- Two 3-dimensional Gaussian balls clustering
- Two 32-dimensional Gaussian balls clustering
- Autoencoder for 32-dimensional data
2023
Press F11 for full screen mode
Plan
3D two Gaussian balls, 80000 data points
component 1 and component 3 distribution and 2d density plot.
Gating on 2D projection
The separation line is calculated from the centers of Gaussian balls.
3D Gaussian balls 3d plot
3D rendering of two 3d Gaussian balls.
2D UMAP plot of 3D Gaussian balls
UMAP separates two clusters better and mis-clustering rate is improved.
32D two Gaussian balls, 80000 data points
component 1 and component 32 distribution and 2d density plot.
32D Gaussian balls - Gating on 2D projection
Gating on 2d projection seems not possible. The calculated line is cutting the oval shape along the shorter direction. The resulting clustering mis-clustered more than 20%.
32D Gaussian balls 3d projection
While the 3d projection shows more separation than the 2d projection, a significant overlap is still in display
2D UMAP of 32D Gaussian balls
UMAP separates two clusters completely and the mis-clustering rate is negligible.
Autoencoder design
Autoencoder design. There are about 600 parameters and the latent space has the dimension of two
Autoencoder training
Reconstructed data
Reconstructed data shows cluster separation even in 1d or 2d projects.
Reconstructed data - Gating on 2D projection
Reconstructed data gated on 2d projects. The mis-cluster rate is extremely low, which was not possible with the original data.
2D UMAP of reconstructed data
Two clusters in the UMAP is closer than the case of the original data. The mis-clustered rate is comparable.
Encoded data in 2d latent space
The encoded data in the latent space with the cluster number as color code. Two clusters are clearly separated.
–> –> –>
–> –> –> –> –> –> –> –>
–> –> –> –> –> –> –> –> –> –> –> –> –> –> –>
–> –> –> –>
–> –>