Semi-supervised recommendation attack detection based on Co-Forest

摘要

In recommendation attack, malicious users attempt to bias the recommendation results by injecting fake profiles into the rating database. To detect such attack, three types of methods, i.e., unsupervised, supervised and semi-supervised, have been proposed. Among these works, the advantage of semi-supervised methods is that they can use the unlabeled user profiles to improve the detection performance. However, the existing semi-supervised methods suffer from low precision. Aiming at this problem, in this paper, we propose a semi-supervised detection approach named SSADR-CoF based on the Co-Forest algorithm. Being different from the existing semi-supervised methods which only use a few of features to train a single classifier for the detection, the proposed approach uses a series of features to train an ensemble of classifiers to detect the recommendation attack. We first use the window dividing and rating behavior statistical methods to extract a series of user rating behavior mode features for training the detection model. Then, we use a small number of labeled user profiles to initialize an ensemble of classifiers, and use the ensemble of classifiers to assign labels to the unlabeled user profiles. Finally, we use the labeled and the newly labeled user profiles to iteratively update the classifiers for the detection. Experiments conducted on three benchmark datasets MovieLens 10M, MovieLens 25M, and Amazon show that the proposed approach can effectively improve the precision of the semi-supervised methods under the condition of maintaining high recall and AUC.