Contextual Modeling of Personality States’ Dynamics in Face-to-Face Interactions Jacopo Staiano ∗ , Bruno Lepri †‡ , Kyriaki Kalimeri † , Nicu Sebe ∗ and Fabio Pianesi † ∗ Department of Information Engineering and Computer Science University of Trento, Italy {staiano,sebe}@disi.unitn.it † Fondazione Bruno Kessler, Trento, Italy {lepri,kalimeri,pianesi}@fbk.eu ‡ Massachussets Institute of Technology, Boston, USA Abstract—In this paper, we investigate the effectiveness of several dynamic graphical models for the task of personality states classification in a meeting scenario. Personality states are defined in social psychology literature as specific behavioral episodes in which a given subject exposes behaviors connected to a certain personality trait. The personality states we are addressing are those corresponding to the Big Five traits. I. I NTRODUCTION In everyday life, people constantly and unconsciously ex- ploit the ability to describe others on the basis of their behaviors: for instance, descriptors such as talkative, bold, sociable refer to the Extroversion personality trait dimen- sion; responsible, attentive, refer to the Conscientiousness personality trait, and so on. This ability of describing others through personality descriptors is crucial for explanation and prediction of others’ behaviors. The importance of personality has been acknowledged in a (increasing) number of studies: it has been showed that the basic dimensions of adaptivity [6] as well as people’s attitude towards computers [16] and conversational agents [15], are influenced by personality traits. In the social networking context, following analysis of text messages, personality matching between users has been proved to increase the chances of successful relationships [3]. In gen- eral, finding means to automatically obtain information about people’s personalities is considered very significant in order to let machines act in a proactive fashion or endowing them with the folk-psychological capability of explaining/predicting people’s behaviors [1]. Several works have explored automated personality analysis [10], [12], [9], often targeting the Big Five model of personality [2]. All these approaches to the automatic recognition of personality have more or less implicitly adopted the so-called person perspective on personality [5]: for a given behavioral sample, classify whether the sample belongs to an extrovert or introvert (or equivalently, to a neurotic or an emotionally stable, and so on). The problem with this approach is that it assumes a direct and stable relationship between, e.g., being extravert and acting extravertedly (e.g., speaking loudly, being talkative, etc.). Extraverts, on the contrary, can often be silent and reflexive and not talkative at all, while introverts Bruno Lepri’s research was supported by PERSI project funded by Marie Curie - COFUND - 7th Framework. can at time exhibit extraverted behaviors. While the person perspective has often dismissed these fluctuations of actual behavior as statistical noise, it has been suggested by Fleeson [5] that they are meaningful. The social psychology literature has coined the term personality states to refer to concrete be- haviors that can be described as having a similar content to the corresponding personality traits. In other words, a personality state describes a specific behavioral episode wherein a person behaves more or less, introvertly/extravertly, more or less neurotically, etc. Personality can then be reconstructed through density distributions over personality states, with parameters such as means, standard deviations, etc., summarizing what is specific to the given individual. In this paper we address the automatic classification of per- sonality states. Hence, we will be concerned with classifying whether people behaved extravertly or introvertly; neurotically or in an emotionally stable manner; and so on. To this end, we exploit graphical dynamic models that may explicitly incorporate hypotheses about the relationships among person- ality, actual behavior of the target and situational aspects. In particular, we will hypothesize that: a) personality states are the only causal determinants of people behavior; b) people behavior is modulated by the behavior and the personality states of their meeting parties (situational factor). The impor- tance of those relationships will be investigated by comparing the performance of three sequential models: generative linear- chain Hidden Markov Models (L-HMM) and discriminative linear-chain Conditional Random Fields (L-CRF) incorporat- ing only hypothesis (a); generative Influence Modeling (IM) incorporating (a) and (b). II. CORPUS Our experiments were performed on a 4-meetings subset of the Mission Survival II corpus [12]; the video stream corresponding to each meeting participant was split into 5- minute long slices. Personality state annotation was performed by volunteers using the Ten Item Personality Inventory [7], a 10-item questionnaire developed to obtain a quick measure of the Big Five dimensions. Annotators were required to assess on a 1 to 7 scale the personality of the participants based on the 5-minute slices containing a close-up view of the subject with 2011 IEEE International Conference on Privacy, Security, Risk, and Trust, and IEEE International Conference on Social Computing 978-0-7695-4578-3/11 $26.00 © 2011 IEEE DOI 896