The collusive fraudulent behaviors on e-commerce platforms lead to proliferation of fraudulent reviews, which disrupt fair competition among merchants and mislead consumers’ shopping decisions. Detection of spammer groups helps purify the e-commerce environment and enhances consumers’ shopping experience. However, existing graph-based methods for detecting spammer groups first learn user node vector representations from the graph, and then use clustering methods to obtain candidate groups. Such separate two-stage detection methods are difficult to obtain high-quality candidate groups, resulting in suboptimal detection performance. Additionally, current graph construction methods used in spammer group detection do not fully consider the characteristics of spammer groups, which limits the detection performance. Aiming these concerns, we integrate heterogeneous graph attention network (HGAN) with label propagation (LP) for detecting spammer groups. First, we build a heterogeneous weighted directed (HWD) graph by analyzing the dataset and assign an initial label to each node. Then, we integrate a HGAN-module with an LP-module to obtain the HWD graph’s node embeddings and simultaneously generate candidate groups. We enhance the quality of embeddings and groups through the collaborative optimization between the predicted labels obtained from the HGAN-module and the pseudo-labels obtained from the LP-module. Finally, we calculate the suspiciousness values of groups using the reconstruction loss of the autoencoder for spammer group identification. Experiments conducted on real-world review datasets, including Amazon, Yelp, and YelpChi, demonstrate that our method achieves significant improvements in average Precision@k and Recall@k metrics compared with state-of-the-art baseline approaches.

Abstract Group shilling attacks are more threatening than individual shilling attacks due to the collusive behaviours among group members, which pose a great challenge to the credibility of recommender systems. Detection of group shilling attacks can reduce the risk caused by such attacks and ensure the credibility of recommendations. The existing methods for detecting group shilling attacks mainly extract features from the rating patterns of users at group level to measure the shilling behaviours of groups. However, they may become ineffective with the change of attack strategy, resulting in a decrease in detection performance. Aiming at this problem, a new solution based on user multi-dimensional features and collusive behaviour analysis is presented for detecting group shilling attacks. First, we employ the information entropy and latent semantic analysis to analyse the user behavioural patterns from dimensions of item, rating, time and interest, and propose a suite of indicators to measure the anomaly behaviours of users. Second, we propose a measure based on the multi-dimensional features of users to capture the collusion of group members from the perspective of their synchronized behaviours and abnormal behaviours, and treat the groups with high collusion as candidate groups. Finally, based on the multi-dimensional features of users, we construct the user behaviour similarity matrix using Gaussian radial basis function (Gaussian-RBF) and adopt the spectral clustering algorithm to spot group shilling attackers in the candidate groups. Experiments show that the detection performance (F1-measure) of the proposed method can achieve 0.965, 0.964, 0.991 and 0.868 on the Netflix, CiaoDVD, Epinions and Amazon datasets, respectively, which is better than that of state-of-the-art methods.

Abstract As an extension of conventional top-K item recommendation solution, bundle recommendation has aroused increasingly attention. However, because of the extreme sparsity of user-bundle (UB) interactions, the existing top-K item recommendation methods suffer from poor performance when applied to bundle recommendation. While some graph-based approaches have been proposed for bundle recommendation, these approaches primarily leverage the bipartite graph to model the UB interactions, resulting in suboptimal performance. In this paper, a dual hypergraph contrastive learning model is proposed for bundle recommendation. First, we model the direct and indirect UB interactions as hypergraphs to represent the higher-order UB relations. Second, we utilize the hypergraph convolution networks to learn the user and bundle embeddings from the hypergraphs, and improve the learned embeddings through a bidirectional contrastive learning strategy. Finally, we adopt a joint loss that combines the InfoBPR loss supporting multiple negative samples and the contrastive losses to optimize model parameters for prediction. Experiments on the real-world datasets indicate that our model performs better than the state-of-the-art baseline methods.

