Analyzing Affect During Virtual Meetings to Improve Quality of Interaction

Social intelligence has a significant effect on the quality of interaction in professional and personal communications. Developing technologies and tools to analyze the emotion of the audience in a classroom or a work meeting and providing the participants with appropriate information could augment the social intelligence of individuals. This augmentation is even more helpful in a virtual setup where elements such as eye contact are absent and hence the emotion and mental state of the audience is more difficult to perceive. One important psychological metric that could be helpful in this regard is ‘affect’. Developing tools to analyze interactions and rate the ‘affect’ of individuals is very valuable. To enable this capability within limited computational resources such as a laptop or a smart phone, the computationally expensive methods such as deep neural networks are not suitable. On the other hand, because of the complicated nature of human behavior and emotion, hand crafting the features for traditional machine learning algorithms is not straight forward and universal for different applications. Therefore, in this work, we leverage the power of neural networks with minimum number of hidden layer and activation nodes to achieve high accuracy in classifying ‘positive’ and ‘negative’ affect. We analyze individual's affect in multiple recorded virtual meetings in completely natural environment without any constraints on the participants in terms of movements, topics to talk about, artificially expressing emotions, etc. Action Units are extracted from video files and principal component analysis is performed to reduce the dimension of the feature space. Aiming at classifying subtle ‘positive’ and ‘negative’ affect (as opposed to only extreme cases of affect), testing accuracy of 76.8 % was achieved.