If you think of the reduced cost as a signal to you to move resources to an activity that is not currently maxed out, (there may be more than one), it makes sense to make the move to the entry that gives the biggest improvement (the largest reduced cost). One thing that can cause confusion though is that different packages have different sign conventions for some of these concepts -- each internally consistent, but you have to stick to one way of doing it. There is a good discussion using the LINDO in the book by Schrage. (Also see the AMPL book.) Let me know if you need further references to find these.

Thank you for the answer, Prof. Yes I understand if the terminology of the reduced cost is similar in simplex method. However, the reduced cost here (in the paper I attached related to zero point method) means the different between original cost and the final cost after some steps (which its figure is zero, so the largest reduced cost is largest unit transportation cost). It means  we allocate from the largest cost first that I do not understand which what we want to do is minimize the overall cost.