Comparison of binary classifier and outlier detection in different equilibrium

Abstract

In the fields of financial risk control and mechanical production, the data sets with abnormal problems are always extremely unbalanced, because the most of abnormal problems occur hardly. Using this unbalanced data set to train the binary classifier, the result is often not ideal. Although there are Ensemble Learning and Grid Search methods to improve the F1 and accuracy of the classifier, in order to simplify the model, it is better to regard this financial risk control problem as an outlier detection problem than a binary classification problem. This paper uses the public data set on Kaggle, and compares the performance of the commonly used binary classification algorithms including Bayesian, Decision Tree, Random Forest, Logistic Regression Classifier, K-Nearest Neighbor (KNN), AdaBoost, One-Class SVM, Isolation Forest and Local Outlier Factor on balanced and unbalanced data sets respectively. According to the experimental results, this paper find that Bayesian is more suitable when the data set is small. Random Forest is more suitable for balanced data. For medium and large data sets with extremely unbalanced data, the effect of using One-Class SVM is better and more stable, and the effect of stable model is more important than that of unstable one.