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

TL;DR

TikTok StitchGraph introduces a first-of-its-kind dataset of $36$ stitch-based graphs collected from TikTok in May $2024$, capturing both video- and user-centric perspectives and enriching edges with sentiment labels. The authors apply frequent subgraph mining and graph embedding (Graph2Vec and Bag-Of-Subgraphs) to characterize stitch communication, finding prevalent star-like motifs, scarce cycles, and limited topology-to-theme separation. A cross-platform comparison with Twitter suggests similar structural tendencies driven by graph size, rather than topic category, highlighting fundamental patterns in short-form video discourse. The work also addresses ethical considerations and provides a data-collection framework using public APIs and scraping, contributing a foundation for future multimodal network analyses on TikTok. Overall, StitchGraph offers a scalable, interpretable view of how stitches shape conversations, with implications for understanding platform-specific discourse and cross-platform dynamics.

Abstract

We present TikTok StitchGraph: a collection of 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.

Figures (5)

• Figure 1: The figure shows the cyclic subgraphs identified within the largest connected components of user graphs, ranging from a triangle to an octagon. The numbers beneath denote the support of the subgraph in TikTok and Twitter in that order.
• Figure 2: The figure shows the frequent undirected subgraphs identified in TikTok user graphs using gSpan on the largest connected component. The lines between subgraphs illustrate their hierarchical relationships, where each subgraph extends from the one above. The numbers are the supports of the given subgraphs in TikTok and Twitter respectively. Simpler structures, such as dyads $S_1$ ($support=36$), form the foundation, while more complex patterns like chains and star-like structures emerge as extensions with lower support. This highlights how TikTok stitch patterns evolve, often around central hubs or sequential interactions.
• Figure 3: The figure shows the frequent undirected subgraphs identified in TikTok user graphs using gSpan on the largest connected component. The lines between subgraphs illustrate their hierarchical relationships, where each subgraph extends from the one above. The numbers are the supports of the given subgraphs in TikTok and Twitter respectively. Simpler structures, such as dyads $S_1$ ($support=36$), form the foundation, while more complex patterns like chains and star-like structures emerge as extensions with lower support. This highlights how TikTok stitch patterns evolve, often around central hubs or sequential interactions.
• Figure 4: Comparison of user graphs across three embedding spaces, labeled by assigned hashtag categories, clustered using HDBSCAN, and subsequently dimensionality reduced with UMAP. They reveal no separation between graphs corresponding to different categories. This observation is further reinforced by the minimal overlap between the assigned categories and the clusters identified by HDBSCAN.
• Figure 5: Comparison of user graph embeddings from TikTok and Twitter across three embedding spaces, dimensionality reduced with UMAP. While no distinct separation is observed for most graphs, larger Twitter graphs appear to occupy a slightly different region compared to TikTok graphs in the Graph2Vec embedding space. No such distinction is evident in either Bag-Of-Subgraphs.