Feature Transformations for Outlier Detection in Classification of Text Documents

In this paper, we investigate the influence of feature transformation on the results of outlier detection of text documents. We tested four outlier detection methods: Local Outlier Factor, Extreme Value Machine, Weibull-calibrated SVM, and the Mahalanobis distance. The analyzed text documents are represented by different feature vectors ranging from TF-IDF, through averaged word embedding (two types), to document embedding generated by the BERT network. Experimenting on two different text corpora, we show how a transformation of the feature space (vector representation of documents) influences the outlier detection results.