If there is a conflict among literature review about which of the variables is dependent and which is independent, is there a method to solve this dilemma, taking into consideration there is no mediating or moderating variable?
Hello, Maha! I believe you are dealing with a multivariate calibration problem. Try using some variable selection technique. There are several tools in the literature for that. I suggest the use of Successive Projections Algorihtm (SPA), Partial Least Squares (PLS), Genetic Algorithms (GA), or Firefly Algorithm (FA). In my profile, you are able to find some useful papers about this issue.
I don't personally think that a machine can become so intelligent from "learning" that it can differentiate between a cause and an effect.
However, what a machine can actually learn, is: by fixing one (say the effect) it can figure out which variable can possibly be the other (i.e. cause), and which cannot be. Mathematically speaking, we fix one variable as the effect (say y), and ask our machine to find out (or learn) which of x1, x2, etc. can be the possible cause of y. ...One can also do the opposite-- i.e. fix a cause, and ask the machine to comment on what can be its possible effects.
For example, take the 2 sets {bacteria, worms, germs, smoking} and {diseases, damage, contamination, cancer}. It is evident for a human to see which is the cause and which is the effect, but what a machine learning algorithm can at most do (to my knowledge), is to find the degree of dependence of a certain disease on a certain foreign organism or human action.
So, the bottom line is: It's my opinion that determining which of the correlated variables is the cause, and which is the effect-- is a matter of human judgement... and only domain knowledge can possibly indicate the differentiation between a cause and an effect.