Suicidal thoughts and behaviors are an urgent public health concern, underscoring the need for effective tools to enable early detection of suicide risk. We address this challenge by developing robust machine learning models that classify Reddit posts into four distinct suicide risk severity levels. Framing this as a multi-class classification task, we propose a RoBERTa–TF-IDF–PCA Hybrid model that integrates deep contextual embeddings from Robustly Optimized BERT Approach (RoBERTa) with statistical term-weighting from TF-IDF, whose features are reduced via Principal Component Analysis (PCA) to enhance accuracy and stability. To mitigate data imbalance and overfitting, we explore a range of data resampling and data augmentation strategies to improve model generalization. We compare our hybrid approach against RoBERTa-only, BERT, and traditional machine learning classifiers. Experimental results demonstrate that our model can achieve improved performance, giving a best weighted \(F_{1}\) score of 0.7512.