Multiview Ensemble Method for Detecting Shilling Attacks in Collaborative Recommender Systems

Abstract

Faced with the evolving attacks in collaborative recommender systems, the conventional shilling detection methods rely mainly on one kind of user-generated information (i.e., single view) such as rating values, rating time, and item popularity. However, these methods often suffer from poor precision when detecting different attacks due to ignoring other potentially relevant information. To address this limitation, in this paper we propose a multiview ensemble method to detect shilling attacks in collaborative recommender systems. Firstly, we extract 17 user features by considering the temporal effects of item popularity and rating values in different popular item sets. Secondly, we devise a multiview ensemble detection framework by integrating base classifiers from different classification views. Particularly, we use a feature set partition algorithm to divide the features into several subsets to construct multiple optimal classification views. We introduce a repartition strategy to increase the diversity of views and reduce the influence of feature order. Finally, the experimental results on the Netflix and Amazon review datasets indicate that the proposed method has better performance than benchmark methods when detecting various synthetic attacks and real-world attacks.

Detecting Shilling Attacks with Automatic Features from Multiple Views

“…When the two strategies are applied in the basic attacks, user random shifting attack, user average shifting attack, target random shifting attack, and target average shifting attack can be deployed. To facilitate the representation, we define the shifting attack as a collection with the same portion of user random shifting attack, user average shifting attack, target random shifting attack, and target average shifting attack [3].…”

“…After that, the latent factors were used as features, and decision tree was used as a classifier algorithm to detect attackers. In [3], based on the rating values and item popularity fused with temporal information, 17 artificial detection features were extracted. Then, these features were divided into several subsets by a feature set partition algorithm to construct multiple optimal classification views.…”

“…Different from [2], we extract the detection features from multiple views including rating values, item popularity, and interaction between users. Unlike the methods based on the human engineering features [3,5,7,16,19,20], we use SDAEs to automatically extract detection features under different corruption rates.…”

“…In convolutional layer, the number of feature maps is set to 6. (7) MV-EDM: A method based on 17 artificially designed features, which take the temporal effects of item popularity and rating values into consideration [3]. Moreover, a multiview ensemble algorithm is used as a classifier method.…”

“…However, the existing features are extracted by human engineering from one piece of usergenerated information (e.g., ratings, timestamps, and item popularity), which mainly focus on some specific types of attacks and need high knowledge costs. Due to the diversity and variability of attackers' strategies, it is very difficult to fully characterize the shilling profiles from one piece of usergenerated information (i.e., single view) [3,4]. Moreover, since the number of fake profiles is rare relative to the genuine ones, the detection of shilling attacks can be seen as an imbalanced classification problem [5].…”