Feature Engineering in the Transformer Era: A Controlled Study on Toxic Comment Classification

Detecting toxic language in user-generated text remains a critical challenge due tolinguistic nuance, evolving expressions, and severe class imbalance. While Transformer-basedmodels have established state-of-the-art performance, their significant computational costs posescalability barriers for real-time moderation. We investigate whether integrating social andcontextual metadata—such as user reactions and platform ratings—can bridge the performancegap between computationally efficient classical models and modern deep learningarchitectures. Using a 40,000-comment subset of the Jigsaw Toxic Comment ClassificationChallenge, we conduct a controlled, two-phase comparison. We evaluatea Baseline configuration (TF-IDF for classical ensembles vs. raw text for ALBERT) againstan Enhanced configuration that fuses text representations with explicit social signals. Ourinvestigation analyzes whether these high-fidelity metadata features allow lightweight models(e.g., LightGBM) to rival the discriminative power of deep Transformers. The findingschallenge the prevailing assumption that deep semantic understanding is strictly necessary forhigh-performance toxicity detection, offering significant implications for the design of scalable,"Green AI" moderation systems.