The proliferation of music streaming platforms has created an urgent need for automatic explicit content detection to protect vulnerable audiences, particularly children, from inappropriate material. This study presents a comprehensive comparative analysis of machine learning approaches for classifying explicit content in song lyrics using a dataset of 108,138 songs from Spotify. We evaluate traditional machine learning models (Decision Tree, Logistic Regression, Naive Bayes, SVM) with two distinct text representation methods: TF-IDF vectorization and pre-trained LyricsBERT embeddings. Additionally, we assess the impact of incorporating musical metadata and acoustic features alongside textual data. Our experimental framework compares feature combinations ranging from lyrics-only to comprehensive multi-modal representations, including one that excludes lyrical data entirely. Results demonstrate that traditional models with TF-IDF vectorization achieve competitive performance, with Decision Tree achieving the highest F1-score of 0.7896 when combining textual, numerical, and categorical features. Surprisingly, metadata-only approaches using Logistic Regression achieved superior performance (F1-score: 0.8715), suggesting that acoustic and categorical features may be more reliable indicators of explicit content than lyrics alone. Deep learning approaches using BERT fine-tuning achieved lower performance (F1-score: 0.7046) at a significantly higher computational cost. These findings challenge the assumption that lyrical content is the primary determinant of explicitness, providing practical insights for developing scalable content moderation systems on streaming platforms.