Videos are increasingly becoming a prominent form of multimedia information readily available online, primarily through social networks. However, misinformation can easily be spread through videos and has the potential to go viral, with severe social consequences. Being able to identify the source of a video can add authenticity to it and help detect and track misinformation. Although different approaches have been proposed to identify the social network used to share a video, each one has flaws such as being dependent on the spatial domain of the video or vulnerable to laundering. One of the most robust techniques is based on the detection of the unique traces left by the compression process applied by the social network by analysing the DCT coefficients of a compressed video. Different social networks compress a video differently, such as using different coding parameters, leading to distinct differences in the DCT coefficients. This work improves upon previous DCT coefficient-based methods by using a novel feature vector created with the interval histograms of the DCT coefficients of all the color components of the I- and P-frames. By training a random forest classifier with this feature vector, significant improvements are achieved, even when videos are shared multiple times or edited.