General Notices

The Department of Computing and Information Technology invites the campus community to the PhD Upgrade Seminar for Mr. Kevin Baboolal, on Friday March 28 at 11 a.m. 

Mr Baboolal will address the topic Data-Dependent Feature Encoding For Enhanced Naive Bayes Classification

Interested persons can join in person at the DCIT Conference Room, Second Floor, Natural Sciences Building.

Abstract

In prior research, we introduced a supervised feature encoding methodology designed to enhance the efficacy of common machine learning regressors. This method involves projecting attribute values into the target space via a data-dependent transformation. The transformation maps an attribute value pair as a decaying average of the target value based on sample separation. This study extends this framework to the Naive Bayes Classifier, addressing a critical limitation.

The Naive Bayes Classifier exhibits compromised performance when encountering attribute values in a test sample that are absent from the training set. Traditional approaches, such as Laplace smoothing and Gaussian Naive Bayes, mitigate this zero-frequency problem using additive smoothing and global statistical information, respectively. In contrast, our approach leverages the proposed data-dependent kernel for prediction, thus providing a more personalised response to unseen attribute values. Furthermore, we augment the performance of the classifier by adapting our previously developed feature encoding technique to this classification context. Finally, we will explore heuristic strategies for the selection of optimal parameters within the proposed kernel framework and methods to reduce runtime.