IKN Public Opinion on TikTok Before and After Efficiency Policy: CNN-LSTM on Imbalanced Data

Growing polarization in Ibu Kota Nusantara (IKN) stems from conventional sentiment analysis tools’ inability to decode TikTok’s contextual complexities, particularly multimodal sarcasm and vernacular-policy relationships (e.g., mangkrak for project cancellations). This study develops a policy-aware hybrid model (CNN-BiLSTM + Policy Knowledge Graph) to decode TikTok’s multimodal sarcasm and vernacular-policy links (e.g., mangkrak), enabling: youth sentiment quantification post-IKN’s 73.3% budget cuts, social criticism-socio-political reality mapping, and evidence-based interventions mitigating Global South strategic project polarization. Using the Knowledge Discovery in Databases framework, we analyzed 2,950 high-engagement TikTok comments (≥10 interactions) from verified accounts (@Polindo.id and @geraldvincentt) across two periods: pre-policy (June-August 2024) and post-policy (January-March 2025). Methodologically, slang normalization, stemming, and minority-class weighting (15×) preceded classification via a CNN-BiLSTM architecture integrated with Policy Knowledge Graphs. Results showed an 18.88% reduction in negative sentiment (83.2%-8.7%), model accuracy of 94.13% (AUC-PR 0.91), and strong correlations between vernacular terms (e.g., mandek [stagnation]) and policy outcomes (r = -0.89; p < 0.01), with investor asing mentions surging 463% post-policy. These validate deep learning-enabled social listening for real-time policy diagnostics, with implications for fiscal transparency dashboards, algorithmic bias mitigation, and context-driven policy communication prioritizing vulnerable groups in SDG infrastructure governance.