Weakly-supervised subspace clustering for sequential data
Being unsupervised methods, subspace clustering algorithms do not generally allow for the incorporation of prior information about a dataset. While this is freedom from prior information can be a strength of such methods, in cases where prior information about a dataset is kn...
Published: |
|
---|---|
Online Access: | http://hdl.handle.net/2047/D20384360 |
Summary: | Being unsupervised methods, subspace clustering algorithms do not generally allow
for the incorporation of prior information about a dataset. While this is freedom from prior
information can be a strength of such methods, in cases where prior information about a
dataset is known or given, this information cannot easily be exploited to better model the
given dataset. This paper explores a way of extending the K-Subspaces clustering algorithm
to incorporate prior information about a dataset and using this information to generate
grammars that can provide weak supervision during the training of a subspace
clustering-based model. We review the problem of subspace clustering, detail a strategy for
adding weak supervision to the existing K-Subspaces algorithm, and compare this new method
with other subspace clustering methods in the domain of video segmentation.--Author's abstract |
---|