Integrating transformer-based credibility signals into neural collaborative filtering for fake review-aware recommendation.

Integrating transformer-based credibility signals into neural collaborative filtering for fake review-aware recommendation - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sci Rep . 2026 Jun 5;16:17510. doi: 10.1038/s41598-026-56220-2 Integrating transformer-based credibility signals into neural collaborative filtering for fake review-aware recommendation Yasmeen Abdelmohsen 1, ✉ , Khaled Wassif 2 , Nagy Ramadan 1 Author information Article notes Copyright and License information 1 Department of Information Systems and Technology, Faculty of Graduate Studies for Statistical Research, Cairo University, Giza, Egypt 2 Department of Computer Science Faculty of Computers and Artificial, Intelligence Cairo University, Giza, Egypt ✉ Corresponding author. Received 2025 Dec 22; Accepted 2026 May 29; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ . PMC Copyright notice PMCID: PMC13237017  PMID: 42243240 Abstract Online recommender systems (RS) face growing trust challenges as deceptive reviews distort user feedback. Although RS optimisation and fake review detection have advanced separately, integrating credibility signals directly into recommendation training remains underexplored. This study proposes the Fake-Review-Aware Recommender System (FRARS), which embeds transformer-based deception probabilities into the training objective of a Neural Matrix Factorisation (NeuMF) model. Among several detectors, DeBERTa-v3-base performed best (ROC-AUC = 0.932 on YelpCHI, 0.921 on YelpNYC). FRARS applies these probabilities through two mechanisms: Hard Filtering removes interactions above a deception threshold, while Soft Weighting proportionally down-weights uncertain ones. We evaluate FRARS on two independent Yelp datasets—YelpCHI (67,395 reviews, ~ 49% deceptive) and YelpNYC (359,052 reviews, ~ 10% deceptive). FRARS-Soft improved NDCG@10 by 20.9% on YelpCHI and 19.5% on YelpNYC, with parallel gains in precision and recall; all transformer-based improvements were statistically significant ( p < 0.001). Detector quality and recommendation gains exhibited a significant monotonic relationship (Spearman ρ = 0.964 on YelpCHI, 0.929 on YelpNYC). These consistent results across different regions, scales, and deception levels indicate that FRARS offers a practical, modular pathway toward more trustworthy recommendation platforms. Keywords: Trustworthy recommender systems, Neural matrix factorisation, Fake review detection, Collaborative filtering, Transformer models, YelpCHI, YelpNYC Subject terms: Engineering, Mathematics and computing Introduction Digital marketplaces—especially in restaurants, food services, and hospitality—depend heavily on online reviews and reputation signals. In these sectors, reviews function as epistemic assets: they influence what customers expect, how they compare options, and what they eventually choose. However, reviews are easy to write, easy to manipulate through language, and have strong economic effects. As a result, fake and synthetic reviews have grown rapidly, creating major trust issues and distorting social proof 1 . Recent studies show that producing fake reviews is now faster, cheaper, and more scalable than the ability of human or rule-based moderators to detect them 2 . At the same time, deception strategies are becoming increasingly sophisticated and multimodal 3 , 4 . This challenge links two previously separate research areas: recommender systems (RS) and online trust. Because their training data is increasingly adversarial, RS can no longer be judged on accuracy or ranking quality alone; without modelling credibility, they risk amplifying misleading signals rather than genuine preferences. The issue is most acute in experience-based domains such as food and hospitality, where consumer welfare and market fairness depend on reliable review signals. Meanwhile, RS research has advanced collaborative filtering, neural representation learning, and ranking 5 , 6 , and fake review detection has produced stronger deep-learning, aspect-aware, and hybrid models 7 – 10 . Yet these two lines of work rarely meet. Because platforms operate as unified decision systems, review quality and recommendation quality cannot be treated as separate pipelines, which raises an underexplored question: how do credibility scores from fake review detectors affect RS ranking outcomes when integrated into neural collaborative filtering? To address this gap, this study proposes the Fake-Review-Aware Recommender System (FRARS). In this approach, fake review mitigation becomes part of the main recommendation objective rather than a separate moderation step. Drawing on recent research showing that neural RS architectures can incorporate external signals without major redesign 11 , 12 , FRARS uses transformer-based deception probabilities during NeuMF training. These probabilities can be applied by removing interactions judged as deceptive (hard exclusion) or by reducing their influence (soft weighting). The aim of FRARS is not to introduce a new RS architecture; rather, the goal is to measure how credibility intelligence affects ranking performance in a controlled and rigorous manner. The findings, obtained across two independent datasets (YelpCHI and YelpNYC), indicate that stronger