An adjustable combination of linear regression and modified probabilistic neural network for anti-spam filtering

Email is a commonly used tool for communication which allows rapid and asynchronous communication. The growing popularity and low cost of e-mails have made spamming an extremely serious problem today. Several anti-spam filtering techniques have been developed but most of them suffer from low accuracy and high false alarm rate due to complexity and changing nature of unsolicited messages. This study proposes an innovative classification framework with comparable accuracy, affordable computation and high system robustness. In particular, an effective feature selection scheme is implemented in conjunction with an adjustable combination of linear and nonlinear learning algorithms. Extensive experiments have indicated that the proposed framework compares favorably to other state-of-the-art methods, especially when misclassification cost is high.