Introduction to Anomaly Detection and One Class Classification

Key Takeaways

what is ood or what is anomaly detection so anomaly detection outlier detection there are so many different names for for something that might belong to one field and it sometimes gets confusing as to what somebody is talking about or what is what is what right like you you have this sort of confusion right there is out of distribution detection which is od anomaly detection outlier detection novelty detection one class classification so much so that you don't know who's the real od right so this confusion exists and even while i was going through the uh through the research and articles uh i i also came across this so i'm trying to just uh put these into different brackets so that it starts to make sense so the first one is out of distribution or anomaly detection they both are used interchangeably in the literature so ore detection deals with detecting whether a test sample is from in distribution that is the training distribution by a classifier or out of distribution if sufficiently different from it what i have found is that out of distribution is like or the anomaly detection problem is like the umbrella term everybody when they want to talk about this they use these terms to represent the entire field right so it is about detecting whether test sample is from in distribution or out of distribution so that is like the high level uh term that is used to capture this field but then if you want to go further in depth in these fields the there are other two definitions which we will see now there's outlier detection and there is novelty detection which you will see in the next uh in the next part so outlier detection is basically an outlier is an observation that deviates so much from other observations as to arouse suspicion that it was generated by a different mechanism and outlier detection deals with detecting of such observations so you have a data set but the data set is not uniform in terms of the distribution of all your data points and if there is a data point that does that to to to an extent does not resemble the other data points so much that you have a suspicion that hey this guy might not belong to this distribution itself might not belong to this data set itself and that is an outlier and the detection of such data points is called outlier detection you have a data set you have a data point in that and now you're like okay does this guy belong to this uh this data distribution or does this guy belong to this data set that is the question you're trying to answer in outlier detection but what you're trying to answer in novelty detection is the identification of new or unknown in distribution data that a machine learning system is not aware of during training so a new data point comes to your uh data set that you have and now you have to say whether this guy is an outlier or not there's a subtle difference between outlier detection and novelty detection outlier detection is when you have the data set already and then you're trying to find the guy who's different from your given general data set but when you what in terms of novelty detection you get a new data point and that new data point you have to make a decision for that new data point saying whether that guy belongs to this data set or not and that is the difference between outlier and power it is okay so this is the essence of different and one class classification i didn't talk about that here is basically you have uh if you have this typical supervised learning machine learning setting where you are trying to classify something uh i'm not going to go with my step like for example you get an email right your email you say whether it is spam or not a spam email so you try and put it into different classes one class classification what it does it says i am going to classify i have data points only on one class like you have emails only on say spam and now you want to say that whether this guy is uh spam or not you don't know about the other class there can be all sorts of other classes so you try and classify with you don't you're not going to say what are the other classes but you will say whether this is spam right so you will say i can for sure say this is spam i cannot say if it is something else so that is one class classification we will see a little bit in depth in one of the techniques but this is just the definition of each of these things okay moving on in short it's about finding things that don't fit a pattern you talk about od anomaly detection novelty detection outlier detection everything converges to this one sentence which says we have some things that fit a pattern and there are some points that will not fit a pattern and it's about finding those data points as simple as that regardless of all the different names that are given regardless of all the different uh techniques that are there what does the problem of anomaly detection or outlier detection or one class classification or all of this uh deal with is it's about finding things that don't fit a pattern or rather in case of one class fascinating about finding things that actually fit one pattern not other patterns but in general outlier detection is about novelty detection is about you have a given pattern i want to find a data point uh if i want to determine if a data point fits this pattern or not and and that's that's what it is all about