Source-linked AI summary
Unbiased Rectification for Sequential Recommender Systems Under Fake Orders
Qiyu Qin, Yichen Li, Haozhao Wang, Cheng Wang, Rui Zhang, Ruixuan Li
TL;DR
Fake orders embedded in genuine user histories can manipulate recommendations, while rectification must distinguish harmful samples from useful information. DITaR combines dual-view detection with influence-guided filtering and targeted gradient ascent; experiments report improved recommendation performance, efficiency, and robustness over state-of-the-art methods.
Problem
Fake orders embedded in genuine user histories can distort recommendations, and rectification must identify harmful samples without discarding beneficial information or disrupting sequential data integrity.
Source: Introduction · pp. 1–2Method
DITaR detects suspicious orders from collaborative and semantic view discrepancies, then uses influence estimates to select harmful samples for targeted gradient-ascent rectification.
Source: Introduction · p. 2Results
Experiments on three datasets report that DITaR improves recommendation performance and computational efficiency over state-of-the-art methods while ensuring robustness.
Source: Introduction · p. 2Takeaways & Limitations
DITaR selectively rectifies harmful samples while preserving beneficial information, data volume, and sequence structure without retraining.
Source: Targeted Rectification with Gradient… · p. 5, Conclusion · p. 7
Abstract
from arXiv · showhide
Fake orders pose increasing threats to sequential recommender systems by misleading recommendation results through artificially manipulated interactions, including click farming, context-irrelevant substitutions, and sequential perturbations. Unlike injecting carefully designed fake users to influence recommendation performance, fake orders embedded within genuine user sequences aim to disrupt user preferences and mislead recommendation results, thereby manipulating exposure rates of specific items to gain competitive advantages. To protect users' authentic interest preferences and eliminate misleading information, this paper aims to perform precise and efficient rectification on compromised sequential recommender systems while avoiding the enormous computational and time costs of retraining existing models. Specifically, we identify that fake orders are not absolutely harmful - in certain cases, partial fake orders can even have a data augmentation effect. Based on this insight, we propose Dual-view Identification and Targeted Rectification (DITaR), which primarily identifies harmful samples to achieve unbiased rectification of the system. The core idea of this method is to obtain differentiated representations from collaborative and semantic views for precise detection, and then filters detected suspicious fake orders to select truly harmful ones for targeted rectification with gradient ascent. This ensures that useful information in fake orders is not removed while preventing bias residue. Moreover, it maintains the original data volume and sequence structure, thus protecting system performance and trustworthiness to achieve optimal unbiased rectification. Extensive experiments on three datasets demonstrate that DITaR achieves superior performance compared to state-of-the-art methods in terms of recommendation quality, computational efficiency, and system robustness.
Introduction
Fake orders are manipulated interactions embedded in genuine user sequences that can distort recommendations and user trust. DITaR identifies suspicious orders, estimates which are harmful, and rectifies those selectively; experiments report improved recommendation performance and efficiency with robustness.
Introduction
- Fake orders are covert interactions embedded in genuine user sequences that disrupt preferences and steer recommendations toward merchant-preferred items.Examples include click farming, context-irrelevant semantic substitutions, and sequential perturbations; unlike fake-user attacks, they manipulate specific recommendation results.Source: Introduction · pp. 1–2
- Fake orders’ artificial nature makes it difficult for them to remain consistent across collaborative and semantic dimensions, creating detectable cross-view discrepancies.The paper contrasts these attacks with prior methods that inject fake users to degrade overall system performance.Source: Introduction · p. 1
- Existing rectification should remove truly harmful samples while preserving beneficial information, but sequential dependencies, sample-wise integrity, and varied fake-order impacts make this difficult.Simple removal can disrupt data integrity because fake orders replace genuine interactions without changing data volume.Source: Introduction · p. 2
- DITaR significantly improves recommendation performance and computational efficiency over state-of-the-art methods while ensuring robustness, according to experiments on three datasets.The paper also reports that DITaR outperforms state-of-the-art methods across multiple evaluation metrics.Source: Introduction · p. 2
- DITaR identifies suspicious orders using collaborative and semantic view discrepancies, then applies influence-guided filtering and gradient ascent to rectify harmful samples.This sample-level approach is designed to retain useful information while avoiding bias residue.Source: Introduction · p. 2
Related Works
Related work covers sequential recommendation, manipulation of genuine user histories, and data rectification. The paper situates DITaR as a framework combining collaborative and semantic identification with influence-based targeted rectification.
