# Generating artificial data by means of LSTM

I got two classes namely positive and negative with 1500 samples on each a total of 3k. A sample sequence is like:

Period = 1min   Pattern = S1>M1>T1>B2>S2>M2>S3>T3>M3>B3
Period = 5min   Pattern = S1>M1>T1>B2>M2>S2>S3>T3>M3>B3
Period = 10min  Pattern = S1>M1>T1>B2>S2>M2>S3>T3>M3>B3
Period = 15min  Pattern = S1>M1>T1>B2>S2>M2>S3>T3>M3>B3
Period = 20min  Pattern = S1>M1>T1>B2>S2>M2>S3>T3>M3>B3
Period = 30min  Pattern = M1>S1>T1>B2>M2>S2>S3>T3>M3>B3
Period = 60min  Pattern = S1>M1>T1>B2>S2>M2>S3>T3>M3>B3
Period = 120min Pattern = S1>M1>T1>B2>S2>M2>S3>T3>M3>B3


So for each sample, we have a total of 8 sequences and within each sequence there are 10 elements which are used only once. I assigned a number for each of the elements from 0 to 9. Later, stored positive and negative samples into numpy arrays with shapes of (1500,8,10). By using LSTM's, I was able to achieve around 96% accuracy, which I'm glad and sufficient for the task.

For now, my task is to create artificial samples and make the LSTM's classify it. I should mention that there are some rules for sequence generation like M1>T1>M2>T3>M3. By applying all of these rules as permutation constraints, I was able to get down to a 12960 sequences that could be used in any of the periods. This makes a total of 12960^8~= Multiple orders of Petabytes of data. Not feasible to store all samples. So I gave up generating each sample even by this means.

What I'm looking for now is a way to generate artificial data, by means of some GAN style LSTM or whatever I don't know. Which will take (1500,8,10) arrays of positive and negative samples and spit out a positive or negative candidate with a label. Thanks.