On class imbalanced learning: design of non-parametric classifiers performance indices and deep oversampling strategies/ Sankha Subhra Mullick

Thesis (Ph. D.) - Indian Statistical Institute, 2021

Includes bibliography

Introduction to Class Imbalance -- Appropriateness of Performance Indices for Imbalanced Data Classification -- Appropriateness of Performance Indices for Imbalanced Data Classification -- Parameter Independent Fuzzy k-Nearest Neighbor Classifier for Balanced and Imbalanced Data Classification -- Generative Adversarial Minority Oversampling

Guided by Prof. Swagatam Das.

A classifier expects to be trained on an equal number of distinct labeled examples from all the classes. This enables the learner to enjoy an equal opportunity to learn about each of the classes and likely enable a similar performance on all of them. However, in many real-life applications, it is common for all events to not occur with equal probability. This in consequence increases the difficulty to gather examples for the classes corresponding to rare events whereas annotated representatives of commonly occurring events are available in abundance. Thus, during the formation of the training set an imbalance between the number of representatives for the different classes is often observed. Such a class imbalanced training set is likely to bias the classifier in favor of the majority classes containing a larger number of labeled instances while the performance deteriorates on the minority class suffering from the dearth of training points. This inspired the classical machine learning and the emerging deep learning communities to consider the problem of class imbalance as a long-standing challenge.
In this introductory chapter, we first formally define the problem of class imbalance, discuss its relevance in detail, and then provide a brief survey highlighting the major research directions and key developments. In this chapter, we also discuss the research problems which motivated us to shape the objective of this thesis. Finally, we illustrate a road map of this thesis highlighting the primary contributions made by the rest of the chapters.