6  /  6

SNA – Machine Learning – Dimensionality reduction

Visualization of a social network analysis (SNA) defined by a high-dimensional parameter space (21 dimensions). Dimensionality reduction (DR) is used here to allow the reduction of the number of variables that defines a system, into a lower-dimensional space (2D).

In this case, I’m using self-organizing maps (SOM), which represent a type of unsupervised artificial neural network, to reduce the high dimensionality of the data whilst also retain the high-dimensional non-linear associations.

Another feature extraction used (side by side image), is a linear feature extraction method -> K-means cluster analysis, (although linear methods don’t retain the associations distributed non-linearly in the high dimensional space).

#owl #machinelearning

 

As the dataset was relatively small, I used only CPU ( 2 x Intel Core i7-4930k @ 3.40GHz) computing for the training phase. If you are dealing with a big data set, a GPU  might be a better way to go to train your dataset.

CPU – 2 x Intel Core i7-4930k @ 3.40GHz

GPU – GPU

03_UnSupervised Learning - autoencoder 577 tensors 002.jpg03_UnSupervised Learning - autoencoder 577 tensors 00403_UnSupervised Learning - autoencoder 577 tensors 012.jpg03_UnSupervised Learning - autoencoder 577 tensors 001.jpg

03_UnSupervised Learning - autoencoder 577 tensors 010.jpg

03_UnSupervised Learning - autoencoder 577 tensors 013.jpg

×