0
$\begingroup$

This is a hypothetical situation.

Let's say you have access to a lot of human behaviors and characteristics (features). Let's say you have a sample of 10000 humans. You know that within this sample, there are some criminals - except you don't know how many of them there are and what distinctive features criminals possess. However, you know for sure that all criminals are similar in some way.

What kind of unsupervised technique could I use to group my observations into basically groups of potential criminals vs. noise. Obviously, after the algo groups the people, I will have to examine the groups myself to see if it makes sense they are actually criminals.

Kmeans doesn't do what I want since you have to specify how many groups there are before hand, and it puts everyone into groups. I only care about groups that may be criminals, the rest I don't care about. So I am entirely ok if the algorithm does not group some observations. I guess with Kmeans, you could try different numbers for K, examine each cluster, repeat until you find something that makes sense? Would that be a good approach?

How could I approach such a problem and what other pieces of information do I need?

Improve this question
$\endgroup$
4
  • 1
    $\begingroup$ Any solution that works would identify any subgroup of humans as "criminals," because "criminals" has no definition or meaning: it's just an abstract term. If this isn't obvious, suppose the same sample has ten criminals and ten saints. What is to prevent the solution from identifying all saints as having common characteristics and labeling them as "criminals"? Now imagine how many different kinds of subgroups that could appear in such a sample, based on all different kinds of shared characteristics, and you should see why this is not an answerable question. $\endgroup$
    – whuber ♦
    Feb 7, 2020 at 20:17
  • $\begingroup$ Well I figured I would have to examine the groups regardless and see if it makes sense - but the point is examining it after the fact. I'm new at this and haven't applied my knowledge to any real world problem yet. But there is one I would like to try and it's similar to this in which going into a data set, I want to be unbiased, even though once the clusters are formed, it'll be easy for me to determine which ones are "criminals" and which ones likely are not - based on some prior knowledge I have about what criminals might be like. $\endgroup$
    – confused
    Feb 7, 2020 at 20:30
  • $\begingroup$ But if it's impossible, I'll have to think of how to solve the problem in another way. Maybe for each observation, I can find the most similar observations to it. And then do something with that. $\endgroup$
    – confused
    Feb 7, 2020 at 20:37
  • $\begingroup$ It might help to look at how real people would do this. Suppose you were to give these data to an intelligent, knowledgeable friend. You still want them to identify criminals for you, but--to put them in a position comparable to the software, which knows no English--your instructions are "within this sample, there are some murgatroyds. I don't know how many of them there are and what distinctive features murgatroyds possess. However, I know that all murgatroyds are similar in some way. Please identify the murgatroyds in this sample." Do think they could?? $\endgroup$
    – whuber ♦
    Feb 7, 2020 at 20:49

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.