Date on Master's Thesis/Doctoral Dissertation


Document Type

Doctoral Dissertation

Degree Name

Ph. D.


Computer Engineering and Computer Science

Degree Program

Computer Science and Engineering, PhD

Committee Chair

Frigui, Hichem

Committee Member

Zhang, Xiang

Committee Member

Nasraoui, Olfa

Committee Member

Park, Juw Won

Committee Member

Zhang, Hui

Author's Keywords

Multiple instance learning; feature selection; local feature selection; multiple instance learning classification; explaining convolutional neural networks; explainable machine learning


Feature selection is a data processing approach that has been successfully and effectively used in developing machine learning algorithms for various applications. It has been proven to effectively reduce the dimensionality of the data and increase the accuracy and interpretability of machine learning algorithms. Conventional feature selection algorithms assume that there is an optimal global subset of features for the whole sample space. Thus, only one global subset of relevant features is learned. An alternative approach is based on the concept of Local Feature Selection (LFS), where each training sample can have its own subset of relevant features. Multiple Instance Learning (MIL) is a variation of traditional supervised learning, also known as single instance learning. In MIL, each object is represented by a set of instances, or a bag. While bags are labeled, the labels of their instances are unknown. The ambiguity of the instance labels makes the feature selection for MIL challenging. Although feature selection in traditional supervised learning has been researched extensively, there are only a few methods for the MIL framework. Moreover, localized feature selection for MIL has not been researched. This dissertation focuses on developing a local feature selection method for the MIL framework. Our algorithm, called Multiple Instance Local Salient Feature Selection (MI-LSFS), searches the feature space to find the relevant features within each bag. We also propose a new multiple instance classification algorithm, called MILES-LFS, that integrates information learned by MI-LSFS during the feature selection process to identify a reduced subset of representative bags and instances. We show that using a more focused subset of prototypes can improve the performance while significantly reducing the computational complexity. Other applications of the proposed MI-LSFS include a new method that uses our MI-LSFS algorithm to explore and investigate the features learned by a Convolutional Neural Network (CNN) model; a visualization method for CNN models, called Gradient-weighted Sample Activation Map (Grad-SAM), that uses the locally learned features of each sample to highlight their relevant and salient parts, and a novel explanation method, called Classifier Explanation by Local Feature Selection (CE-LFS), to explain the decisions of trained models. The proposed MI-LSFS and its applications are validated using several synthetic and real data sets. We report and compare quantitative measures such as Rand Index, Area Under Curve (AUC), and accuracy. We also provide qualitative measures by visualizing and interpreting the selected features and their effects.