Main content

Multiple Instance Learning with Applications to Concept Learning, Classification and Structural Pattern Recognition

Show simple item record

dc.contributor.advisor Chiu, David
dc.contributor.author Xu, Tao
dc.date.accessioned 2017-05-05T19:28:02Z
dc.date.available 2017-05-05T19:28:02Z
dc.date.copyright 2017-04
dc.date.created 2017-02-27
dc.date.issued 2017-05-05
dc.identifier.uri http://hdl.handle.net/10214/10405
dc.description.abstract In multiple instance learning (MIL), a class label is assigned to a collection (called bag) of instances instead of the individual instances. To learn concepts from a set of collectively labeled bags, MIL assumes that a bag is positive if and only if at least one instance in the bag is relevant to the target concept or negative otherwise. This research focuses on MIL in generating concept estimates (probably composed of multiple sub-concepts) that are not only expressive in terms of describing the class characteristics, but also useful in classifying new samples. For applications where sub-concepts are statistically independent and structurally unrelated, we propose two criteria: 1) adaptive kernel diverse density estimate (AKDDE) method that considers sub-concepts as the “peaks” of dense or high probabilistic regions of diverse positive samples; and 2) maximum partial entropy (MPE) that compares both the within-class similarity and the between-class dissimilarity. Taking into account both the statistical dependencies and the structural relations, we propose a common random subgraph (CRSG) model that represents a commonly observed subgraph pattern with probabilistically varying attributes of a class. The experimental results on both artificial and benchmark datasets demonstrate the adequacy and the advantage of the proposed methods, in describing the conceptual components, and at the same time effective in classifying new samples. en_US
dc.language.iso en en_US
dc.subject multiple instance learning en_US
dc.subject concept learning en_US
dc.subject partial entropy en_US
dc.subject adaptive kernel diverse density estimate en_US
dc.subject common subgraph modeling en_US
dc.subject common random subgraph en_US
dc.subject Research Subject Categories::TECHNOLOGY en_US
dc.title Multiple Instance Learning with Applications to Concept Learning, Classification and Structural Pattern Recognition en_US
dc.type Thesis en_US
dc.degree.programme Computer Science en_US
dc.degree.name Doctor of Philosophy en_US
dc.degree.department School of Computer Science en_US


Files in this item

Files Size Format View Description
Xu_Tao_2017_phd.pdf 1.051Mb PDF View/Open Main article

This item appears in the following Collection(s)

Show simple item record