PaperlayerView exact revision on arXiv

Source-linked AI summary

GCN-Based User Representation Learning for Unifying Robust Recommendation and Fraudster Detection

Shijie Zhang, Hongzhi Yin, Tong Chen, Quoc Viet Nguyen Hung, Zi Huang, Lizhen Cui

arXiv:2005.10150v1cs.SIcs.IR

TL;DR

Shilling attacks undermine recommender systems by injecting misleading feedback, motivating methods that jointly preserve recommendation quality and identify fraudsters. GraphRfi uses coupled GCN-based representation learning for both tasks, and experiments report superior performance and robustness across real-world evaluations.

  • Problem

    Shilling attacks can manipulate recommendations and undermine the assumption that collected ratings faithfully represent user preferences.

  • Method

    GraphRfi jointly learns user preference and reliability with an end-to-end GCN framework coupling robust recommendation and fraudster detection.

  • Results

    GraphRfi outperforms state-of-the-art baselines on robust rating prediction and fraudster detection, with minimal impact across attack types.

  • Takeaways & Limitations

    The experiments support GraphRfi as an effective and practical unified approach for recommendation and fraudster detection under shilling attacks.

  • Takeaways & Limitations

    The formulation targets predicting ratings that non-malicious users would give to unrated items in the presence of fraudsters and shilling attacks.

Abstract

from arXiv · show

In recent years, recommender system has become an indispensable function in all e-commerce platforms. The review rating data for a recommender system typically comes from open platforms, which may attract a group of malicious users to deliberately insert fake feedback in an attempt to bias the recommender system to their favour. The presence of such attacks may violate modeling assumptions that high-quality data is always available and these data truly reflect users' interests and preferences. Therefore, it is of great practical significance to construct a robust recommender system that is able to generate stable recommendations even in the presence of shilling attacks. In this paper, we propose GraphRfi - a GCN-based user representation learning framework to perform robust recommendation and fraudster detection in a unified way. In its end-to-end learning process, the probability of a user being identified as a fraudster in the fraudster detection component automatically determines the contribution of this user's rating data in the recommendation component; while the prediction error outputted in the recommendation component acts as an important feature in the fraudster detection component. Thus, these two components can mutually enhance each other. Extensive experiments have been conducted and the experimental results show the superiority of our GraphRfi in the two tasks - robust rating prediction and fraudster detection. Furthermore, the proposed GraphRfi is validated to be more robust to the various types of shilling attacks over the state-of-the-art recommender systems.

1 INTRODUCTION

Online recommender systems help users navigate information overload, but shilling attacks can corrupt ratings and degrade recommendation quality. GraphRfi unifies robust recommendation and fraudster detection through coupled GCN-based user representations.

  • Motivation: Recommender systems predict unrated-item ratings from historical behavior to generate personalized recommendations.They help users choose among increasingly large numbers of online products.
  • Problem: Shilling attacks inject fake feedback that can manipulate recommendations and make learned user or item representations biased and unreliable.Fraudsters may promote or defame targeted items through unfair ratings.
  • Research gap: Existing research separately improves recommendation robustness or detects and removes fraudsters, leaving their complementary integration largely unexplored.Robust methods may rely on restrictive assumptions, while removal can mistakenly exclude genuine users.
  • Approach: GraphRfi is an end-to-end GCN-based framework with coupled components for robust recommendation and fraudster detection.Its GCN models user-item interactions while integrating predefined fraudster-indicative user features.
  • Approach: The framework learns user representations that jointly capture preference and reliability information from local user-item graph neighborhoods.GCN aggregation combines node features with graph topology during representation learning.
  • Evaluation: GraphRfi is evaluated for recommendation accuracy, recommendation robustness, and fraudster detection accuracy, with experiments reporting superiority across these objectives.The contribution summary identifies all three evaluation dimensions.

2 GRAPHRFI: THE MODEL

GraphRfi jointly models robust rating prediction and fraudster detection with GCN-based user representations that combine behavioral features and rating-graph structure. Its attention-based aggregation and differentiable fraudster component support end-to-end coupling between the two tasks.

