<![CDATA[The rapid growth of social media, particularly the X (formerly Twitter) platform, has made it a primary space for public discourse in Indonesia, yet its openness has also become a systematic loophole for the spread of hate speech that threatens social cohesion. Conventional text-based detection models fail to capture hidden linguistic nuances such as slang and regional euphemisms, and they disregard the social dimension of hate propagation reflected in user interaction patterns. This study proposes a hybrid IndoBERT-GAT architecture that integrates textual semantic representation with social graph structure into a single unified framework. The research method employed 16,646 government-related tweets collected via Apify crawling, labeled through a hybrid approach (manual annotation and pseudo-labeling), then represented through IndoBERT embeddings for textual features and a Graph Attention Network to model a heterogeneous tweet-user graph, before being combined via a feature fusion mechanism and evaluated through an ablation study across four model scenarios. The results show that the full Hybrid IndoBERT-GAT model achieved the highest test F1-score (65%), outperforming the same architecture without metadata (61.8%), the GAT+metadata-only baseline (42.2%), and the MLP+metadata baseline (39.9%), demonstrating that IndoBERT's semantic representation is the dominant contributor while graph structure and metadata serve as complementary signals, although the model still tends to overpredict hate speech in politically sarcastic content that uses sharp language without genuinely hateful intent.]]>
