Text compression works by accumulating context information in a memory buffer. The accumulated context information is used for predicting the next symbol/bit. STARLIT makes better predictions by first ordering the articles according to similiarity. From the GitHub page:
> During the searching phase, firstly, each article is mapped to a feature vector using a Doc2Vec model. Secondly, considering each feature vector to be a point in a Euclidean space, the Traveling Salesman Problem is solved resulting in the shortest path visiting each point. As a result of these operations, the found shortest path represents the order of all articles where similar articles are placed nearby.