Create a discrete variable from continuous one Have any sense, in order to solve a classification problem (only 2 possibles target classes) to create a new discrete column from a continuous one? 
For example, I have one column that is Age, that has integer numbers. I have created Age_Over variable with 2 possible values 1 or 0.
Is it useful for use it in a machine learning problem maybe with certain models or just a a bad idea?

Let me explain more. I have data from workers of a company. I'm trying to solve a classification problem to predict the employee attrition.
For example I'm tempted to do this with one variable called YearsAtCompany. I have calculated the median of this column and created a new column with 1 if the value is grater than the median and a 0 otherwise.
I have noticed that the rate of attrition in the employees with less years at company than the median is significantly greater than employees with more years than the median.
What do you think?
 A: Unless the variable in question is strongly bimodal, I would suggest against dichotomising it using some arbitrary threshold.
Some of the obvious issues of dichotomising is that the estimated values will have reduced precision as well as that the dichotomisation assumes that the relationship between the dichotomised variable and other variables is constant within dichotomisation intervals (an obviously strong assumption that usually is unrealistic).
There is a significant body of literature presenting problems deriving from such dubious discretisation approaches. See for example: Ragland (1992) Dichotomizing continuous outcome variables: Dependence of the magnitude of association and statistical power on the cutpoint, Buettner et a. 1997 Problems in defining cutoff points of continuous prognostic factors: Example of tumor thickness in primary cutaneous melanoma and Altman et al. (1994) Dangers of using 'optimal' cutpoints in the evaluation of prognostic factors. F. Harrell's RMS book has the aptly name subsection 2.4.1: "Avoiding Categorization".
It could potentially be OK to partially dichotomising a variable into Low-/Medium-/High- value ranges and use only the Low- and High- value ranges if the variable in question is really noisy and contains a high degree of error but this is an edge case and should rely on particular domain knowledge. M. Kuhn discusses this shortly in his APM book in Chapt. 20.4 "Discretizing Continuous Outcomes".
