0.0 1.0 1.0 0.0 ...
3 202630 3.0 1.0 1.0 0.0 1.0 1.0 ...
4 207000 2.0 4.0 1.0 6.0 3.0 7.0 ...
I want to use IsolationForest for anomaly detection on each row of my data. 我想使用IsolationForest对数据的每一行进行异常检测。
First I convert each row to a new DataFrame and fit data on that, one by one for every row and I will add the the result to a list: 首先,我将每一行转换为新的DataFrame并在其上拟合数据,每一行一个接一个,然后将结果添加到列表中:
def find_anomaly(trx1,outliers_fraction):
scaler = StandardScaler()
np_scaled = scaler.fit_transform(trx1)
data = pd.DataFrame(np_scaled)
# train isolation forest
model = IsolationForest(contamination=outliers_fraction)
model.fit(data)
trx1['anomaly'] = pd.Series(model.predict(data))
return(trx1)
#This for is slow
list_terminal_trx = []
for i in range(0,len(trx)-1):
trx1=trx.iloc[i,1:].reset_index()
trx1.columns=['day','count']
trx1['day']=trx1['day'].astype(float)
list_terminal_trx.append(find_anomaly(trx1,outliers_fraction))
print('Learning for record',i)
The code above works fine but it is slow I wanted to know if there is a better way? 上面的代码工作正常,但是很慢,我想知道是否有更好的方法?
Edited1: thanks to @AT_asks advise I set n_jobs=-1 and now It is faster But is there any alternative to my for loop? Edited1:感谢@AT_asks建议我将n_jobs = -1设置为现在,它更快,但是我的for循环还有其他选择吗?
Edited2: with some modification I used what @AT_asks suggested to use apply() but I got no performance differences: For version takes 3:29:00 Apply Version Takes 3:25:28 Edited2:经过一些修改,我使用了@AT_asks建议使用apply()的方法,但没有任何性能差异:对于版本需要3:29:00 Apply版本需要3:25:28
Edited3: using iterrows() instead of for brings the same result: 3min 16s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each) Edited3:使用iterrows()代替for会带来相同的结果:每个循环3分钟16s±0 ns(平均±标准偏差,运行1次,每个循环1次)
You might get some improvements if add this parameter 如果添加此参数,您可能会得到一些改进
model = IsolationForest(contamination=outliers_fraction, n_jobs=-1)
Also, we could try this. 另外,我们可以尝试一下。
# Do not create instance every time
scaler = StandardScaler()
def find_anomaly(trx1,outliers_fraction):
np_scaled = scaler.fit_transform(trx1)
data = pd.DataFrame(np_scaled)
# train isolation forest
model = IsolationForest(contamination=outliers_fraction, n_jobs=-1)
model.fit(data)
trx1['anomaly'] = pd.Series(model.predict(data))
return(trx1)
# not loop but apply
list_terminal_trx = trx.apply(lambda x: find_anomaly(x,outliers_fraction), axis =1).values
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.