Gradient Boosting-Based Predictive Click Fraud Detection Using Manifold Criterion Variable Elimination

Abstract

Online advertising models are vulnerable to click fraud, which occurs when an individual or a group repeatedly clicks on an online advertisement with the intent to generate illegitimate clicks and make money from the advertiser. In machine learning-based approaches for detecting click fraud, the performance of the models can be affected by the presence of collinear, redundant, and least significant features in the dataset. These types of features can lead to overfitting, where the model becomes too complex and fails to generalize well to new data. Therefore, a Manifold Criterion Variable Elimination method is proposed in this work to select significant features utilizing the potential of six filter-based feature selection techniques for the discrimination of fraud and genuine publishers. Experimentations are conducted on the online advertisement user click dataset in two modes, first considering all extracted features and second considering only selected features. An extraction of 103 statistical features from the user-click dataset is performed for each class instance labelled with OK, Fraud and Observation. The Manifold Criterion Variable Elimination method selects the top 15 most relevant features. Individual and ensemble learning models are trained with selected feature-set and tuned parameter values. The performances of learners are evaluated using standard evaluation measures. The results demonstrated that, in general, the performance of all learners improved with the selected feature set. Particularly, the Gradient Tree Boosting (GTB) ensemble model performed superiorly by improving the weak learners by minimizing the model's loss via merging the weak learners into a strong one iteratively.