Detecting collusive spammers who collaboratively post fake reviews is extremely important to guarantee the reliability of review information on e-commerce platforms. In this research, we formulate the collusive spammer detection as an anomaly detection problem and propose a novel detection approach based on heterogeneous graph attention network . First, we analyze the review dataset from different perspectives and use the statistical distribution to model each user's review behavior. By introducing the Bhattacharyya distance , we calculate the user-user and product-product correlation degrees to construct a multi-relation heterogeneous graph. Second, we combine the biased random walk strategy and multi-head self-attention mechanism to propose a model of heterogeneous graph attention network to learn the node embeddings from the multi-relation heterogeneous graph. Finally, we propose an improved community detection algorithm to acquire candidate spamming groups and employ an anomaly detection model based on the autoencoder to identify collusive spammers. Experiments show that the average improvements of precision@k and recall@k of the proposed approach over the best baseline method on the Amazon , Yelp_Miami, Yelp_New York, Yelp_San Francisco, and YelpChi datasets are [13%, 3%], [32%, 12%], [37%, 7%], [42%, 10%], and [18%, 1%], respectively.

Label propagation-based overlapping community detection algorithms have been widely used in complex networks due to their simplicity and efficiency. However, such algorithms need to randomly choose neighbor nodes and do not fully take the network’s topology into consideration, resulting in low stability and accuracy. Aiming at this problem, we propose an overlapping community detection approach based on DeepWalk and the improved label propagation. We first use the DeepWalk model to learn the network’s topology to obtain low-dimensional vector representations that reflect the spatial location of nodes and construct the weight matrix through vector dot product operation. Then, we design a label propagation algorithm with a preference selection strategy, which can obtain stable overlapping communities by exchanging information with fixed neighbors on the basis of preserving the nodes’ own labels. The experimental results on the real network and synthetic datasets show that the proposed approach has better accuracy and stability than the baseline methods.

Information security is one of the key issues in e-commerce Internet of Things (IoT) platform research. The collusive spamming groups on e-commerce platforms can write a large number of fake reviews over a period of time for the evaluated products, which seriously affect the purchase decision behaviors of consumers and destroy the fair competition environment among merchants. To address this problem, we propose a network embedding based approach to detect collusive spamming groups. First, we use the idea of a meta-graph to construct a heterogeneous information network based on the user review dataset. Second, we exploit the modified DeepWalk algorithm to learn the low-dimensional vector representations of user nodes in the heterogeneous information network and employ the clustering methods to obtain candidate spamming groups. Finally, we leverage an indicator weighting strategy to calculate the spamming score of each candidate group, and the top-k groups with high spamming scores are considered to be the collusive spamming groups. The experimental results on two real-world review datasets show that the overall detection performance of the proposed approach is much better than that of baseline methods.

To protect recommender systems against shilling attacks, a variety of detection methods have been proposed over the past decade. However, these methods focus mainly on individual features and rarely consider the lockstep behaviours among attack users, which suffer from low precision in detecting group shilling attacks. In this work, we propose a three-stage detection method based on strong lockstep behaviours among group members and group behaviour features for detecting group shilling attacks. First, we construct a weighted user relationship graph by combining direct and indirect collusive degrees between users. Second, we find all dense subgraphs in the user relationship graph to generate a set of suspicious groups by introducing a topological potential method. Finally, we use a clustering method to detect shilling groups by extracting group behaviour features. Extensive experiments on the Netflix and sampled Amazon review datasets show that the proposed approach is effective for detecting group shilling attacks in recommender systems, and the F1-measure on two datasets can reach over 99 percent and 76 percent, respectively.

Collaborative recommender systems are vulnerable to shilling attacks. To address this issue, many methods including supervised and unsupervised have been proposed. However, supervised detection methods require training classifiers and they only apply to detect known types of attacks. The existing unsupervised detection methods need to know the prior knowledge of attacks, otherwise they suffer from low detection precision. In this paper, we present BS-SC, an unsupervised approach for detecting shilling profiles, which does not need to know the attack size or to label the candidate spammers. BS-SC starts from an in-depth analysis of user behaviors and uses two key mechanisms (i.e., behavior features extraction and behavior similarity matrix clustering) to distinguish shilling profiles from genuine ones. The behavior features reflect the behavior difference between genuine and shilling profiles, and the behavior similarity matrix clustering is to cluster shilling profiles based on their highly similar behaviors. Experimental results on the MovieLens and the sampled Amazon review datasets indicate that BS-SC outperforms the baseline unsupervised approaches, even when the prior knowledge is given for them.

