A Cost-Sensitive Hybrid Model of ALBERT Model and Convolutional Neural Network for Personality Classification
Rilo Chandra Pradana, Derwin Suhartono · CommIT (Communication and Information Technology) Journal · 2025
A tremendous amount of text data from social media activity can be used to extract information about a user’s personality, including the Myers-Briggs Type Indicator (MBTI). The MBTI personality type is extensively used to identify individual traits, which helps to solve problems in human resources and mental health awareness. Nonetheless, constructing an effective model for classifying MBTI types that are insensitive to unbalanced data remains a major challenge, as certain types dominate the social media environment. The research proposes a hybrid classification model that combines the transformer-based language model A Lite Bidirectional Encoder Representations from Transformers (ALBERT) with a Convolutional Neural Network (CNN), leveraging cost-sensitive learning to address class imbalance. The model is trained on the PersonalityCafe dataset and evaluated across the four MBTI dimensions. Experimental results show that the proposed ALBERT+CNN model achieves an overall F1-score of 77.67%, outperforming baseline models such as Bidirectional Encoder Representations from Transformers (BERT), Bidirectional Long Short-Term Memory (BiLSTM), and traditional CNN. When integrated with cost-sensitive learning, the model reaches an improved F1-score of 80.50%, surpassing the performance of oversampling techniques like Random Oversampling (ROS) and Synthetic Minority Oversampling Technique (SMOTE). The exponential cost function proves to be the most effective in weighting misclassifications for minority classes. In addition to higher accuracy, the proposed model demonstrates balanced prediction performance across personality dimensions, reducing bias toward dominant classes. These findings highlight the potential of hybrid deep learning and cost-sensitive strategies for personality classification in imbalanced textual data.