i am using cluster method to detect outlier in stream data and i want to calculate
detection rate =
Number of Outlier Detected/the total number of outliers in data
if i use data set like" Iris data set" how i know the real total number of outliers in data .
You will have to calculate the detection based on available data points. For example, in first 100 samples, 10 outliers are detected. Keep on calculating for next 100 samples. This is the way, it will help you to compute the estimated detection rate.
The detailed Example:
Initial 100 samples, 10 outliers
Next 100 samples, 14 outliers
Next 100 samples, 15 outliers
Next 100 samples, 9 outliers
Next 100 samples, 12 outliers
Calculation: In 500 samples, there are 60 outliers.
Detection rate: 60/500 * 100 = 12%
The more the number of experiments, the better estimation would be calculated.
https://ac.els-cdn.com/S2212017312005713/1-s2.0-S2212017312005713-main.pdf?_tid=cd53a1c5-485e-44af-9d8c-cefebf2bd126&acdnat=1532422799_6d59d97c02b4e232351abe5373e3c7bf
Article Outliers detection and classification in wireless sensor networks
I suggest you perform a Bootstrap aggregating (bagging) to ensure your detection is a "true" outliers and not the result of the sampling method.
thanks all for your answers
the problem that i need to know
"real total number of outliers in data set "
for example if if my method get 10 outlier and real number outlier 20 so 10/20 is my method detection rate so i need to know the real number of outlier
do i need to use data set from outlier benchmarks or generate synthetic data sets that i know total number ?
This simple formula might help you .
Rate of Detection = (NTP/ NTP+NFP)
where NTP is Number of True Positive for you its number of detected outliers .
NFP is Number of False Positive for you its samples which are not outliers.
- Badges
- Science topic
- Similar topics
- Computer Science
- Data Mining