---

Supervised and Semi-Supervised Learning for Clustering

Usually Clustering is an unsupervised learning approach. In this approach every data is unlabeled and the algorithms decides clusters with no supervision, using a similarity/distance (see bregman divergences).

With supervised/semi-supervised we have some subset of the training data that is labeled. A transformation or metric (also called embedded) is learned from this subset and then it is used to label data into groups in a unsupervised way (using algorithms like K-Means).

Instead of assuming a fixed representation of the data, the method adapts the space so that standard clustering algorithms produce partitions aligned with labels.

An example is Learning Vector Quantization algorithm. Instead of classes we have prototypes vectors (or codebook vectors). Class lables guide how prototypes move and where decision boundaries are formed.

ML - Paper Presentation - Deep Spectral Clustering Learning (TODO): in classical spectral clustering we have the problem that depends heavily on the fixed representation (the original features of data). With Deep Spectral Clustering (DSC) a deep neural network learns the representation and the clustering structure. If supervised information is available it canguide the representation so that spectral embeddings separate classess more cleanly.