Reaction videos from YouTube provide a range of possibilities when looking to investigate human behavior. This research aims to present a method of creating a Spontaneous Facial Expression dataset from YouTube reaction videos. In this work, we use Convolutional Neural Networks to classify emotions in facial expressions, as well as feature extraction tools to support this classification. To understand the behavior of faces that react to a given group video, Agreement and Continuity analyzes were designed to identify spontaneous emotions that cannot be classiifed by Neural Networks with high assertiveness. Analyzing the behavior of the faces in a group, it was identified that only 26% of the total amount of faces react in similar ways and in the same time, however, 71% of the faces keep their emotion between transitions of a given moment of reaction, although not in the best frame.