Learning Personalised Models for Automatic Self-Reported Personality Recognition

This paper presents our approach presented at the ICCV 2021 Understanding Social Behavior in Dyadic and Small Group Interactions Challenge: Automatic Self-reported Personality Recognition Track. Previous research has revealed differences in personality traits among different age groups, genders, and even cultures. However, existing methods have focused on one-fits-all approaches only and performed personality recognition without taking into consideration the user's profiles. We propose to learn personalised models of self-reported big five personality traits. In particular, our proposed approach automatically learns deep learning architectures for different user profiles using Neural Architecture Search (NAS) for predicting the Big Five personality traits scores from multimodal behavioural features. We experiment with two different user profiling criteria: gender and age, and compare the results of our approach with those of the participating teams. Our results show that personalised models improve the performance as compared to the generic model overall. Particularly, gender based user profiling combined with bimodal features reduce the prediction error by 0.128, achieving the state-of-the-art performance on the UDIVA dataset.