In this thesis, prediction of romantic, social and sexual attraction between two people using bodily coordination features is studied. Attraction is one type of interest that can occur between interacting people and understanding the modeling of it can help understand how to model other types of interests in human-human interactions because similar methods can be used to model the non-verbal behavior that reveals interest. Previous research in psychology and social signal processing fields showed that synchrony and convergence of both audio features and body movements of people during an interaction are indicators of interest. However, audio features require recording people's voices during interactions and it can be disturbing for people especially during more personal conversations. In addition, capturing nonverbal cues from video mostly requires recording people with a camera from front and it might make people more aware of being recorded and intervene with the naturality of the interaction. Moreover, processing the videos to extract specific nonverbal behavior can be costly. Based on these ideas, we decided to use motion channel and hypothesize that movement synchrony and convergence features can be used to automatically quantify and predict attraction. We propose a novel method of estimating romantic, social and sexual attraction between two people by quantifying their bodily coordination using wearable sensors in a speed-date setting. We developed simple synchrony and convergence features, inspired from the literature and specifically adapted to be extracted from accelerometer data. To our knowledge, this is the first time that motion convergence is used for estimating attraction. Our features could predict one-way social attraction with a 73% Area under the ROC curve (AUC), out-performing previous work in a similar setting. We also showed that prediction performance increased when the male and female data are separated, aligning with the theories in psychology studies. We could also predict mutual romantic attraction with an AUC of 80%. We found that different types of attraction can be estimated better using different feature types, more specifically we could predict social attraction better using movement correlation features whereas for romantic and sexual interest mimicry features were better indicators. Moreover, features extracted from different types of signals recorded from accelerometers showed varying performances for different attraction types. Additionally, asymmetric features out-performed the symmetric features and our synchrony features showed better performance than convergence features. Finally, we a
have seen that motion convergence can occur in people having an interaction regardless of attraction.