Toward an efficient fuzziness based instance selection methodology for intrusion detection system

Building a high quality classifier is one of the key problems in the field of machine learning (ML) and pattern recognition. Many ML algorithms have suffered from high computational power in the presence of large scale data sets. This paper proposes a fuzziness based instance selection technique for the large data sets to increase the efficiency of supervised learning algorithms by improving the shortcomings of designing an effective intrusion detection system (IDS). The proposed methodology is dependent on a new kind of single layer feed-forward neural network (SLFN), called random weight neural network (RWNN). At the first stage, a membership vector corresponding to every training instance is obtained by using RWNN for computing the fuzziness. Secondly, the training instances (along with their fuzziness values) according to the actual class labels are grouped separately. After this, the instances having low fuzziness values in each group are extracted, which are used to build a reduced data set. The instances outputted by the proposed method are used as an input for ML classifiers, which result in reducing the learning time and also increasing the learning capability. The proposed methodology exhibits that the reduced data set can easily learn the boundaries between class labels. The most obvious finding from this study is a considerable increase in the accuracy rate with unseen examples when compared with other instance selection method, i.e., IB2. The proposed method provides the better generalization and fast learning capability. The reasonability of the proposed methodology is theoretically explained and experiments on well known ID data sets support its usefulness.