Article Preview
TopIntroduction
Machine learning (ML) is continuously involved in the varied application (Zhou et al., 2017). According to the Forbes report, the global ML market in 2017 was $1.58B and expected to reach $30B by 2024 (Moorning, 2017). ML enables organizations in cost reduction up to 25%, better customer acquisition and retention, and new revenue generation across services and products (Lee and Shin, 2020). Advancements in big data have helped ML to use in healthcare; natural language processing, speech recognition, and computer vision (Qiu et al., 2016). The information provided by big data is useful for building predictive analysis (Wazurkar et al., 2017) and pattern extraction (Moreno et al., 2016). However, challenges for ML algorithms to handle voluminous records are of data compression (Azar et al., 2019; Narwane and Sawarkar, 2019), data cleaning (Gudivada et al., 2017), task scheduling (Ji and Wang, 2017), workflow management (Zhou et al., 2017), and class imbalance (Holden, 2016). Class imbalance problem affects sensitive domains healthcare, cyber security, and software defect predictions (Leevy et al., 2018).
In ML variety of methods were tried to handle class imbalance problem. These methods are broadly classified as data level, algorithm level, and hybrid level. In the data level approach, the data set was divided into a majority and minority class (Kalsoom et al., 2018; Bellinger et al., 2019). Mostly under sampling and oversampling algorithms were used in this approach (Hassib et al., 2019). It was very difficult to apply the same ML algorithm to every data set to handle class imbalance. As per the nature of the data set, one needs to choose an algorithm or combination of algorithms (Sitompul et al., 2018; Hsu et al., 2018). Popularly used ML algorithms are decision trees, support vector machines, and naïve Bayes (Feng et al., 2018). Limitations of data level are data redundancy in oversampling (Hassib et al., 2019) and loss of data in under-sampling (Akkasi et al., 2017). The algorithm level is more efficient than the data level; however, this method has its limitations. Bagging bad classifier (Feng et al., 2018), sensitivity towards outliers and noisy data (Cho et al., 2019), and lack of proper training (Kumar et al., 2019) are some of the limitations of algorithm level.
The hybrid method is a combination of data and algorithm method. The hybrid method can work efficiently if proper combinations of data level/algorithms level/both are used (Bader-El-Den et al., 2019; Al Majzoub et al., 2020). Previous researches worked on the support vector machine (SVM) modeling algorithm (Kumar et al., 2016), sampling techniques (Patil et al., 2020), and ensemble methods (Galar et al., 2011). Synthetic minority oversampling technique (SMOTE) is often used by researchers to balance the uneven distribution of classes (Al Majzoub et al., 2020). Through a synthesis of the literature, it has been observed that the previous studies rarely investigated the following: i) all standard ML algorithms, ii) data set from balanced to unbalanced, and iii) the impact of various class distributions on data set. To address these research gaps, this study addresses the following research questions (RQs):
RQ1: What are the different ML methods used to address the class imbalance problem?
RQ2: Describe a practical approach for consideration of class imbalance?
RQ3: How to analyze the impact of the class imbalance problem in the dataset?
To address these RQs, a detailed literature survey was carried out on the following methods ‘data level’, ‘algorithm level’, and ‘hybrid level’. Data sets were collected from the Kaggle dataset repository. The study was performed on two datasets namely, Titanic and Pima Indian Diabetes Prediction. Standard ML algorithms used to analyze the effects of balanced to unbalanced datasets. The Research Objectives (ROs) of paper are as given below.
RO1: To understand previous studies of ML methods for class imbalance problem.
RO2: To investigate train and test data accuracy for different class distribution.
RO3: To analyze the impact of class imbalance using for different class distribution