RandomGroupSpliter

What is RandomGroupSpliter?

This class provide a simple way to split the multi dataset into train and test dataset.

How to use it?

from RandomGroupSpliter import RandomGroupSpliter
RGS = RandomGroupSplitter(noise = True, force_bound = False, tolerance = 0.1, split_rate = 0.75, label_matrix = label_matrix)
label_distribution, result_idx = RGS.get_one_combination() # get one combination
remain = RGS.get_remain_combination(result_idx) # get all remain combination based on the result_idx

top_k_distribution, top_k_group_combination = RGS.get_multi_combination(noOfCombination = 10, top_k = 1000) # get 10 combination and only consider top 1000 combination in each iteration 
tried = RGS.get_tried_combination() # get all tried combination

caution

• do not set the top_k too large, it will be very time consuming.
• top_k = 1000 will cause the program run about 30 seconds for 40 datasets.
• if the value of best result does not change for a long time, please stop the program since it may be some bug in the program.

How to realize it?

by the idea of greedy algorithm, we can get the local optimal solutions by the guide of the reward function.