Dilated Separable Residual Network (DSRNet) for Personality Recognition using Textual Data

Personality significantly influences our attitudes/reactions/choices towards various situations and social interactions, affecting task suitability and team compatibility. Personality trait identification and analysis are vital in personal development, job selection, and performance enhancement. The Myers-Briggs Type Indicator (MBTI) is one of the effective models for understanding individual personality traits. Traditionally, MBTI assessments relied on psychologists, thus introducing subjectivity and bias. After post-pandemic, with increased online interactions using social media platforms, identifying traits without direct intervention through natural interactions is becoming inevitable before formal testing. Recently, deep learning algorithms have shown promise in developing such automated techniques for Personality Recognition; however, they face challenges with existing imbalanced data. We propose an efficient oversampling strategy using the transformer model GPT-2 and Synthetic Minority Oversampling Technique (SMOTE) to address imbalanced data problems with existing MBTI datasets. We introduce a novel lightweight CNN architecture called Dilated Separable Residual Network (DSRNet) that employs depth-wise separable convolutions to minimize computational cost while achieving good accuracy for Personality Recognition.