Graduate Thesis Or Dissertation
 

Learning with Partial Supervision for Clustering and Classification

Public Deposited

Downloadable Content

Download PDF
https://ir.library.oregonstate.edu/concern/graduate_thesis_or_dissertations/8910jz74b

Descriptions

Attribute NameValues
Creator
Abstract
  • In the field of machine learning, clustering and classification are two fundamental tasks. Traditionally, clustering is an unsupervised method, where no supervision about the data is available for learning; classification is a supervised task, where fully-labeled data are collected for training a classifier. In some scenarios, however, we may not have the full label but only partial supervision about the data, such as instance similarities or incomplete label assignments. In such cases, traditional clustering and classification methods do not directly apply. To address such problems, this thesis focuses on the task of learning from partial supervision for clustering and classification tasks. For clustering with partial supervision, we investigate three problems: a) constrained clustering in multi-instance multi-label learning, where the goal is to group instances into clusters that respect the background knowledge given by the bag-level labels; b) clustering with constraints, where the partial supervision is expressed as "pairwise constraints" or "relative constraints", regarding similarities about instance pairs and triplets respectively; c) active learning of pairwise constraints for clustering, where the goal is to improve the clustering with minimum human effort by iteratively querying the most informative pairs to an oracle. For classification with partial supervision, we address the problem of multi-label learning where data is associated with a latent label hierarchy and incomplete label assignments, and the goal is to simultaneously discover the latent hierarchy as well as to learn a multi-label classifier that is consistent with the hierarchy.
  • Keywords: Classification, Partial Supervision, Active Learning, Clustering
License
Resource Type
Date Available
Date Issued
Degree Level
Degree Name
Degree Field
Degree Grantor
Commencement Year
Advisor
Committee Member
Academic Affiliation
Non-Academic Affiliation
Rights Statement
Publisher
Peer Reviewed
Language
Replaces
Embargo reason
  • Existing Confidentiality Agreement
Embargo date range
  • 2017-10-21 to 2018-06-20

Relationships

Parents:

This work has no parents.

In Collection:

Items