  • User behavioral statistics initialize GCN user nodes, allowing representations to capture preference and reliability information from local user-item neighborhoods.Features include rating, activity, helpfulness, timing, feedback-length, and sentiment statistics.
  • GraphRfi formulates robust rating prediction and fraudster detection as two jointly learned tasks.
  • Rating-specific embeddings are combined with item representations to pass edge-specific messages before aggregating information into updated user representations.The aggregation uses the user node and rated items as the relevant neighbor set.
  • Attention is incorporated into the aggregation function so interacted items can contribute unequally to the learned user representation.The design accounts for varied characteristics and contributions of different interacted items.
  • Prediction errors are appended to user embeddings as fraudster-detection features, because deviations between ratings and predictions can signal malicious users.The enhanced vector concatenates the updated user representation with the mean squared rating-prediction error.
  • A differentiable neural random forest enables end-to-end fraudster classification, while a joint loss balances rating and fraudster-detection objectives.The fraudster probability guides rating optimization, and predicted ratings provide an auxiliary fraudster-detection feature.

3 EXPERIMENTS

The experiments evaluate GraphRfi on fraudster detection and robust rating prediction using Yelp and Movies & TV, with research questions covering performance, attack robustness, and hyper-parameters.

  • Experiments evaluate GraphRfi on fraudster detection and robust rating prediction using the Yelp and Movies & TV datasets.
  • RQ1 compares rating prediction with and without fraudsters, while RQ2 compares detection of different fraudster types against baselines.
  • RQ3 examines GraphRfi’s recommendation robustness for each specific shilling-attack type.
  • RQ4 studies how hyper-parameters affect GraphRfi’s performance across the two tasks.
  • Yelp labels users with fake reviews as fraudsters and users without fake reviews as genuine.
  • Movies & TV labels users with at least 20 votes as benign above 0.7 helpful-vote proportion and fraudulent below 0.3.

3.2 Evaluation Protocols

The evaluation measures rating-prediction accuracy and robustness by combining MAE/RMSE assessment with progressively injected fraudsters in the training data.

  • MAE and RMSE measure rating-prediction accuracy, with smaller values indicating better accuracy.
  • Robustness testing begins with models trained only on genuine ratings, then progressively adds fraudsters to observe performance fluctuations.
  • The protocol randomly assigns 20% of genuine ratings to testing and uses the remainder as the initial training set.
  • GraphRfi uses five depth-three decision trees, while embedding size and λ are selected from predefined search sets.
  • The rating-prediction component uses a hidden layer of size 100, ReLU activation, and three-layer neural components.

3.4 Baselines

The baselines span fraudster detection, robust recommendation, matrix factorization, collaborative filtering, autoencoding, and graph-based interaction modeling.

  • Recommendation baselines: Recommendation baselines include robust matrix factorization, graph auto-encoding, social recommendation, autoencoding, probabilistic matrix factorization, item-based filtering, and matrix factorization.
  • Recommendation baselines: RCF uses robust M-estimators in matrix factorization to resist shilling attacks.
  • Recommendation baselines: GCMC models the bipartite interaction graph through message passing in a graph auto-encoder framework.
  • Recommendation baselines: GraphRec contributes its user-item modeling component because user-user interactions are unavailable in the evaluation setting.
  • Fraudster-detection baselines: OFD combines an autoencoder with a deep neural random forest and is described as the state-of-the-art fraudster-detection method.
  • Fraudster-detection baselines: Fraudster-detection comparisons include REV2, DegreeSAD, FAP, and OFD.
  • Evaluation displays: Figure 2 reports rating-prediction results on Yelp and Movies & TV, where spam profiles refer to dataset fraudsters.
  • Evaluation displays: Table 3 reports fraudster-detection performance on Yelp and Movies & TV.

3.5 Prediction Performance (RQ1)

GraphRfi consistently outperforms the comparison recommenders in rating prediction, including when shilling attacks are introduced, while GCN-based models generally perform strongly.

  • 0.8%, 1.1%, 6.5%, 10.9%, 13.9% and 15.5% improvements are reported for GraphRfi over MF on Yelp RMSE as spam profiles increase.
  • GraphRfi significantly and consistently outperforms all baselines on rating prediction, including under shilling attacks.
  • GraphRec and GCMC generally outperform other comparison methods on both datasets.
  • MF performs best among non-GCN models on Yelp with 0%, 20% and 40% fraudsters, whereas AutoRec leads at 60%, 80% and 100%.
  • RCF outperforms the other four non-GCN models on Movies & TV.

