How to cluster users in recommendation systems use to detect attacks?

Good day! Interested in the question of how to identify suspect accounts with clustering users to their scores (low or high) are not counted in Recommender systems.
Enough information to separate users according to their interests, but can not find how to understand these are real people or this attack (for example, someone wanted to understate the product of its competitor, and its increase).
Share your thoughts or articles.
If you do not understand the question, can you reformulate.
Thank you
April 19th 20 at 12:22
3 answers
April 19th 20 at 12:24
the standard deviation
the upper and lower quartiles

conventional statistical methods, well, just a scrap on unrealistic parameters, like growth of 2.5 meters, to comment in a minute after the publication time of the movie and so on
April 19th 20 at 12:26
The easiest way to identify the average ratio of the number of likes to the number of hits for each unique.
And make allowance width in the median "corridor" 50% of all unique users.
All who find themselves outside this "corridor" - are cheaters.
April 19th 20 at 12:28
Share your thoughts or articles.
If you do not understand the question, can you reformulate.

Yes, there is a sort of intelligence gathered so that you can not reformulate. But if you want to - of course we can. And while you will answer briefly restate what you know.
Personally, this task is not doing, but know people who do it professionally, ie, for serious customers. So no you open your results will not tell - as soon as such information becomes open - instantly are particularly mentally gifted, who will try to circumvent the protection. Who cares?
And so, the analysis is carried out by conventional methods from the field of Fraud Detection. Such methods and tools a lot of books on this account I write. But it's "gentle introduction" to the topic, the distant approaches to real case studies. Well as in the banking systems of all heard about the methods of catching fraudulent transactions, about which I write. It seems like here it is, the information is available to all - deviations, search anomalies, 3 Sigma, spatial gaps, etc. - and how really it functions in the real banks - alas "know how" and a mystery.

Find more questions by tags Recommender systemsClustering