Partitioning a matrix intro train and test in MATLAB in an efficient manner -

i working user-rating matrix r real data many users u in rows , many items i in columns ratings r(u,i) rating of user u item i. ratings 1-5. want partition data 2 sets train (80%) , test (20%) following requirements:

1. each user contains minimum 6 positive ratings (positive ratings 4,5)
2. on test, need have minimum 5 positive ratings each user. (because used calculating recall in evaluation section)
3. on train, minimum 1 positive rating required each user.
4. while maintaining condition 2 , 3, need participating proportion 80-20 withheld.

i start program follows:

• for condition 1, remove users ratings less 6.
• for condition 2 , 3, can start single user. find items has interacted , find positive ratings. randomly put 5 ratings in test , rest (1 or more) in train. repeat process users.
• this process can work can have 2 problems:

i. how make sure partitioning scheme 80-20 kept. not probably.

ii. how can make process fast not need loop on users? (for example using function @bsxfun).

this process done in matlab , not have , memory problem.

thanks opinions in advance