TikTok StitchGraph: Characterizing communication patterns on TikTok through a collection of interaction networks

Mads Høgenhaug, Marcus Friis, Morten Pedersen, Luca Rossi · 2025

Read TikTok StitchGraph: Characterizing communication patterns on TikTok through a collection of interaction networks by Mads Høgenhaug, Marcus Friis, Morten Pedersen, Luca Rossi (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-18661.txt.

Introduction

Mads Hogenhaug and co-authors map conversation networks on TikTok in a 2025 paper. TikTok stitches let users reply to videos with videos of their own. Those chains form graphs of who answers whom. Video-first platforms hide such structure, so it has had little study.

The team built 36 graphs from platform APIs and web scraping. User graphs use dots for users and arrows for stitch links. Mining for common small patterns shows stars rule the space. Cycles are rare, and one-way stars beat mixed forms. Sentiment tags ride on the links. A side check finds these graphs look much like Twitter reply networks, a surprise bridge between platforms.

Catalog record

CREATORMads Høgenhaug, Marcus Friis, Morten Pedersen, Luca Rossi
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE36 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-18661

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

TikTok StitchGraph: Characterizing communication patterns on TikTok through a collection of interaction networks Mads Høgenhaug, Marcus Friis, Morten Pedersen, Luca Rossi Abstract We present TikTok StitchGraph: a collection of 36 36 graphs based on TikTok stitches. With its rapid growth and widespread popularity, TikTok presents a compelling platform for study, yet given its video-first nature the network structure of the conversations that it hosts remains largely unexplored. Leveraging its recently released APIs, in combination with web scraping, we construct graphs detailing stitch relations from both a video- and user-centric perspective. Specifically, we focus on user multi-digraphs, with vertices representing users and edges representing directed stitch relations. From the user graphs, we characterize common communication patterns of the stitch using frequent subgraph mining, finding a preference for stars and star-like structures, an aversion towards cyclic structures, and directional disposition favoring in- and out-stars over mixed-direction structures. These structures are augmented with sentiment labels in the form of edge attributes. We then use these subgraphs for graph-level embeddings together with Graph2Vec, we show no clear distinction between topologies for different hashtag topic categories. Lastly, we compare our StitchGraphs to Twitter reply networks and show that a remakable similarity between the conversation networks on the two platforms. Introduction In the 21st century, the landscape of public discourse has undergone a significant transformation with the rise of short-form video content platforms such as TikTok, Instagram Reels, and YouTube Shorts. Among these, TikTok has seen a significant increase in popularity, becoming one of the most popular social media services, with more than 1.67 billion active monthly users worldwide ( Howarth 2024 ) . In particular, teens have heavily adopted this new format type ( Pew Research 2023 ) . With the recent introduction of the TikTok Research API ( TikTok 2023 ) , performing large scale studies of TikTok has become a possibility. The short-form video format has been adopted for all communicative purposes, with content ranging from comedic sketches to serious political debates. This multifaceted nature of TikTok’s content makes it a compelling platform to study. The shift from traditional text-dominated social media, such as Facebook and X (formerly known as Twitter), to short-form video content presents new challenges in computationally analyzing public discourse. While natural language processing methods have sufficed for older platforms, the multimodal character of short-form video content introduces analytical challenges that require a combination of visual, auditory, and textual analysis. This paper presents TikTok StitchGraph - a network of stitches - : possibly the first graph dataset detailing stitches on TikTok. A stitch network is constructed comprising 36 different hashtags, containing all stitches created in May 2024 using one of said hashtags. The dataset aims to explore and improve understanding of how people communicate using stitches on TikTok. The topological structure of TikTok stitch networks and the content of individual videos are analyzed to uncover patterns and insights into this emerging form of public discourse. Additionally, these findings are compared to Twitter to highlight structural differences and commonalities in how users interact across these two social media platforms. The approach combines basic network analysis, sentiment analysis, frequent subgraph mining, and network embeddings to gain an introductory understanding of the stitch networks collected from TikTok. Specifically, this paper aims to answer the following research questions: • How can stitch communication on TikTok be studied through the lens of network analysis? • What stitch patterns are characteristic of TikTok? • How do the stitch patterns vary between different content themes on TikTok? • How do the discovered properties of TikTok stitch networks compare with Twitter reply networks? TikTok and stitches TikTok users consume the platfomr through a personalized video feed, curated by TikTok’s algorithm. TikTok presents videos to users, and based on user engagement, it adapts the recommended content. Users can interact with the content they see in a variety of ways: liking it, commenting on it, re-posting it or via more complex interactions. Among the most interesting forms there are those based on the remix of the original video content: duets and stitches . Duets enable users to create collaborative videos in a side-by-side or picture-in-picture format, facilitating interaction through real-time reactions or joint performances. Stitches , introduced in in 2020 , allows users to create a ”reaction” videos in response to another video, incorporating up to five seconds of the original - stitched - video. This content remixing functionality yields an explicit network structure, where content can have direct connections to other content. This is akin to X’s reply and quote functionality, where posts can be explicit reactions to other posts. Importantly, and differently from what happens on other platforms, users cannot stitch a video that is already a stitch of another video. Moreover, TikTok also gives users control over this feature, allowing them to disable stitching for specific videos or for all of their content. Data and Materials To create a graph of TikTok stitches we started by using TikTok’s API to collect all the videos posted with one of 36 selected hashtags that have been selected to cover a diverse set of issues 1 1 1 The hashtags were chosen based on three criteria: comparable size, topic/community focus, and predominantly English language content. These are guiding criteria, but were not strongly enforced in our selection. The size criterion is there to limit the impact of size differences on the analysis of TikTok graphs. The topic/community focus criterion means that the hashtag should be centered on a topic, event, or community of people. The intent is to find hashtags where a conversation occurs such that users communicate using stitches. Lastly, the English language criterion guarantees that we can work with the auditory component of the videos, facilitating NLP analysis. The hashtags are: #comedy, #booktok, #anime, #storytime, #lgbt, #palestine, #gaming, #maga, #football, #catsoftiktok, #news, #trump2024, #kpop, #makeup, #gaza, #dogsoftiktok, #gym, #israel, #learnontiktok, #movie, #challenge, #blacklivesmatter, #science, #conspiracy, #election, #watermelon, #biden2024, #asmr, #minecraft, #prochoice, #tiktoknews, #plantsoftiktok, #abortion, #climatechange, #jazz, #guncontrol . The data collection focused on videos created in May 2024 . Since the API does not provide direct support for identifying stitches, we leverage TikTok’s default behavior of automatically including ”#stitch with @ username ” with an hyperlink to the source video. This allowed us to scrape the original stitched video. Once the stitch relationships are collected, metadata must be collected via TikTok API. At the end of this process, we have constructed a graph structure with the edge list scraped from TikTok and vertex metadata gathered through the TikTok API. The result is a dataset composed of 36 different graphs, with video metadata and audio data. While many many forms of data augmentation could be done we worked purely with speech content. Specifically, we extracted the sentiment of each video, adding an extra dimension to the graph data. Sentiment analysis is chosen because it captures the emotional tone of speech, offering a meaningful feature that aligns with our goal of enriching the graph without overcomplicating the task. To this end, we first use the Python library scenedetect ( Breakthrough 2024 ) , which detects scene changes by analyzing variations in the HSL (Hue, Saturation, Lightness) color space to detect the scene shift that transitions from the stitchee part to

Read the complete text