Sequential Recommendation
- Sequential recommendation research models dynamic interests and temporal dependencies, using attention-based models, multimodal features, and semantic embeddings from large language models.Examples include SASRec and BERT4Rec.Source: Related Works · p. 2
User History Manipulation
- Prior work documents manipulation of genuine histories through web injection and item substitution; this paper groups three embedded manipulation types as fake orders.The paper frames these histories as complex across collaborative patterns, semantic associations, and temporal dependencies.Source: Related Works · p. 2
Rectification Methods
- Rectification methods include influence functions, gradient-based and Shapley-based approaches, clustering, and graph methods; this paper targets harmful items in sequential recommendation.Prior work also addresses noisy-label correction and large-language-model lifecycles.Source: Related Works · p. 2
DITaR Framework
- DITaR combines collaborative and semantic views to identify fake orders, then uses influence functions to filter items and apply gradient ascent to harmful ones.The framework has Dual-view Identification and Targeted Rectification components.Source: Figure 2 · p. 3
Methodology
DITaR identifies fake orders through complementary collaborative and semantic views, then uses influence estimates to select harmful samples for targeted rectification. Its two modules combine multi-signal detection with gradient ascent and clean-data regularization.
Methodology
- DITaR has a Dual-view Identification module and a Targeted Rectification stage that filters items by influence before applying gradient ascent to harmful fake orders.The approach is intended to avoid retraining costs while retaining data volume and sequence structure.Source: Methodology · p. 3, Targeted Rectification with Gradient… · p. 4
Problem Formulation
- Sequential recommendation predicts a user’s next item from historical interactions; direct inverse-Hessian computation is infeasible for large models, so IHVP approximates the needed product.The approximation reduces computational complexity using recursion depth J and parameter dimension p.Source: Problem Formulation · p. 3
Dual-view Identification Module
- Cross-view inconsistencies and behavioral anomalies jointly identify suspicious orders, with temporal smoothing and a validation-tuned threshold stabilizing per-position decisions.Signals include representation and prediction disagreement, popularity deviation, and local sequential-pattern disruption.Source: Dual-view Identification Module · p. 4
Targeted Rectification with Gradient Ascent
- Positive influence indicates that removing a suspicious interaction decreases clean-validation loss, while negative influence suggests a beneficial augmentation effect.The influence estimate uses the suspicious-sample gradient, Hessian, and average clean-validation gradient.Source: Targeted Rectification with Gradient… · p. 5
- Rectification targets only confirmed harmful interactions, then applies a smaller clean-data gradient-descent step to preserve normal recommendation modeling.The method alternates updates and monitors validation performance to prevent overfitting.Source: Targeted Rectification with Gradient… · p. 5
Experiment
Across three datasets, DITaR is compared with existing rectification methods for recommendation quality, efficiency, detection, and robustness. The experiments also examine how fake-order types affect performance and how DITaR’s components contribute to rectification.
Experiment Setup
- Experiments use ML-1M, Amazon-Beauty, and Yelp2018, with repetitive, semantic, and sequential fake orders replacing genuine interactions without changing sequence length.Evaluation includes HR@10/20 and NDCG@10/20, convergence epochs, and robustness; the methods are tested on SASRec, GRU4Rec, and BERT4Rec.Source: Experiment Setup · p. 5, Experiment · p. 5
Performance Overall
- DITaR consistently outperforms baseline methods, performs comparably to retraining, and occasionally surpasses original clean-data performance.The performance gap between retraining and original data grows more severe as fake-order frequency and intensity increase.Source: Table 2 · p. 6, Performance Overall · p. 6
- DITaR requires significantly fewer convergence epochs than baselines, using gradient-ascent fine-tuning on pretrained models rather than training from scratch.Sharding efficiency can decline as shard count increases, creating a trade-off with accuracy.Source: Table 3 · p. 6, Performance Overall · pp. 6–7
- Across two sequential recommendation models and datasets, DITaR achieves high detection precision and strong recall, with occasional false positives or missed detections.The detected candidates provide input for subsequent targeted rectification.Source: Figure 3 · p. 7, Performance Overall · p. 7
- Fake-order effects differ by type: semantic orders degrade performance, while sequential swaps can improve it and moderate repetition can reinforce interaction patterns.The impact magnitude increases with user ratio and intensity.Source: Figure 4 · p. 7, Performance Overall · p. 7
- Removing either detection view reduces final rectification effectiveness, with a larger impact from removing the collaborative view; influence filtering prevents ascent on beneficial or neutral candidates.The ablation study evaluates Amazon-Beauty across SASRec, GRU4Rec, and BERT4Rec.Source: Table 4 · p. 7, Performance Overall · p. 7
Conclusion
The conclusion presents DITaR as an efficient rectification framework for sequential recommenders under fake orders. It combines dual-view detection with selective rectification and reports better performance than state-of-the-art methods across multiple metrics.
Conclusion
- DITaR identifies fake orders through cross-view discrepancies and uses influence functions to selectively rectify genuinely harmful samples while preserving beneficial information.The approach performs sample-level rectification without retraining or altering data structures.Source: Conclusion · p. 7
- The framework is designed for efficient, unbiased rectification of sequential recommender systems under fake orders.Source: Conclusion · p. 7
- Comprehensive experiments show that DITaR significantly outperforms state-of-the-art methods across multiple evaluation metrics.Source: Conclusion · p. 7