An Effective Concept Drift Detection Technique with Kernel Extreme Learning Machine for Email Spam Filtering

The increase in the number of undesirable emails named spam has posed a major requirement to develop a highly dependent and robust antispam filters. This paper presents a novel email spam filtering technique with the capability of adapting with the dynamic environment. Concept drift detector attempts to determine the position of the concept drift in large data stream for replacing the baseline learner next to the modifications in the data distribution and therefore enhances accuracy. The proposed method detects the concept drift depending upon the computation of variation in the email content distribution using Statistical Test of Equal Proportions (STEPD) technique. The STEPD is a simpler commonly available model that identifies the concept drift with respect to a hypothesis test among two proportions. The SPEPD technique is used to determine the criteria of the concept drift for all unknown emails that assist the filtering technique in the recognition of the occurrence of the spam. In addition, the kernel extreme learning machine (KELM) based classification model is applied to classify the instances into two class labels namely spam and non-spam correspondingly. The experimental results of the STEPD-KELM model are tested against Enron dataset and the results are examined interms of distinct aspects. The experimental values indicated that the STEPD-KELM model has resulted to a maximum precision of 93.78%, recall of 96.54%, and accuracy of 95.33%.