Fig. 1

Illustration of the proposed Adaptive Focal Loss with Personality-Stratified Splitting for Hard Class Imbalance in Multi-Dimensional Personality Recognition. Following the proposed dataset stratification and splitting approach, the training process begins with feature processing, including proposed novel datasets processing techniques along with text tokenization, feature extraction (LIWC and TF-IDF), and tensor conversion. A batch balancing inspector assesses class distribution, and a weight adjuster computes instance-specific weights, which are later utilized alongside difficulty scaling in the Adaptive Focal Loss function. During backpropagation, binary cross-entropy (BCE) loss is computed, and an adaptive focal factor is applied to adjust gradient scaling dynamically.