Assigning manually a score based on some statistics?

Question

I am trying to assign manually scores to identify bad accounts (not in Stacks community!) based on some specific conditions, for example if they have account names containing only numbers and a limited number of review, a past history poor, poor score, banned, and so on.

To do it, as I mentioned, I am doing the following:

-if number of reviews is < 2 then assign -1;
or
-if past history is poor then assign -1;
or 
-if account score is less than 3 then assign -1;

and so on.

Since I have variables such as number of reviews, past history, account score, for all the users in my dataset, I am just wondering it I should assign a score in a different way, maybe using a more statistical approach based on average. My goal is to determine an algorithm which can predict if an account is bad or not thought time based on the above conditions.

I would appreciate it if you could let me know what you think.

Thanks

Answer 1

If you use a ML model to learn a score that you built based on your features the best thing that could happen is that your model learn exactly the rules you applied on your features. It would not be usefull as it would be equivalent to applying your rules in the first place. Supervised ML is only usefull to learn unknown rules.

Basically you know have two approaches :

• Don't use ML and keep your rules as is. Even if it wouldn't be the most predictive model it might already be usefull to observe unusual behaviors.
• If you want to go the ML way you need to build relevant target. Generally speaking that means having a solid, undiscutable definition of your target. In your case the target could be an action from the moderation. Then ML could help you find the behavior that leads to moderator intervention.