Cost-Aware Benchmarking of Classical and Deep Learning Models for Twitter Sentiment Analysis

Sentiment analysis of Twitter data has progressed significantly with the adoption of deep learning and transformer-based architectures. However, many existing studies emphasize accuracy improvements without systematically analyzing computational efficiency or ensuring controlled experimental comparison. This paper presents a cost-aware benchmarking study of classical machine learning models (Logistic Regression and Support Vector Machines) and deep learning architectures (CNN, BiLSTM, BiGRU, and DistilBERT) for binary sentiment classification on English-language Twitter data. All models were evaluated under identical preprocessing conditions using balanced subsets of the Sentiment140 dataset to ensure fair comparison. Experimental results demonstrate that DistilBERT achieves the highest accuracy (80.02%), outperforming recurrent networks (~76.6%) and classical baselines (~76%). However, transformer-based inference introduces significantly higher computational cost compared to traditional approaches. The findings reveal that performance gains from transformer models are moderate in short-text classification and must be evaluated alongside efficiency considerations. This study provides a reproducible, resource-efficient experimental framework and offers practical insights for model selection in latency-sensitive deployment scenarios.