Improving the Accuracy of Spam Message Filtering using Hybrid CNN Classification

Spam messages are growing day by day due to the invention of low-cost messaging and emailing solutions. Due to this, the identification of genuine messages from the spammy ones requires a lot of learning. This learning includes training the system for spam messages, and then training another system for non-spam or genuine messages. Once these systems are trained, then a probabilistic classifier is needed, which can find out the probability of the message to either be spam or genuine. Such a network is called as two-stage convolutional neural network. In this paper, we have designed a two-stage convolutional neural network, that first trains one network with spam messages, and then trains another network with non-spam messages. These stages are cascaded, and the outputs of each stage is given to a decision unit. The unit evaluates the probabilities of spam and non-spam messages, and finally classifies the input text into either spam or non-spam. The proposed system is tested on the standard UCI spam text dataset, and it has achieved more than 90% accuracy for classification of spam messages.