The Japanese Society for Artificial Intelligence

Current research on emotion recognition has mainly focused on content dependent emotion recognition, where a model is trained and tested using user data from the same content sources (e.g., watch a movie or play a game). To provide cross-content services due to users’ emotions anywhere, it is necessary for a model to recognize users’ emotions in different content sources (i.e., content independent). Since limited studies have focused on content independent recognition, whether such emotion recognition has a competitive performance with content dependent emotion recognition is still unclear. To address this issue, we performed a comparative study of content dependent and independent emotion recognition by building CNN-based models from DEAP dataset. The DEAP dataset contains physiological data collected from 32 individuals while they were watching different videos. The data collected while watching a specific video is regarded as a single content. We built content independent model with leave-one-content-out approach. That is, using physiological data from one specific content for testing, and using the data from the remaining contents for training. As a result, we noticed that the performance of content independent recognition is significantly lower than that of content dependent recognition. We also identified that users’ emotions can be easily recognized in certain contents.