Mapping News Categories: An Experiment on Louvain Community Discovery and Machine Learning

The study designs an explainable, lexical feature–based pipeline that examines algorithmic visibility and editorial homogenization in Turkish digital news through measurable indicators. The analytical background data were compiled by capturing, at regular intervals, the headlines and standfirsts (lead texts) of news items featured in the homepage slider/headline area of selected news websites. Following cleaning steps aimed at reducing repetitive template patterns and improving comparability, the data were converted into embeddings. Next, top-level category classification and within-category sub-labeling were applied, making the data traceable in terms of broad agenda clusters as well as micro-themes within categories. Proximity among subcategories was identified using a similarity graph and the Louvain community detection algorithm. The degree of divergence was evaluated via the number of communities and modularity (Q). Community-level representative terms not only characterised dominant discursive patterns within each sub-area but also supported the examination of homogenization dynamics. Overall, the findings indicate that a limited number of top-level categories concentrate on homepage visibility; while more fragmented sub-community structures appear in some areas, integrated structures predominate. This approach, applicable to datasets that do not require large language models beyond labelling, offers a methodological and practical framework for digital journalism research.