Synthetic minority oversampling in addressing imbalanced sarcasm detection in social media

Recent developments in sarcasm detection have been emerged as extremely successful tools in Social media opinion mining. With the advent of machine learning tools, accurate detection has been made possible. However, the social media data used to train the machine learning models is often ill suited due to the presence of highly imbalanced classes. In absence of any thorough study on the effect of imbalanced classes in sarcasm detection for social media opinion mining, the current article proposed synthetic minority oversampling based methods to mitigate the issue of imbalanced classes which can severely effect the classifier performance in social media sarcasm detection. In the current study, five different variants of synthetic minority oversampling technique have been used on two different datasets of varying sizes. The trustworthiness is judged by training and testing of six well known classifiers and measuring their performance in terms of test phase confusion matrix based performance measuring metrics. The experimental results indicated that SMOTE and BorderlineSMOTE - 1 are extremely successful in improving the classifier performance. A thorough analysis has been performed to better understand the effect of imbalanced classes in social media sarcasm detection.