Machine Learning Models Stacking in the Malicious Links Detecting

Abstract

An analysis of the performance of various classifiers on address and network groups of features was performed. A new classification model is proposed, which is a stacking of 3 models: kNN, XGBoost and Transformer. The best model for stacking was experimentally determined: Logistic Regression, which made it possible to improve the result of the best available model by 3%. The hypothesis that stacking a larger number of worse models has an advantage over stacking a smaller number of more productive models on the used data set was confirmed: regardless of the choice of stacking meta-algorithm, stacking of three models showed better results than stacking two.