PERSONALIZED ESTIMATION OF DAILY EMOTIONS AND INTERPERSONALCONFLICT BETWEEN ROMANTIC PARTNERS VIA METRIC LEARNING

Abstract

A novel personalized model using metric learning via siamese Neural Network is implemented to estimate daily emotions and detect interpersonal couples’ conflict using moment-to-moment multimodal bio-behavior signals (i.e., physiological, linguistic and acoustic signals) and additional relationship characteristics that includes individuals’ relationship satisfaction and attachment information. The ambulatory couples’ data has high inter-participant variability because each participant have a different distribution of data. Hence, a personalized model that has ability to eliminate the inter-participant variability and preserve the behavior characteristic is likely to perform well. Personalized learning implemented using metric learning via siamese neural network have innate ability to rank the pair of inputs after learning the personalized embeddings. Variants of the proposed personalized model and loss functions have been implemented and explored in this study. The performance of these proposed models are compared against that of non-personalized models such as feed-forward neural network.