Analysis of X and Threads Responses Based on Single Keywords Using Graph Neural Network

Abstract

This study analyzes user responses on the social media platforms X and Threads regarding radicalism, based on the single keyword "radicalism." Differing interaction characteristics between the two platforms motivate a comparison of network structure using a graph-based approach and Graph Neural Networks (GNN). Data were collected through scraping of public content on X and Threads, with 5,000 raw posts each, yielding 1,327 reply interactions on X and 1,273 on Threads. Research stages included data collection, preprocessing, construction of the user-post graph, network metric analysis, and implementation of a Graph Autoencoder with a Graph Convolutional Network (GCN) encoder to generate node embeddings. The resulting graphs comprised 1,471 nodes and 1,223 unique edges for X, and 1,491 nodes and 1,081 unique edges for Threads, with X showing a denser structure (density 0.000565; average degree 1.663) than Threads (density 0.000487; average degree 1.450), while Threads was more fragmented (412 weak components versus 299 on X). The Graph Autoencoder was evaluated via link prediction using AUC and Average Precision (AP): X achieved AUC 0.5397 and AP 0.5879, slightly above the random-guessing baseline, while Threads achieved AUC 0.4987 and AP 0.5507, indicating a structure harder to reconstruct due to fragmentation. These quantitative results reinforce the network-metric findings that X forms a more connected network while Threads fragments into smaller groups. Practically, the findings offer an empirical basis for a decision-support system monitoring radicalism-related discourse, favoring dominant-cluster monitoring on X and parallel, cross-cluster monitoring on Threads. This study does not aim to detect or label accounts or content as radical, but to analyze interaction patterns and network characteristics of user responses.