If we have a dependent variable like yes or no, and independent variables which are continuous, how important is it to transform the independent variables into categorical? Is it something which always improve the accuracy of the machine learning algorithms? I have seen some papers doing these steps.
No, it does not guarantee improved accuracy. In fact, it may significantly harm accuracy if the discretisation process removes distinguishing features. Having said that, it can help combat overfitting.