Existing shilling attack detection approaches focus mainly on identifying individual attackers in online recommender systems and rarely address the detection of group shilling attacks in which a group of attackers colludes to bias the output of an online recommender system by injecting fake profiles. In this article, we propose a group shilling attack detection method based on the bisecting K-means clustering algorithm. First, we extract the rating track of each item and divide the rating tracks to generate candidate groups according to a fixed time interval. Second, we propose item attention degree and user activity to calculate the suspicious degrees of candidate groups. Finally, we employ the bisecting K-means algorithm to cluster the candidate groups according to their suspicious degrees and obtain the attack groups. The results of experiments on the Netflix and Amazon data sets indicate that the proposed method outperforms the baseline methods.

Collaborative filtering (CF) recommender systems have been shown to be vulnerable to shilling attacks. How to quickly and effectively detect shilling attacks is a key challenge for improving the quality and reliability of CF recommender systems. Although many recent studies have been devoted to detecting shilling attacks, there are still problems that require further discussion, especially the improvement of the detection performance on real-world unlabelled datasets. In this work, we propose an unsupervised approach that exploits item relationship and target item(s) for attack detection. We first extract behaviour features based on the item relationship. Then, we distinguish suspicious users from normal users and construct a set of suspicious users. Finally, we identify target item(s) by analysing the aggregation behaviour of suspicious users, based on which we detect attack users from the set of suspicious users. Extensive experiments on the MovieLens 100K dataset and sampled Amazon review dataset demonstrate the effectiveness of the proposed approach for detecting shilling attacks in recommender systems.

Collaborative recommender systems have been known to be extremely vulnerable to shilling attacks. To prevent such attacks, many detection approaches including supervised and unsupervised have been proposed. However, the supervised approaches are only suitable for detecting known types of attacks and the unsupervised approaches require a priori knowledge to ensure the detection performance. To address the limitations, the authors propose an unsupervised approach for detecting shilling attacks based on user rating behaviours. They first use Gibbs latent Dirichlet allocation model to extract latent topics of user preferences from user rating item sequences, then they use mixture transition distribution model to construct the user's preference model and present several metrics to capture the diversity between genuine and attack users in rating behaviours. In the case of unknown attack size, the number of attack users is obtained by analysing the critical point of rating behaviour suspicious degrees between genuine and attack users, and based on which the attack users are identified. The experimental results on the MovieLens 1 M dataset show that the proposed approach outperforms the baseline methods in terms of recall and precision metrics.

Due to the openness of the recommender systems, the attackers are likely to inject a large number of fake profiles to bias the prediction of such systems. The traditional detection methods mainly rely on the artificial features, which are often extracted from one kind of user-generated information. In these methods, fine-grained interactions between users and items cannot be captured comprehensively, leading to the degradation of detection accuracy under various types of attacks. In this paper, we propose an ensemble detection method based on the automatic features extracted from multiple views. Firstly, to collaboratively discover the shilling profiles, the users’ behaviors are analyzed from multiple views including ratings, item popularity, and user-user graph. Secondly, based on the data preprocessed from multiple views, the stacked denoising autoencoders are used to automatically extract user features with different corruption rates. Moreover, the features extracted from multiple views are effectively combined based on principal component analysis. Finally, according to the features extracted with different corruption rates, the weak classifiers are generated and then integrated to detect attacks. The experimental results on the MovieLens, Netflix, and Amazon datasets indicate that the proposed method can effectively detect various attacks.

