Mapping Public Conversation on Indonesia's Free Nutritious Meal Program on Twitter (X) Using TF-IDF and K-Means Clustering

Abstract

Public programs now generate their own online commentary, and that commentary is worth reading systematically rather than anecdotally. This paper maps the topic structure of Indonesian-language conversation about the Free Nutritious Meal program (Makan Bergizi Gratis, MBG) on Twitter (X). We scraped 3,459 tweets posted in January and February 2025, and cleaned them through case folding, filtering, tokenization, stopword removal, and stemming, which left 2,926 documents, or 84.59% of the raw set. These were weighted with Term Frequency-Inverse Document Frequency (TF-IDF) into a 2,926 x 5,000 document-term matrix and partitioned with K-Means. The Elbow Method placed the optimal solution at K = 3, with an SSE of 2,760.3617 and a Silhouette Score of 0.0092. Three topics emerged: institutional endorsement tied to food security and self-sufficiency (94 tweets, 3.21%), everyday talk about how the program is being delivered and paid for (2,012 tweets, 68.76%), and the benefits of the program for children and their health (820 tweets, 28.02%). The Silhouette Score sits close to zero. We read this as a property of the corpus rather than a failure of the procedure: every document discusses the same program, so the clusters share a large common vocabulary and their boundaries are soft. The partition should therefore be treated as a first thematic map rather than a clean separation, and contextual representations such as Word2Vec, FastText, or IndoBERT are the obvious next step.