Optimizing Sentiment Classification on IKN Development: A Comparative Study of TF-IDF, Word2Vec, and FastText Embeddings

Abstrak

The relocation of the National Capital City (IKN) has instigated significant polarization of public discourse on YouTube, presenting substantial challenges for sentiment analysis due to the high variability of non-standard linguistic patterns. Existing scholarship, how-ever, has frequently overlooked the bias inherent in accuracy metrics within imbalanced datasets, while also neglecting the critical alignment between feature characteristics and algorithmic geometry. To address these methodological limitations, this study conducts a comparative analysis of TF-IDF, Word2Vec, and FastText feature extraction techniques applied to Naïve Bayes, Support Vector Machine (SVM), and Random Forest algorithms, utilizing a dataset of 3441 comments. Empirical results demonstrate that the synergy be-tween SVM and FastText yields the most robust performance, achieving an accuracy of 85.2% and outperforming other model combinations. To mitigate the bias in accuracy metrics due to class imbalance, this study further incorporates Precision, Recall, and F1-Score as additional evaluation metrics, with SVM + FastText achieving a Precision(+) of 89.7%, Recall(+) of 83.2%, and F1(+) of 86.3%. These findings underscore the efficacy of margin maximization on semantic vectors over conventional probabilistic approaches in processing informal text, thereby offering precise insights for policymakers regarding the social legitimacy of the IKN development.

Identifier

• OpenAlex: https://openalex.org/W7164905127
• DOI: https://doi.org/10.3390/engproc2026137020