ML method to identify data subset with lower average values? I have a task where I need to find anomalous groups of observations [y,X] by the average value of y, using the feature set X.
A good interpretation is that these are customers and I want to find the most valuable ones.
(Note, this is actually separate from a classification problem, which I would use decision trees.) 
This lead me to "Subgroup Discovery", but this is a field I'm unfamiliar with and I don't know the mainstream methods in it. For example, regressions, decision trees and NN are methods in classification methods, but I don't know the standard methods for SD. 
Can someone confirm subgroup discovery is what I want and give point me in the right direction with some concepts (maybe just search keywords) of common methods used in this field for my problem?
 A: Generally this type of problem can be viewed in two ways:


*

*as an anomaly detection problem (as suggested by 7kemZmani), if we can assume that there is some kind of pattern that we expect the overall population to follow and we are interested in identifying the instances that deviate noticably from this pattern. 

*as a clustering problem, if we expect multiple subgroups in the population we're examining and want to identify their location and members. 


Now based on these two views, I suppose it's up to you to identify which fits your problem better.
If you are not sure that your data is the result of a single underlying process, I would suggest you experiment with clustering approaches and on examination of the results decide further steps. 
In particular it would be interesting to have a look at how clustering methods that don't require a prior assumption of the number of groups (like k-means does) cluster your data and with how many groups these end up. For this there's multiple options, right now DBSCAN, hierarchical clustering and (a bit far out for this but I've worked with it before) 'Bayesian Gaussian Mixture Models' come to my mind.  
These three methods are examples of methods that will present you a grouping of your data into an unspecified number of groups. Examine the results you get from this, with hierarchical clustering for example you can visualize the result of the clustering with a dendrogram to examine which clustering is closer to the actual structure. With a bayesian GMM, you can have a look at the resulting mixture weights of the model.
If in these two examples this examination points towards a single cluster describing the data best, then you should be safe to assume that there actually is only a single process generating your data.
Then, applying anomaly detection makes sense and allows you to identify outliers from the 'expected outcome'.   

Edit:
While I am not familiar with subgroup discovery, I can point you towards a nice survey article on the topic. This should give you a good introduction to the topic and point you towards common techniques and methods for such problems.   
