Self-Adaptive Probability with Long Short-Term Memory for Early Detection of Depression in Web Document

Abstract

The depression is increasingly prevalent across all ages groups due to fast pace of life and many people use social media to express their feelings. Consequently, social media provides valuable data for early detection of stress and depression. However, challenges arise in text feature sensitivity, as unwanted text in documents lead to overfitting, making it difficult to capture complex relationships due to affect the detection accuracy. In this research, proposed Self-Adaptive Probability – Long Short Term Memory (SD-LSTM) technique for detecting depression is efficiently capture sequence data and achieves better accuracy. The SD-LSTM ability to adjust based on input patterns helps to detect different types of text data, minimizes chances of overfitting thereby enhancing accuracy. Pre-processing involved stemming, stop word removal, lowercase, removal of non-character to eliminate unwanted text and improve clarity of sentences. The feature extraction used XLNet, capturing nuanced meaning of words in context, which is particularly essential in depression detection where word meanings change significantly depending on context. The proposed SD-LSTM achieve better accuracy of 96.45% on Dreaddit, 96.85% of accuracy on Depression Mixed and 0.92 RSDD dataset. The existing method such as Gates Recurrent Unit (GRU), Logistic Regression (LR) is evaluated the proposed method.