SMS Spam Detection using NLP and Deep Learning Recurrent Neural Network Variants

The ubiquity of Short Messaging Services (SMS) has been accompanied by an upsurge in spam messages, necessitating effective detection mechanisms. This paper delves into the intricate task of SMS spam detection using a Kaggle dataset of 5570 samples. As initial step, the challenge of class imbalance was addressed by applying oversampling technique ADASYN. Later, TF-IDF embedding technique was employed to capture semantic meaning and contextual information within SMS messages. A conventional ML classifier, Random Forest has achieved a good accuracy of 92.3%. Later, various RNN architectures, including Simple RNN, LSTM, Bi-LSTM and Gated Recurrent Unit (GRU), applied for SMS spam detection. LSTM variants demonstrated superior performance with 93% accuracy for RNN, 98% for LSTM and 98.2%, 98.3% for BI-LSTM and GRU. To further enhance the accuracy, a hybrid architecture combining Bi-LSTM and GRU layers was explored, resulting in improved accuracy of 99%. The proposed model outperformed traditional ML approaches for SMS spam detection. This paper sheds the efficacy of LSTM variants in addressing the SMS spam detection challenge. Emphasizing the significance of handling class imbalances and employing effective embedding techniques, the achieved accuracies hold promising implications.