PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey of Methods for Managing the Classification and Solution of Data Imbalance Problem

Khan Md. Hasib, Md. Sadiq Iqbal, Faisal Muhammad Shah, Jubayer Al Mahmud, Mahmudul Hasan Popel, Md. Imran Hossain Showrov, Shakil Ahmed, Obaidur Rahman

arXiv:2012.11870v1cs.LG

TL;DR

Class imbalance can skew classification toward the majority class, leaving minority classes poorly represented in model performance. This survey reviews 24 studies and their single, ensemble, and hybrid approaches, finding that differences among studies prevent a clear comparison of methods' superiority.

Abstract

from arXiv · show

The problem of class imbalance is extensive for focusing on numerous applications in the real world. In such a situation, nearly all of the examples are labeled as one class called majority class, while far fewer examples are labeled as the other class usually, the more important class is called minority. Over the last few years, several types of research have been carried out on the issue of class imbalance, including data sampling, cost-sensitive analysis, Genetic Programming based models, bagging, boosting, etc. Nevertheless, in this survey paper, we enlisted the 24 related studies in the years 2003, 2008, 2010, 2012 and 2014 to 2019, focusing on the architecture of single, hybrid, and ensemble method design to understand the current status of improving classification output in machine learning techniques to fix problems with class imbalances. This survey paper also includes a statistical analysis of the classification algorithms under various methods and several other experimental conditions, as well as datasets used in different research papers.

Introduction

Class imbalance means class labels are unevenly distributed, with fewer minority-class examples than majority-class examples. This survey reviews prior approaches and statistical comparisons for addressing the resulting classification challenge.

Introduction

  • A dataset is imbalanced when its class distribution is nonuniform; the less represented class is the minority and the remainder is the majority class.The passage links imbalance to impaired classification performance and predictive-model efficiency.Source: Introduction · p. 1
  • Class disparity can bias classifiers toward the majority class, reducing minority specificity while preserving high majority-class accuracy.The passage also notes that some small-imbalance metrics were beneficial.Source: Introduction · p. 1
  • Multiclass datasets contain more than two classes, and multiclass imbalance remains less addressed than binary imbalance.The passage also notes study of imbalance in standard and multilevel classification.Source: Introduction · p. 1
  • External approaches commonly use sampling, while internal approaches include Genetic Programming and Extreme Learning Machine methods.The survey frames these as approaches to handling data imbalance.Source: Introduction · pp. 1–2
  • The survey includes 24 studies published in 2003, 2008, 2010, 2012, and 2014–2019, with statistical comparisons of methods and algorithms.It also discusses function-selection steps and identifies future research issues.Source: Introduction · p. 2

Research Paper Overview

The overview organizes imbalance methods around sampling, classifier-level changes, ensembles, and hybrids. It describes representative procedures for modifying class distributions and combining sampling with learning algorithms.

Sampling

  • Sampling balances class distributions by resizing training data, using under-sampling or over-sampling techniques.Over-sampling increases minority-class instances; the passage states that new data can be appended while preserving the primary dataset.Source: Data Level Methods · p. 2
  • SMOTE increases the minority class by generating synthetic examples from existing data.The passage says it can produce patterns with a distribution similar to the true one, while other methods continue to improve it.Source: Data Level Methods · p. 2
  • Under-sampling removes majority-class examples, reducing training-set size and potentially shortening model training time.Random Under-Sampling (RUS) removes majority examples randomly; its simplicity, speed, and efficiency motivate its use in boosting.Source: ADASYN · p. 2

Algorithm-level approaches

  • Algorithm-level approaches target minority-class learning by adjusting probability estimates or class-specific costs.These are described as internal approaches that modify current classifier algorithms.Source: Algorithm Level Methods · p. 3

Ensemble approaches

  • Bagging integrates multiple learners using averaging, while boosting changes sample weights so some samples are run more frequently than others.Source: Ensemble Methods · p. 3

Hybrid approaches

  • Hybrid approaches combine sampling with boosting; RUSBoost is presented as a simple, quick, less complex method for imbalanced data.LIUBoost combines under-sampling with cost-sensitive learning, weighting instances by hardness to reduce information loss.Source: Hybrid Methods · p. 3
  • SMOTEBoost generates minority-class synthetic examples during boosting, while RHSBoost combines random under-sampling and ROSE sampling.HUSBoost is described as using data cleaning, data balancing, and classification.Source: Hybrid Methods · pp. 3–4

Statistical Comparison of Related Work

The survey describes 24 studies across publication years and compares single-level, ensemble, and hybrid approaches to class imbalance. It also summarizes datasets used and notes limits on comparing reported performance.

