Detecting Coordinated Behaviour on Video-First Platforms: The Challenge of Multimodality and Complex Similarity on TikTok

TL;DR

This paper tackles the problem of detecting coordinated online behaviour on video-first platforms, specifically TikTok, where multimodal content and complex similarity complicate traditional coordination detection. It introduces a multilayer network framework in which each layer encodes a distinct co-action type across different modalities (visual frames, audio transcripts, music, descriptions, URLs, hashtags), coupled with carefully tuned, layer-specific similarity functions. The method is validated on German politics surrounding the 2024 European Elections, revealing cross-modal coordination signals and highlighting heterogeneous coordination patterns across layers, including cross-layer account participation and music-driven alignment. Limitations arising from TikTok's API and platform design are discussed, along with ethical considerations and avenues for future refinement, such as expanding similarity measures and cross-layer pattern analyses to improve robustness and reduce false positives.

Abstract

Research on online coordinated behaviour has predominantly focused on text-based social media platforms. However, the rise of video-first platforms such as TikTok introduces distinct challenges. The multimodal nature of video posts, combining visuals, audio, and text, allows for coordination across various modalities and complicates comparison between posts. This paper proposes an approach to detecting coordination that addresses these characteristic challenges. Our methodology, based on multilayer network analysis, is tailored to capture coordination across multiple modalities, and explicitly handles complex forms of similarity inherent in video and audio content. We test this approach on German political posts regarding the 2024 European Elections retrieved via the TikTok Research API. Our results demonstrate the ability of our approach to identify coordination within the constraints of the API, while also critically highlighting potential pitfalls and limitations.

Figures (5)

• Figure 1: Examples of videos that are considered equivalent in the Audio Layer, thanks to our fine-tuned similarity function
• Figure 2: Visualisation of the different filtered layers of the network
• Figure 3: Examples of pairs of videos found in the same clusters using clique percolation
• Figure 4: Examples of videos from two clusters found in the Video Description and Hashtag Sequence layers using CP(4,2)
• Figure 5: Examples of videos from two clusters found with GLouvain