Impact of outlier detection on classification performance in sentiment analysis

Nowadays, people spend a significant amount of time to expressing their emotions and thoughts through various online platforms, as a result of the widespread use of the internet. Social media, where people can freely and easily express their opinions and views, provides the opportunity to find out about customers' positive and negative thoughts regarding brands and products. In this study, the effect of outliers on sentiment analysis was examined using Amazon's "Pet Supplies" and IMDB data sets. Within the scope of this work, bidirectional encoder representations from transformers (BERT) model was used as the basis model, and the outlier detection Methods including local outlier factor (LOF), one class support vector machine (OCSVM), and isolation forest (IForest) were applied. Outliers were examined using four different strategies: removed from only training data, removed from both training data and test data, marking outliers by adding a token begining of the text, and replacing them with a fixed expression. In addition, the effect of t-distributed stochastic neighbor embedding (t-SNE) and principal component analysis (PCA) dimensionality reduction methods on the outlier detection process was also evaluated in the strategy where only outliers were eliminated from the training data. The obtained results indicate that the proposed method contributes to classification accuracy; however, this contribution appears to be limited under certain conditions.