Multimodal mixture density boosting network for personality mining

Knowing people’s personalities is useful in various real-world applications, such as personnel selection. Traditionally, we have to rely on qualitative methodologies, e.g. surveys or psychology tests to determine a person’s traits. However, recent advances in machine learning have it possible to automate this process by inferring personalities from textual data. Despite of its success, text-based method ignores the facial expression and the way people speak, which can also carry important information about human characteristics. In this work, a personality mining framework is proposed to exploit all the information from videos, including visual, auditory, and textual perspectives. Using a state-of-art cascade network built on advanced gradient boosting algorithms, the result produced by our proposed methodology can achieve lower the prediction errors than most current machine learning algorithms. Our multimodal mixture density boosting network especially perform well with small sample size datasets, which is useful for learning problems in psychology fields where big data is often not available. Copyright © 2018 Springer International Publishing AG, part of Springer Nature.