Statistical Comparison of Related Work

  • The surveyed papers span 2003, 2008, 2010, 2012, and 2014–2019; 2017 had the most, with 6 papers.Three papers each were addressed for 2014 and 2019.Source: Distribution of Papers by Year of… · p. 4, Figure 1 · p. 4
  • The survey classifies 10 papers as single-level, 7 as ensemble, and 7 as hybrid methods.These categories account for the 24 enlisted papers.Source: Distribution of Papers by Year of… · p. 5, Table 1 · p. 5
  • Single-level studies use sampling classifiers and traditional algorithms, ensemble studies combine algorithms, and hybrid studies combine approaches to address imbalance.The survey names SMOTE, DT, SVM, KNN, ENN, and Naïve Bayes among commonly used single-level methods; Table 4 summarizes hybrid methods and their reported results.Source: Single Level Method · p. 5, Hybrid Method · p. 5
  • The surveyed dataset inventory includes KDD Cup 1999, Glass, Ecoli2, Abalone, Wisconsin breast cancer, Yeast, Ionosphere, and Liver-disorder-bupa.Table 5 gives their year-wise distribution; the survey also notes use of private or non-public datasets and medical and banking datasets.Source: Used Dataset in Researches · p. 5, Table 5 · p. 8
  • Performance comparisons remain inconclusive because studies use separate datasets and imbalance levels, report contradictory assessments, and may depend on problem complexity and metrics.The survey calls for broader experiments across datasets and replicated trials; it reports that only about one-third of experiments specify repetitions.Source: Discussion of Surveyed Works · p. 9

Conclusion and Future Works

The survey presents imbalanced-data classification as a continuing concern, offers a comparison of prior work while acknowledging that it is not in-depth, and identifies directions for future research.

Conclusion and Future Works

  • The survey says its comparison of prior papers is fair but not in-depth, and calls for broader evaluation of methods across data sets and imbalance levels.It also recommends complementary performance indicators and statistical evidence to help define optimal deep-learning and traditional machine-learning approaches.Source: Conclusion and Future Works · p. 9
  • Removing redundant and irrelevant features during training is identified as a way to increase system performance.Source: Conclusion and Future Works · p. 9
  • The authors identify feature selection as important and describe hybrid or ensemble classifiers as more feasible for performance measurement than a single classifier.Source: Conclusion and Future Works · p. 9
  • Future work may test new hybrid and cluster-based approaches alongside deep learning for class imbalance in big data and class-rareness settings.Source: Conclusion and Future Works · p. 9

Author’s Contributions

Author contributions include work across paper sections, design and implementation, data gathering, experimentation, organization, and revisions responding to reviewers.

Author’s Contributions

  • Khan Md. Hasib contributed across all sections, worked on design and implementation, and helped correct the paper and respond to reviewers.Source: Author’s Contributions · p. 9
  • Md. Sadiq Iqbal and Md. Imran Hossain Showrov contributed to data gathering, experimentation, and paper organization.Showrov also conceived the idea, worked on implementation, and participated in corrections and reviewer responses.Source: Author’s Contributions · p. 9
  • Faisal Muhammad Shah coordinated data gathering, design, and experiment execution, contributed across all sections, and helped revise the paper and respond to reviewers.Source: Author’s Contributions · p. 9
  • Jubayer Al Mahmud, Mahmudul Hasan Popel, and Shakil Ahmed coordinated data gathering, design, and experiment execution and helped with corrections and reviewer responses.Source: Author’s Contributions · pp. 9–10

Ethics

The corresponding author states that the article is original, unpublished, approved by all authors, and has no ethical issues.

Ethics

  • The corresponding author states that the article contains unpublished material, all authors approved the manuscript, and no ethical issues are involved.Source: Ethics · p. 10
Loading 2012.11870v1…