A novel hybrid SMOTE oversampling approach for balancing class distribution on social media text

Abstract

Depression is a frequent and dangerous medical disorder that has an unhealthy effect on how a person feels, thinks, and acts. Depression is also quite prevalent. Early detection and treatment of depression may avoid painful and perhaps life-threatening symptoms. An imbalance in the data creates several challenges. Consequently, the majority learners will have biases against the class that constitutes the majority and, in extreme situations, may completely dismiss the class that constitutes the minority. For decades, class disparity research has employed traditional machine learning methods. In addressing the challenge of imbalanced data in depression detection, the study aims to balance class distribution using a hybrid approach bidirectional long short-term memory (BI-LSTM) along with synthetic minority over-sampling and Tomek links and synthetic minority over-sampling and edited nearest neighbors’ techniques. This investigation presents a new approach that combines synthetic minority oversampling technique with the Kalman filter to provide an innovative extension. The Kalman-synthetic minority oversampling technique (KSMOTE) approach filters out noisy samples in the final dataset, which consists of both the original data and the artificially created samples by SMOTE. The result was greater accuracy with the BI-LSTM classification scheme compared to the other standard methods for finding depression in both unbalanced and balanced data.