Enhancing Accuracy of MBTI Personality Prediction Using Deep Ensemble Models and Data Augmentation Techniques

Personality traits prediction from text has broad applications in various fields such as recruitment, job performance analysis, adaptive learning and personalised systems. Although traditional psychological assessments are widely used today, they may be subjective and impractical for large-scale deployment because they require the physical presence of a psychologist. This study presents an automated personality prediction model utilising text data. To address class imbalance, a significant factor that degrades model performance on the personality text dataset, a two-tier oversampling strategy has been implemented. The primary contribution of this study is to systematically evaluate the efficacy of various Deep Learning Architectures, including Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Bidirectional LSTMs, for MBTI prediction. Additionally, we have explored various ensemble learning approaches by combining separable CNNs, LeNet-5, and LSTM and BiLSTM models, thereby further improving prediction accuracy and generalisation. The experimental results show that integrating the proposed oversampling technique ensemble with the ensemble learning framework achieves higher accuracy, exceeding 87 percent, and outperforms previous models based solely on a single architecture or machine learning methods. The proposed method enables large-scale personality assessments to be deployed anywhere, at any time, reducing the need for the physical presence of psychologists.