A Probabilistic Approach To Multiple-Instance Learning

This study introduced a probabilistic approach to the multiple-instance learning (mil) problem. In particular, two bayes classication algorithms were proposed where posterior probabilities were estimated under dierent assumptions. The rst algorithm, named instance-vote, assumes that the probability of a bag being positive or negative depends upon the percentage of its instances being positive or negative. This probability is estimated using a k-nn classication of instances. In the second approach, embedded kernel density estimation (ekde), bags are represented in an instance induced (very high dimensional) space. A parametric stochastic neighbor embedding method is applied to learn a mapping that projects bags into a 2-d or 1-d space. Class conditional probability densities are then estimated in this low dimensional space via kernel density estimation. Both algorithms were evaluated using musk benchmark data sets and the results are highly competitive with existing methods.

Name thesis:degree_name

M.S. in Engineering Science

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer and Information Science

Year dc:date.available

2017

Author dc:creator

• Zhang, Silu

Contributors dc:contributor

• Yixin Chen
• Xin Dang
• Dawn Wilkins

• Data Visualization
• Multiple-Instance Learning
• Non-Linear Dimensionality Reduction
• Computer Sciences

Repository record dc:identifier

https://egrove.olemiss.edu/etd/944

OAI identifier oai:identifier

oai:egrove.olemiss.edu:etd-1943