Unsupervised feature learning for spam email filtering

An excessive number of features may negatively affect the performance of a learning classifier. In addition, the computational time for processing the data during the training process may be prolonged. Therefore, a preprocessing stage that includes feature extraction and feature reduction processes in the field of machine learning is a vital role for speeding up computation and improving classification accuracy. The problem considered in this study is related to data transformation, prior to machine learning classifiers. Feature representation that preserves class separability with lower dimensional space for identifying spam is being proposed. The major advantage regarding the proposed feature representation is its robustness that enables classifiers like Random Forest, Support Vector Machines, and the decision tree C4.5 to identify an incoming email as spam or non-spam where the feature size is very small with a good generalization irrespective of the data source. (C) 2019 Elsevier Ltd. All rights reserved.