Computational Personality Analysis with Interpretability Empowered Prediction

Personality analysis can help individuals gain self-awareness, improve decision-making skills, and enhance relationships, while also providing valuable insights in fields like psychology, human resources, and marketing. Computational models, including traditional machine learning and deep learning models, have been beneficial in analyzing the impact of personality in sociological studies, especially deep learning models with interpretability. However, different computational models may produce different explanation results, which poses a challenge for social studies researchers in selecting appropriate explanations, as each model provides distinct prediction accuarcy values and explanation. To this end, this work introduces a computational personality analysis framework that incorporates Local Interpretable Model-Agnostic Explanations (LIME) to investigate analysis methods. The framework covers various computational personality models, encompasses pre-processing techniques, feature embedding, and focuses specifically on Myers-Briggs Type Indicator (MBTI) personality analysis, enabling valuable insights from predictions generated by diverse computational models. Cosine similarity is employed to evaluate the variations in explanation results produced by different computational models in relation to personality analysis. Our experiment reveals the key finding: although they have comparable predictive accuracy, there are not small explanation differences among the computational models.