Uncovering anomalous rating behaviors for rating systems

Abstract

Personalization collaborative filtering recommendation plays a key component in online rating systems, which also suffers from profile injection attacks in reality. Although anomalous rating detection for online rating systems has attracted increasing attention in recent years, detection performance of the existing methods has not reached an end. Eliminating the impact of interfering information on anomaly detection is a crucial issue for reducing false alarm rates. Moreover, detecting anomalous ratings for unlabeled and real-world data is always a big challenge. In this paper, we investigate a two-stage detection framework to spot anomalous rating profiles. Firstly, interfering rating profiles are determined by comprehensively analyzing the distributions of user activity, item popularity and special ratings in order to eliminate sparse ratings. Based on the reserved rating profiles, combining target item analysis and non-linear structure clustering is then adopted to further determine the concerned attackers. Extensive experimental comparisons in diverse attacks demonstrate the effectiveness of the proposed method compared with competing benchmarks. Additionally, discovering interesting findings including anomalous ratings and items on two real-world datasets, Amazon and TripAdvisor, is also investigated.

Introduction

Online rating systems have been significantly developed in parallel with the social networks in the last decade. Rating data is ubiquitous on the well-known E-commerce websites including Amazon, Taobo, TripAdvisor, Yelp and etc. [5], [6], [11], [12], [21], [32], [48], [55], [61]. Personalization collaborative recommender systems play a crucial role in handling the increasingly prominent problem of information overload, which automatically suggest to a user items that might be of interest to her [1], [4], [23], [24], [25], [26], [27], [28], [41]. However, collaborative filtering recommender systems (CFRSs) are highly vulnerable to outside attacks, called profile injection attacks (a.k.a. shilling attacks) [6], [34], due to the fact that recommender systems are entirely based on the input provided by users or customers [3], [7], [8], [14], [16], [19], [22], [37], [44], [49], [56], [61]. Profile injection attacks, in which attackers manipulate biased ratings in order to influence future recommendations, have been demonstrated to be effective against collaborative filtering recommendation engines. According to the intention of attackers, shilling attacks can be classified in two basic categories: inserting malicious profiles which rate a particular item highly are called push attacks, conversely inserting malicious profiles aimed at downgrading the popularity of an item are termed nuke attacks [31]. Anonymous or pseudonymous users in online systems can multiply their profiles and identities nearly indefinitely, which utilize well-designed rating profiles to produce recommendation behaviors that the attackers desire. Therefore, proactively identifying the malicious rating profiles is extremely significant and meaningful for personalized collaborative recommendations.

Securing collaborative filtering recommender systems from malicious attacks have become an important issue with increasing popularity of recommender systems [13], [47]. Although previous researches have shown promising results, defending such attacks is still an unresolved technique, and has not reached a full level of performance [13], [36], [47], [57], [61], [62], [63], [64]. In particular, how to construct a strategy that can be used to spot anomalous ratings for real-world data is also extremely desired. Furthermore, developing detection method which can effectively defense diverse shilling attacks is always a big challenge. Moreover, compared with the number of genuine profiles (authentic profiles), the number of attack profiles is very small in rating systems. The distinct difference between the numbers of genuine and attack profiles is call imbalanced distribution of rating profiles [6], [53]. The imbalanced distribution makes a challenging task for abnormality detection due to the difficulty of characterizing rating behaviors of users. How to eliminate a part of genuine profiles (interfering rating profiles) and reduce imbalanced distribution before anomaly detection is a concerned task especially for large-scale and real-world data. With respect to unlabeled and real-world datasets, investigating abnormality forensics metrics for determining the concerned users or items is a realistic problem that cannot be ignored.

In this paper, we present a two-stage detection framework to spot anomalous rating profiles. Facing with the imbalanced distribution of rating profiles, interfering rating profiles are first determined by comprehensively analyzing the rating distribution of user activity, the distribution of item popularity and special ratings in order to eliminate sparse ratings. The goal of the first stage is to filter out interfering rating profiles (genuine profiles) [36], [52] as many as possible and simultaneously reserve all attack profiles. Based on the remaining rating profiles, combining target item analysis and non-linear structure clustering is then adopted to further determine the concerned attackers. Since shilling attackers mimic rating details of authentic user to manipulate attack profiles, it is difficult to identify them. A robust multiple kernel data clustering method is employed to distinguish the attack profiles from authentic profiles in an appropriate feature space while the clusters are not linear separable in the original space. Moreover, we also explore evaluation metrics of abnormality forensics for discovering interesting findings in two real-world datasets including TripAdvisor and Amazon. More importantly, analyzing the internal relationship between historical ratings and reviews of items is provided to spot anomalous items. Extensive experimental comparisons on diverse attack datasets demonstrate the effectiveness of the proposed detection method compared with competing benchmarks. In addition, discovering interesting findings including anomalous items, ratings and etc. on Amazon and TripAdvisor datasets is investigated.

The main contributions of this paper are four-fold as follows:

• Eliminating interfering profiles according to the distributions of users’ activity, items’ popularity and sparse ratings in advance provides a feasible idea for abnormality detection faced with the imbalanced distribution of rating profiles, which is also favorable to characterize rating behaviors of users.

• Combining target item analysis and non-linear structure clustering is effective to reduce the scope of determining anomalous users. The false alarm rate of the proposed approach can be further reduced.

• To discover suspicious items or ratings on unlabeled and real-world datasets, suspected items detected by the proposed approach are further determined by comprehensively analyzing intrinsic association between overall rating and each aspect rating on the same item, rating behavior aggregation, rating intention distribution and topological structure analysis of suspicious items.

• Extensive experiments on both synthetic datasets in 10 different attacks and real-world datasets including Amazon and TripAdvisor are conducted to demonstrate the effectiveness of the proposed approach.

The rest of the paper is organized as follows. Section 2 discusses related work. Section 3 introduces the proposed method in detail. In Section 4, experimental results are reported and analyzed. Finally, we briefly conclude the paper with a brief summary and discuss our future work.