3.6 Fraudster Detection (RQ2)

The study evaluates fraudster detection on original mixed-attack datasets and simulated attack-specific datasets. GraphRfi shows superior detection performance across these scenarios.

  • Attack settings: The experiments cover random, average, and hate attacks, using original datasets for mixed real-life scenarios and simulated datasets for attack-specific evaluation.Target items receive manipulated ratings alongside filler-item ratings to create stronger recommendation bias.
  • Detection results: On Yelp-Random and Yelp-Hate, GraphRfi has slightly lower Recall than OFD but significantly higher Precision and F1 Scores.This indicates stronger performance on the latter two reported detection metrics in those settings.
  • Baseline comparison: OFD improves over DegreeSAD by 16% in Precision, 21% in Recall, and 18% in F1 Score on average across both datasets.The comparison is used to highlight the advantages of neural decision trees for fraudster detection.
  • Detection results: GraphRfi accurately detects fraudsters under both type-specific shilling attacks and mixed real-life scenarios.The reported superiority applies across the evaluated dataset constructions.

3.7 Robustness Performance (RQ3)

GraphRfi is evaluated by training and testing recommenders under varying proportions and types of injected fraudsters. It shows minimal impact from attacks and consistently accurate rating predictions.

  • Experimental setting: The robustness experiments train and evaluate recommenders on datasets containing random, average, or hate attacks.The proportion of inserted fraudsters varies from 0% to 100% in increments of 20%.
  • Robustness results: GraphRfi shows minimal impact from all tested attack types and generates consistently accurate predicted ratings.The authors identify it as the most robust model in the comparison.
  • Mechanism and comparison: The NRF classifier bonds the impact of fraudsters, preventing the recommender from being corrupted by shilling attacks.Hybrid models are generally more robust, while traditional collaborative-filtering approaches are highly sensitive to fraudulent ratings.
  • Baseline comparison: GraphRec and GCMC show less severe performance fluctuations as fraudster prevalence increases than other methods.Their predicted ratings account for external domain knowledge such as side information in addition to collaborative filtering.

3.8 Impact of Hyper-parameters (RQ4)

The study examines GraphRfi’s sensitivity to embedding dimension e and parameter λ on Yelp using rating-prediction and fraudster-detection metrics.

  • Parameter settings: The experiments vary embedding dimension e across 50, 100, 150, 200, and 250, and λ across 1, 3, 5, 7, and 9.Results are plotted with RMSE and MAE for rating prediction and Precision, Recall, and F1 Score for fraudster detection.
  • Sensitivity analysis: Figure 4 presents performance differences across the tested settings for both rating prediction and fraudster detection.The passage states that the best results for both tasks are identifiable from the parameter-sensitivity analysis.

4 RELATED WORK

Related work covers collaborative-filtering recommenders, graph-convolutional approaches, robust recommendation under shilling attacks, and behavior- or network-based fraudster detection.

  • Recommender systems: Collaborative filtering recommends items using activity histories from similar users or similar items.Recent work also applies autoencoders and graph convolutional networks for recommendation and feature extraction.
  • Robust recommendation: Conventional recommenders treat input data equally, an assumption challenged when shilling attacks introduce fraudulent ratings.Robust methods such as RCF modify matrix-factorization optimization with M-estimators to improve attack resistance.
  • Fraudster detection: Fraudster detection research commonly distinguishes behavior-based methods from network-based methods.Behavior-based approaches extract features from review text, while online fake feedback motivates the need to assess review reliability.

5 CONCLUSION

GraphRfi jointly optimizes robust rating prediction and fraudster detection through coupled user representation learning. On two real-world datasets, it outperformed state-of-the-art baselines on both tasks.

  • GraphRfi jointly optimizes robust rating prediction and fraudster detection in an end-to-end user representation learning model.Its graph convolutional network captures user preference and node side information, while neural random forests support fraudster detection.
  • The fraudster classifier and rating prediction system mutually enhance one another during learning.The classifier is expected to discourage large shifts in rating prediction, while accurate rating prediction benefits fraudster detection.
  • On two real-world datasets, GraphRfi outperformed state-of-the-art baselines on both robust rating prediction and fraudster detection.
Loading 2005.10150v1…