dc.description.abstract | Apparent personality detection has emerged as a prominent research area within deep learning. While numerous
deep learning solutions have been developed to predict personality accurately, the lack of transparency in how
these models derive predictions based on facial features undermines trust in their results. This study focuses on
identifying and differentiating facial features that contribute to the Big-Five personality traits, addressing
transparency in model predictions. To conduct our experiments, we utilised the ChaLearn First Impressions V2
dataset, with background removed frames ensuring models focused more on human features than background in
the learning process. We began by developing Convolutional Neural Networks architectures using pre-trained
VGGFace and VGG19 models. Subsequently, we employed the Grad-CAM and Guided Grad-CAM model
explainable AI techniques on the test and validation datasets, utilising the trained models. Furthermore, we
employed the "SelectKBest" feature selection method to analyse the outcomes of the interpretability techniques.
VGG19 achieved higher accuracy (90%) compared to VGGFace (89%). Our investigation reveals that personality
prediction extends beyond facial features, with XAI techniques emphasizing non-facial aspects such as
background information. Statistical analysis across deep learning architectures shows no significant correlation
between features identified by XAI techniques by giving different F1-scores. Despite VGG19's superior accuracy,
it exhibits a stronger inclination towards non-facial data, while VGGFace prioritizes facial features, highlighting
the nuanced nature of personality prediction and suggesting avenues for further research. | en_US |