Fig. 8

The effects of hyper-parameters on clustering performance of scGAL. For each hyper-parameter, the evaluation is conducted three times on the SA501 dataset by using different seeds. A Numbers of hidden layers in {1, 2, 3, 4} are tested for comparing different network architectures. The results suggest the network architecture with 4 hidden layers delivers the best results. B Different latent dimensions in {2, 3, 4, 5, 6} are compared, and the results indicate the best performance is obtained with latent dimension of 3. C Through evaluation of the effect of the hyper-parameter \(\lambda\) that balances the reconstruction loss of the AE and adversarial loss of the GAN, it is found that setting \(\lambda\) to 5 gives the best clustering results. D The size of buffer pool used to train the GAN has significant effect on model training, and buffer size of 64 is appropriate to get better clustering results