Enlarge microbiome dataset using stack_GAN

Background

The 16S rRNA abundance timeseries data used for prediction is small(hurdreds). So it is easy to suffer from overfitting. To solve the issue, the stack_GAN(generative adversarial network) is used to conduct data augment. Then the Recurrent neural network(to be more specific, Long short-term memory) is used to do prediction. The dataset used is the same as the Microbiome_Prediction, which serve as the baseline for the prediction accurary.

Three Loss Type

• Adversarial loss
• Conditional loss
• Entropy loss

Model List

model	input	output
discriminator1	real/fake sample	[the probability that sample is real,z1 reonstruction]
discriminator0	real/fake feature0	[the probability that feature0 is real,z0 reonstruction]
generator0	fake z0, real lable	fake feature0
generator1	fake z1, feature0	fake sample
encoder0	real sample	real feature0
encoder1	real feature0	predicted lable
adversary0	gen0_inputs	[dis0(gen0_outputs),enc1(gen0_outputs)]
adversary1	gen1_inputs	[dis1(gen1_outputs),enc0(gen1_outputs)]

Training workflow

1. train the encoder
2. train the discriminator
3. train the adversary0

Process
i) feed [real lables,z0] to gen0, get gen0_ouputs
ii) feed gen0_ouputs to dis0,enc1
iii) from dis0(gen0_outputs) ->get loss: binary,mse
iv) from enc1(gen0_outputs) ->get loss: categorical
note: the ground truth in iii) iv) is [real/fake, fake z0, real labels]

4. train the adversary1

Process
i) feed [feature0,z1] to gen1, get gen1_ouputs
ii) feed gen1_ouputs to dis1,enc0
iii) dis1(gen1_outputs) ->loss function: binary,mse
iv) enc0(gen1_outputs) ->categorical
note: the ground truth in iii) iv) is [real/fake, fake_z1, real_feature0]

Design

Result

0, 10%, 20%, ···, 1 of data is added to the original dataset. As image shows, it increase the preformance of LSTM model when adding 30% of data.

Reference

Snow, D. (2020). MTSS-GAN: Multivariate Time Series Simulation Generative Adversarial Networks. Available at SSRN.