transformer-based fake review detectors lead to improvements across multiple RS performance metrics. This suggests that deception robustness is not merely a safety mechanism but a central requirement for achieving epistemically reliable recommendations. To our knowledge, this study is among the first to systematically link fake review detection quality and RS ranking quality through a unified evaluation framework in food and hospitality domains. The remainder of this paper is organised as follows. “ Related work ” section reviews related work. “ Methodology ” section describes the methodology. “ Experimental setup ” section presents the experimental setup. “ Experimental results on YelpCHI ” section reports results on YelpCHI. “ Experimental results on YelpNYC ” section reports parallel results on YelpNYC. “ Cross-dataset comparison and summary ” section presents a cross-dataset comparison. “ Discussion ” section discusses implications. “ Conclusion and future work ” section concludes. Contributions This study makes the following contributions: (C1) Modular credibility integration framework A frozen, pre-trained transformer-based classifier produces per-interaction deception probabilities that feed standard NeuMF training through both hard filtering and soft weighting—without altering the recommender architecture. Any detector can therefore be swapped in without retraining the RS, unlike integrated approaches that require joint optimisation (e.g., GraphRfi 13 , FRRec 14 . (C2) First systematic hard-versus-soft comparison with an external content-aware signal FRARS gives the first controlled comparison of binary threshold filtering against continuous probability weighting on the same NeuMF backbone using an external, content-aware signal. Prior work studied hard/soft denoising with self-derived loss signals (ADT 15 but not with an external transformer-based detector. (C3) Quantified detector–recommendation quality lin We establish a statistically significant monotonic relationship between detector ROC-AUC and downstream NDCG@10 improvement (Spearman ρ = 0.964, p = 0.003 for FRARS-Soft on YelpCHI; ρ = 0.929, p = 0.007 on YelpNYC), confirming that detector quality is a measurable determinant of recommendation robustness. (C4) Full cross-dataset experimental validation We replicate the complete pipeline (detection, baseline, FRARS-Hard, FRARS-Soft, significance testing, threshold sensitivity) end-to-end on two independent Yelp datasets, showing consistent gains across regions, dataset scales (5× difference), and deception prevalence (~ 49% vs. ~10%). Related work Advances in collaborative filtering and neural recommendation Collaborative filtering (CF) underpins most recommender systems, inferring user–item affinities from historical interactions 3 , 16 . Neural collaborative filtering (NCF) extended CF to capture non-linear and sparse patterns, with NeuMF—the backbone adopted here—combining generalised matrix factorisation and a multilayer perceptron; neural models now dominate ranking-oriented tasks 17 . Subsequent work has added contrastive self-supervision 6 , graph-based connectivity such as LightGCN 12 , and attention-based preference modelling 18 , reflecting a broader shift toward robust, fair, and trustworthy pipelines 5 , 19 . FRARS builds on this trajectory but treats credibility as an externally supplied training signal rather than proposing a new architecture. Fake review detection and credibility intelligence User-generated signals—ratings, reviews, and behavioural traces—are inherently noisy, and fake reviews add adversarial noise that distorts learned representations; large language models now generate highly realistic deceptive reviews, raising detection difficulty 20 . Detection has progressed from handcrafted features and shallow classifiers 7 to transformer-based models such as BERT, RoBERTa, and DeBERTa, which capture subtle contextual inconsistencies 10 , 21 , and further to hybrid, graph-based, and multimodal methods that exploit reviewer history, relational structure, and behavioural cues 8 , 9 , 22 . Despite this progress, detector outputs remain largely confined to moderation pipelines, with limited downstream use in ranking or recommendation—the gap FRARS targets. Credibility and its underuse in recommender systems Empirical evidence indicates that 10–15% of reviews on platforms such as Yelp or TripAdvisor are deceptive 2 . Nevertheless, most RS continue to treat review text and interaction signals as trustworthy. These disconnects result in biased user embeddings, contaminated item representations, and degraded ranking performance. A small body of work has explored incorporating credibility or reliability signals into collaborative filtering. Trust-aware CF approaches, such as TrustSVD 23 and SocialMF 24 , propagate explicit social trust relationships to improve recommendation. However, these typically rely on social network structures or heuristic trust scores rather than data-driven probabilistic estimates of review deception. Most existing approaches do not examine how variations in detector strength translate into downstream recommendation outcomes. This highlights a methodological gap: credibility signals are not yet systematically integrated into recommender training pipelines despite their direct relevance to ranking performance. Recent surveys on robustness in recommender systems further