Outlier Detection Model Based on SOM for Classification Problem

Abstract

In many practical classification problems, it often contains some outliers in the data set, which may affect the performance of classification model. To solve this problem, this paper combines the self-organizing mapping network (SOM), the pruning technique and the local outlier factor (LOF), constructs the outlier detection model based on SOM (SOD). Firstly, it clusters with SOM on the training set, and then obtains the new training set by pruning the clustering results. Finally, it detects the outliers by the local outlier factor of each sample on the new training set. The empirical results show that the SOD model has better detection performance compared with some existing outlier detection models, and it can improve the classification accuracy more efficiently through the models trained without the outliers.