Enhancing Topic Modeling Coherence through Aggregation: A Case Study of Indonesian Rupiah Inflation on Social Media

The purpose of this research is to enhance coherence through the application of an aggregate topic model using social media data X (Twitter) with the inflation rupiah study case. In April 2024, Indonesia has experienced inflationary. Interrelated domestic and international factors influence the prevailing inflationary wave. Simultaneously, social media, especially X, has emerged as an important platform for individuals to articulate their perspectives on the consequences of inflation. However, with so many tweets, it becomes a challenge for the public or certain stakeholders to know and understand the issues. To know and understand what public opinion is like, Latent Dirichlet Allocation (LDA) is used to extract topics, but the coherence result of LDA is 0.422, which is quite low and makes topics biased and difficult to understand, so the use of topic model aggregation to generate topics with optimal coherence levels and Large Language Models (LLM) to interpret topics, is an effective solution. After testing, it was found that in 7 out of 10 topic modeling trials, the aggregation activity successfully alleviated the topics from a coherence score of 0.422 to 0.644. The general conclusion of this research is that aggregation has been shown to improve topic coherence and, thus, a $21 \%$ increase in scores. These results emphasize that incorporating aggregation models in various topics has brought about a marked improvement compared to non-aggregation methods. With increased coherence, topics will be easier for the public or relevant stakeholders to interpret and understand.