Hierarchical Clustering

[azhe825]

Active Learning with Clustering

Hierarchical sampling for active learning

Active learning using pre-clustering

[azhe825]

Active Learning with Clustering

Different scenarios of active learning:

• pool-based active learning
• query synthesis
• stream-based selective sampling

Query strategy (how to measure the informativeness of an example):

• probabilistic output of a classifier
• agreement between the members of a committee
• estimated reduction of error

Spectral clustering:

• use spectrum (eigenvalues) of the similarity matrix of the data to do dimensionality reduction
• then clustering on reduced dimensions.

[azhe825]

Hierarchical sampling for active learning

Sampling bias:

• If initial labels are chosen randomly

Build a hierarchical clustering Top down, sample from current leaves to query labels, sample less from purer leaves. (what if unbalanced?) Label clusters with majority labels observed Fit a classifier on the resulted data. (Why do we need this?) May query more labels around decision boundary

[azhe825]

Active learning using pre-clustering

P(y|x)=P(y|k)*P(k|x)