<![CDATA[This study aims to analyze class-level model behavior under metric disagreement in imbalanced multi-label Indonesian emotion classification, using the divergence between Macro F1 and Micro F1 as a diagnostic signal rather than a mere performance indicator. A machine-translated Indonesian version of the GoEmotions dataset, comprising approximately 58,000 samples across 28 fine-grained emotion categories, is used as the experimental setting. The translated dataset was not manually revalidated, and findings are scoped to this translated GoEmotions setting. Two transformer-based models are evaluated: IndoBERT, a monolingual Indonesian model, and DistilBERT, a multilingual model, both fine-tuned with class-specific threshold optimization. The results reveal opposing divergence patterns: IndoBERT achieves higher Micro F1 than Macro F1, indicating performance concentrated on high-frequency classes, while DistilBERT exhibits the reverse pattern, suggesting broader but less precise label activation. Per-class analysis further shows that most minority classes consistently fall into unstable or non-functional performance regimes across both models. This study concludes that aggregate metrics alone are insufficient for evaluating model behavior in imbalanced multi-label settings. A behavior-oriented interpretation framework for Macro–Micro F1 divergence and a regime-based class reliability categorization are proposed to support more structured and informative evaluation practices.]]>
