Clustering and salience sensitivity of the Context–Aware Media Presence Index

It reports a simulation study of how the Context-Aware Media Presence Index (CMPI) responds to the two choices the framework leaves to the user: the algorithm that groups documents into stories, and the salience function that scores how central an entity is to each story.</p><p dir="ltr">The study uses synthetic news corpora about a fictional politician, generated with a local large language model, so the true story of every document is known.

Three clustering algorithms (leader–follower, Leiden community detection and k-means) were run over parameter grids, and CMPI was compared with its value under the true grouping. Both threshold-based algorithms kept CMPI within about 10% of that value over a wide range of thresholds. K-means did not.