It's going to depend on your application.  Are outliers erroneous data or true extreme values (such as Bill Gates getting included in a random survey of annual income)?  Do you know values have to be in a particular range? Are you better off dropping outliers or imputing more realistic data?

It's going to depend on your application.  Are outliers erroneous data or true extreme values (such as Bill Gates getting included in a random survey of annual income)?  Do you know values have to be in a particular range? Are you better off dropping outliers or imputing more realistic data?

@Mitchell Maltenfort  : I have filtered data considering various factors. So i want to find outliers from true data. But sometimes density of my data is more at one region, so need to find out exceptional cases.This is performed on realistic data for the traffic analysis. so density for the samples gets divided into many regions. from that i need to find outliers.

I generally use boxplot, but you can also use outliers package in r which contains many statistical test for detecting outliers. In addition, you might find this helpful

http://www.dbs.ifi.lmu.de/~zimek/publications/KDD2010/kdd10-outlier-tutorial.pdf

@Yuanhang Liu: I tried using boxplot, but boxplot produces good results in calse of unimodal model and i am having multi modal density model. in this case boxplot looses our some actual data with outliers which should not happen. I need the way to find solution for such cases.

Thanks for your suggestion, it will definitely going to help me in my problem.

Mitchell Maltenfort is correct in his statement, and his questions will provide a useful guide to progress.

In a very general way, outliers are outliers with respect to "something." The discussions I have read about outliers seem to avoid dealing with the fact that there are two steps in outlier detection: defining what it means to be an outlier and then identifying the outliers.  The latter is easy. The former is hard because it means one has to commit, not because it is technically difficult. Can anyone define "outlier" in an absolute manner? I mean beyond "a value that is highly unlikely given an assumed distribution." What is the mental model behind stating that an observation is an outlier?? Is there a single model for all situations? Probably not. My concern is very practical because honest people looking for answers in the internet will come across "shooting from the hip" answers that sound pretty technical and convincing but that actually should not be considered for their specific problem. Certainly, a random process that generates uniformly distributed data will have a lot of outliers if one is thinking "normal."

Here a simple explanation of a classic method;

https://www.youtube.com/watch?v=bRdC1u9veg8

Article Fine-Tuning Some Resistant Rules for Outlier Labeling

https://en.wikibooks.org/wiki/Data_Mining_Algorithms_In_R/Classification/Outliers

This helps me a lot, maybe it can also help you. Thanks.

Hi,

Library called "AnomalyDetection" from Twitter will help identify local as well as global anomalies and outliers. Might help.

Link: https://github.com/twitter/AnomalyDetection

Mayur Narkhede, I'll second Lakmal Meegahapola and recommend looking into an anomaly detection method. But, to be more specific, you might look into anomaly detection in time series. There are a few packages out there that can be helpful. I do think, however, you'll need to do some work for "real-time" (e.g., scheduled and streaming data). You'll likely need to think about how you might bound the time series—"moving anomaly detection" based on time or n.

Best of luck on this interesting analysis.