Comparing SVD and SDAE for Analysis of Islamist Forum Postings

We analyze postings in the Turn to Islam forum using techniques based on singular value decomposition (SVD) and the deep learning technique of stacked denoising autoencoders (SDAE). Models based on frequent words and jihadist language intensity are used, and the results compared. Our main conclusion is that SDAE approaches, while clearly discovering structure in document-word matrices, do not yet provide a natural interpretation strategy, limiting their practical usefulness. In contrast, SVD approaches provide interpretable models, primarily because of the coupling between document and word variation patterns.