Media Text Analysis Based on One-Dimensional Hashtag Embeddings

We propose a new methodology for analyzing topics of interest in media space. The method includes several steps. First, hashtags (or keywords) are extracted from the corpus of texts under study for each publication to build a bipartite network of hashtags and publications. Such a network representation is given by a matrix, the rows of which correspond to tags, and the columns to articles. Then, transform this representation to square matrix of hashtag co-mentions. Next, the algorithm builds a projection of this space into a one-dimensional representation, solving the optimization problem of minimizing the distance of the path around all the vertices of the graph (the traveling salesman problem). Thus, we get a ranked looped list of hashtags. Given the ability to easily visualize and interpret the results, this approach can also be used to examine changes in the structure of the media space over time.