Improved detection of email and SMS spam using a hybrid deep and shallow learning model

This study addresses the critical challenge of detecting sophisticated spam in digital communications by developing an advanced hybrid machine learning framework. The primary objective is to enhance detection accuracy through the synergistic integration of traditional and deep learning approaches, thereby overcoming the limitations of existing single-method systems. Our methodology combines shallow linguistic features (BoW, TF-IDF, lexical statistics) with deep contextual embeddings from a fine-tuned BERT model, processed through a novel Deep-Shallow Hybrid Model (DSHM) architecture. The system employs a multilayer perceptron for feature fusion and a Gradient Boosting meta-learner to ensemble predictions from five baseline classifiers (GBAM, RL-SPAM, HMAM, SSHM, MSSL) with DSHM outputs. The research uses the SMS Spam Collection v.1 dataset, containing 5,574 messages, to evaluate the proposed framework under a realistic imbalanced spam distribution. Data collection followed standard preprocessing protocols, including tokenization, normalization, and stratified sampling to maintain dataset balance. Experimental results demonstrate that the proposed hybrid ensemble consistently outperforms the individual baseline models on the SMS Spam Collection v.1 dataset. The final ensemble achieves an accuracy of 98.8%, a ROC-AUC of 0.993, and an F1-score of 0.956 for the spam class while maintaining a low false positive rate (0.0058). The proposed Deep-Shallow Hybrid Model (DSHM) also demonstrates strong standalone performance, confirming that combining shallow linguistic features with contextual BERT embeddings substantially improves spam detection robustness. These findings indicate that the proposed hybrid feature-fusion and ensemble-learning strategy provides an effective and scalable solution for real-world spam detection.