Enterprise AI Analysis

Recognizing the limitations of NLP-based approaches, this study focused on image-based modeling of handwritten text. An end-to-end Convolutional Neural Network (CNN) was proposed to detect the five core personality traits (Openness, Conscientiousness, Extroversion, Agreeableness, and Neuroticism). Additionally, extensive experimentation was conducted using transfer learning approaches with multiple pre-trained CNN architectures (VGG16, VGG19, ResNet variants, InceptionV3, MobileNet, DenseNet variants). Transfer learning leverages knowledge from models trained on vast datasets like ImageNet to enhance performance on the HiEnWrite dataset, significantly speeding up the learning process and improving generalization. The pipeline involves feature extraction using these pre-trained models, followed by a regression head for personality score prediction.

Our experimental results demonstrate the efficacy of the proposed approach, with VGG19 emerging as the top-performing model. VGG19 achieved a remarkable maximum Pearson Correlation Coefficient (PCC) of 0.416 on the validation set, indicating a robust linear relationship between predicted and actual personality trait values. This model also exhibited commendable efficiency with a training time of 209.126 seconds and a testing time of 5.509 seconds. Data augmentation played a crucial role, enhancing model robustness and generalizability by providing additional diversity in training samples and mitigating overfitting. Residual analysis further validated the model's consistent accuracy across the dataset.

To understand which parts of handwriting contribute most to personality trait predictions, Grad-CAM (Gradient-weighted Class Activation Mapping) was applied to both English and Hindi samples. The resulting heatmaps revealed a consistent spatial bias: model attention was primarily focused on the central horizontal region of handwriting. Visual inspection indicated the model's reliance on interpretable handwriting features such as letter slant, character size consistency, and intra-word spacing. This pattern was robust across languages and traits, suggesting that CNN-based models implicitly learn and leverage these structural cues, opening possibilities for identifying trait-specific handwriting characteristics through deep learning methods.