Personality prediction via multi-task transformer architecture combined with image aesthetics.

This article focuses on predicting human personality traits based on image aesthetics from social media content, specifically leveraging the Big-Five personality model (Openness, Conscientiousness, Extroversion, Agreeableness, Neuroticism). The authors propose a two-stage deep learning architecture combining a multi-task encoder/decoder network—integrating a Swin Transformer and convolutional neural network (CNN)—to jointly predict image aesthetics and personality traits, followed by an attention mechanism that refines personality prediction using aesthetic scores. Evaluated on the PsychoFlickr and Flickr-AES datasets, the model outperforms several state-of-the-art methods, achieving a Spearman Rank Order Correlation Coefficient (SROCC) of 0.776 for image aesthetics and 0.6730 for personality traits, with a 7.02% accuracy improvement when incorporating aesthetic influence. The study also highlights correlations between personality traits and image preferences, noting ethical considerations in dataset use and suggesting future work incorporating multimodal data for enhanced personality prediction.