Learning Person-specific Network Representation for Apparent Personality Traits Recognition

Recent studies show that apparent personality traits can be reflected from human facial behavior dynamics. However, most existing methods can only encode single-scale short-term facial behaviors in the latent features for personality recognition. In this paper, we propose to recognize apparent personality recognition approach which first trains a person-specific network for each subject, modelling multi-scale long-term person-specific behavior evolution of the subject. Consequently, we hypothesize that the weights of the network contain the person-specific facial behavior-related cues of the subject. Then, we propose to encode the weights of the person-specific network to a graph representation, as the personality representation for the subject, allowing them to be processed by standard Graph Neural Networks (GNNs) for personality traits recognition. The experimental results show that our novel network weights-based approach achieved superior performance than most traditional latent feature-based approaches, and has comparable performance to the state-of-the-art method. Importantly, the produced graph representations produce robust results when using different GNNs. This paper further validated that person-specific network's weights are correlated to the subject's personality.