Abstract
<title>Abstract</title> <p>As online social networks (OSNs) play an increasingly influential role in political communication, accurately detecting political stance from noisy, short-form content remains a significant challenge in computational social science (CSS). This paper presents a comparative multi-feature framework for political stance detection during the 2024 U.S. presidential election, focusing on candidate affiliation toward Donald Trump and Joe Biden-Kamala Harris. Using nearly 100,000 posts from X, we introduce a candidate affiliation feature that extends beyond coarse sentiment polarity and leverages topic modeling to support scalable annotation and thematic characterization. Under a unified preprocessing, labeling, and evaluation pipeline, we compare classical machine learning models, sentence-embedding-based neural architectures, and a domain-adapted transformer. Results show that appropriate preprocessing and task-aligned text representation often exert greater influence on performance than model complexity alone. While fine-tuned BERTweet achieves the strongest in-domain performance, carefully engineered classical models, particularly linear support vector machines, remain highly competitive while offering substantial advantages in computational efficiency, interpretability, and reproducibility. Multi-feature and ablation analyses further indicate that expanded metadata provides only modest and inconsistent gains beyond strong text representations. Overall, this work offers practical guidance for designing scalable, interpretable, and resource-efficient political stance detection systems while emphasizing the importance of cautious, deployment-aware evaluation under dataset-specific conditions.</p>