Integrating Social Network Analysis and Topic Modeling in Social Media Research: A Systematic Literature Review of Open-Access Scopus Studies (2016-2026)

Social media leaves behind two distinct kinds of footprints every time we scroll, post, or share. One footprint is the content itself: the words we type, the opinions we voice, and the stories we tell. The other is the connection: the act of retweeting, replying, or following someone, which reveals who is listening to whom and how ideas travel between people. For years, researchers have studied these footprints separately. Some teams mapped the connections to see how communities form and how influence spreads, while others analyzed the text to understand what people are actually saying. But to truly understand the digital world, one must look at both the map of relationships and the map of conversation at the same time. The challenge has been figuring out how to weave these two different maps together without losing the details of either.

A recent systematic review by Dwi Jaya Saputra, a researcher at the University of Brawijaya, takes a comprehensive look at how scientists have attempted to combine these two methods over the last decade. The study examined 194 open-access research articles published between 2016 and 2026 that tried to merge social network analysis with topic modeling. The goal was not just to count how many studies existed, but to understand how they were built, what they found, and where the field is heading. The review reveals a field that is growing rapidly but often stopping at the surface, where researchers run two separate analyses side by side rather than weaving them into a single, deeper understanding.

The review began by searching a major academic database for studies that used both network analysis and topic modeling on social media data. After filtering for English-language articles that were freely available, the researchers narrowed a list of 560 potential papers down to 194 that met strict criteria. These studies covered a wide range of topics, from health discussions and political debates to business trends and environmental concerns. The data came from many platforms, though one platform, X (formerly Twitter), provided the information for more than half of the studies. Over time, the use of other platforms like Reddit, YouTube, and Weibo has increased, suggesting a broadening of the research landscape beyond a single source.

When the researchers looked at how these studies were conducted, they found a clear pattern in how the two methods were combined. In the vast majority of cases, about 71 percent, the researchers used a "parallel" approach. This means they ran a topic analysis to see what people were talking about and a network analysis to see who was connected, and then simply compared the two results. It is like looking at a map of a city's roads and a list of its popular shops separately, then noting that they seem to overlap, without actually tracing how the roads lead to the shops. While this method is common and useful for getting a general overview, the review found that it rarely tested whether the connections between the two were real or just a coincidence.

Only a small fraction of the studies, about 9 percent, attempted a "joint" approach, where the network and the text are analyzed together in a single mathematical framework. These studies are more complex because they try to estimate the structure of the community and the content of the conversation simultaneously. Another group of studies used a "network-to-topic" approach, where they first identified groups of people and then analyzed what those specific groups were saying. A rare fourth group, making up less than 5 percent, used a "topic-to-network" approach, building the connections between people based on the similarity of the topics they discussed. The review suggests that while the field is moving toward more sophisticated methods, the simple parallel approach remains the default for most researchers.

The tools used to analyze the text have also evolved. For a long time, a method called Latent Dirichlet Allocation, or LDA, was the standard for finding hidden themes in large collections of text. It appeared in about one-third of the studies reviewed. However, newer methods that use advanced language understanding, such as BERTopic, have started to appear more frequently, especially in recent years. Despite this technological shift, the review noted a significant gap in transparency: half of the studies did not name the specific algorithm they used in their abstracts. Furthermore, very few studies included a clear explanation of how they checked the accuracy of their findings. Most abstracts did not mention whether the researchers tested if their results would change if they used different settings or different tools.

The researchers concluded that while combining these two methods is now a routine part of social media research, the integration is often shallow. The field has grown from just one article in 2016 to 36 articles in 2025, showing a clear acceleration in interest. Yet, the way these studies are designed has not kept pace with the growth. The review proposes a new framework for future research that encourages scientists to move beyond simply running two separate analyses. Instead, they should choose a specific path for how the network and the text interact, whether that is building a network from topics or analyzing topics within specific communities. Crucially, the author argues that researchers must validate their combined findings, testing whether the link between the people and the words holds up under different conditions.

Looking ahead, the review outlines an agenda for the next decade of research. It calls for more studies on platforms other than X, including video-based sites and messaging apps that are currently under-represented. It also urges researchers to expand their work to include languages and regions that have been largely ignored, moving beyond the dominance of studies from the United States and China. Finally, the review suggests that the most promising path forward lies in developing models that treat the structure of the network and the content of the conversation as a single, interconnected system. By doing so, researchers can move from simply describing what is happening on social media to truly understanding how the connections between people shape the ideas they share.