It's the result of floating point rounding in the two different ways you generated the datenum values --
In
full_data_time_steps=begin_date_time:interval:end_date_time;
the interval value is an approximation and over the vector it will accumulate noticeable differences.
The solution is to always use integer values of whatever granularity is needed to generate sequences of dates such that the internal calculations are reproducible.
Presuming you're using ML release R2014b or later use datetime instead of datenum; there are many features that make it more flexible.