To defend recommender systems, various methods have been proposed to detect shilling profiles, which can be categorised as user‐ and item‐based detection methods. Most of the user‐based methods identify shilling profiles via statistical signatures of rating values and suffer from low precision when detecting different types of attacks. Most of the item‐based methods use temporal information to detect the anomaly items, but they assume that the fake ratings were injected in short periods. So they are invalid for the long duration and decentralised injection attacks. To address these limitations, the authors extract the multidimensional profile temporal features and present a shilling detection method. First, from the user profile view, user rating behaviours are characterised by corrected conditional entropy and the dissimilarity with the rest‐rating model. Second, from the item profile view, the user features are extracted according to item temporal popularity. Third, the features based on weighted deviation from dynamic mean are extracted according to the fact that the items mean changes with time. Finally, support vector machine is exploited to detect shilling profiles based on the proposed features. Experimental results on the Netflix dataset indicate that the performance of the proposed method is better than that of the benchmark methods.

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.

Abstract Collaborative filtering systems are vulnerable to shilling attacks in which malicious users bias the systems' recommendation output by inserting fake profiles. While many approaches have been proposed to detect shilling attacks, they suffer from low precision. To solve this problem, an ensemble method for detecting shilling attacks based on ordered item sequences is proposed. Firstly, by analyzing the differences of rating patterns between genuine and attack profiles, we construct ordered popular item sequences and ordered novelty item sequences, and based on which, the popular and novelty item rating series are constructed for each user profile. Secondly, we propose six features to characterize the attack profiles. Particularly, we extract two features based on the popular and novelty item rating series. We partition the item set according to the ordered item sequences and combine them with mutual information to extract another four features. Finally, we propose an ensemble framework to detect shilling attacks. In particular, we create base training sets with great diversities using bootstrap resampling technique. Based on these base training sets, we train decision tree algorithm to generate diverse base classifiers. The simple majority voting strategy is used to combine the predictive results of these base classifiers. Experimental results indicate that ensemble method for detecting shilling attacks based on ordered item sequences can significantly improve the precision while maintaining a high recall. Copyright © 2015 John Wiley & Sons, Ltd.

The existing collaborative recommendation algorithms based on matrix factorisation (MF) have poor robustness against shilling attacks. To address this problem, in this study the authors propose a robust collaborative recommendation algorithm based on kernel function and Welsch reweighted M‐estimator. They first propose a median‐based method to calculate user and item biases, which can reduce the influence of shilling attacks on user and item biases because median is insensitive to outliers. Then, they present a method of similarity computation based on kernel function, which can obtain the information of similar users by non‐linear inner product operation. Finally, they combine the user and item biases based on median and the similarity based on kernel function with MF model, and introduce the Welsch reweighted M‐estimator to realise the robust estimation of user feature matrix and item feature matrix. The experimental results on the MovieLens dataset show that the proposed algorithm outperforms the existing algorithms in terms of both recommendation accuracy and robustness, and the improvement of its robustness is not at the expense of recommendation accuracy.

The existing supervised approaches suffer from low precision when detecting profile injection attacks. To solve this problem, the authors propose an ensemble detection model by introducing back propogation (BP) neural network and ensemble learning technique. Firstly, through combination of various attack types, they create base training sets which include various samples of attack profiles and have great diversities with each other. Secondly, they use the created base training sets to train BP neural networks to generate diverse base classifiers. Finally, they select parts of the base classifiers which have the highest precision on the validation dataset and integrate them using voting strategy. Uncorrelated misclassifications generated by each base classifier can be successfully corrected by the ensemble learning. The experimental results on two different scale of the real datasets MovieLens and Netflix show that the proposed model can effectively improve the precision under the condition of holding a high recall.

Personalized support for users becomes even more important, when service access takes place in open and dynamic service-oriented environment. This paper shows how to realize personalized service support in cross-system/service environments based on ontology and Web service technologies. First, we introduce the related approaches for supporting cross-system personalization and give their insufficiency respectively. Aimed at the problems we propose a Web service based architecture for cross-system personalization. The loosely coupled structure of Web service can easily integrate personalization service from various information systems and provide seamless access to the users. In order to reuse user models we also present an ontology based user model to support cross-system personalization. Compared with the existing approaches, our approach can effectively support the various existing personalized systems and user models, and the realization of cross-system personalization is more simply and efficiently