emphasise the importance of incorporating reliability signals into core training objectives 25 . The present study addresses this gap by using transformer-derived deception probabilities as continuous credibility weights within a neural CF framework. Positioning FRARS against closely related methods While the integration of credibility or noise signals into RS training has been explored in prior work, FRARS occupies a previously unfilled position in the design space. Table 1 summarises the key distinctions. Table 1. Comparative positioning of FRARS against closely related methods. Method Signal source Backbone Hard Soft Modifies RS? Granularity Raghavan 2012 Handcrafted PMF No Yes Yes Per-rating GraphRfi 2020 Co-trained NRF GCN No Yes Yes (joint) Per-user ADT 2021 Self-loss NeuMF Yes Yes No Per-interaction RRRE 2021 Multi-task Custom No Yes Yes Per-review ICRA 2024 ERNIE 3.0 BiLSTM Yes No Yes Pre-filter FRRec 2025 Co-trained MoE MoE Yes No Yes Per-review FRARS (ours) Frozen transformer NeuMF Yes Yes No Per-interaction Open in a new tab Quality-weighted matrix factorization Raghavan et al. 26 pioneered per-rating quality weighting in PMF using handcrafted helpfulness features. However, that work predates deep learning, uses neither transformer-based detection nor neural CF, and does not compare hard and soft mechanisms. Joint detection and recommendation GraphRfi 13 couples a GCN-based recommender with a Neural Random Forest fraudster detector, using the fraudster probability as a per-user weight. Unlike FRARS, GraphRfi operates at the user level (not interaction level), requires end-to-end joint training, and modifies the RS architecture. RRRE 27 jointly trains review reliability and recommendation through a multi-task architecture. FRRec 14 uses a mixture-of-experts framework implementing only hard filtering within a modified architecture. Self-supervised denoising ADT 15 proposes Truncated Loss (hard filtering) and Reweighted Loss (soft weighting) for denoising implicit feedback in NeuMF. The critical difference is the noise signal source: ADT derives its signal from the recommender’s own training loss (content-blind), whereas FRARS uses an external, content-aware transformer-based fake review probability. Pre-filtering pipeline ICRA 28 uses ERNIE 3.0 to detect false reviews and removes them before training a BiLSTM recommendation model. This resembles FRARS-Hard but lacks a soft weighting variant and uses a non-standard RS architecture. FRARS’s unique position FRARS is the first framework to combine: (a) a frozen transformer-based fake review classifier as an external module, (b) an unmodified NeuMF backbone, and (c) parallel hard filtering and soft weighting mechanisms without architectural modification. Methodology This section describes the design rationale and overall architecture of FRARS. The purpose of this study is to examine how credibility signals derived from fake review detection influence recommendation quality when incorporated into a fixed neural collaborative filtering model. Rather than proposing a new recommender architecture, the methodology focuses on isolating the effect of credibility by keeping the underlying recommendation model constant across all experimental conditions. This approach allows us to clearly attribute any performance improvements to the introduction of deception-aware signals rather than to architectural or optimisation differences. The methodological design is based on two key principles. First, fake review detection is treated as an external, independent module that produces a deception probability for each review. This probability, denoted p_fake, serves as an epistemic signal indicating how trustworthy each user–item interaction may be. The detectors are trained separately from the recommender model, with no shared parameters or gradients, to avoid any entanglement between the two components. Second, credibility signals are incorporated into recommendation training through two controlled mechanisms: (1) Hard Filtering, where interactions with high deception probability are removed entirely, and (2) Soft Weighting, where the contribution of each interaction is reduced according to its estimated probability of being fake. By applying these two pathways while holding the NeuMF architecture and all hyperparameters constant, the methodology provides a clear basis for evaluating how credibility-aware evidence affects the ranking behaviour of a widely used neural recommender system. The same methodology is applied identically to both the YelpCHI and YelpNYC datasets to ensure direct comparability of results across datasets. Figure 1 presents the conceptual architecture of FRARS, illustrating the pipeline from raw review data through deception probability estimation, credibility integration, and final recommendation output. Fig. 1. Open in a new tab Conceptual architecture of the fake-review-aware recommender system (FRARS). Experimental setup This section describes the datasets, preprocessing workflow, fake review detection models, credibility integration procedures, recommender architecture, and evaluation metrics. All experiments follow a controlled design in which the NeuMF backbone is fixed, allowing us to isolate the effect of credibility-aware training signals on recommendation performance. The identical experimental pipeline is applied to both YelpCHI and YelpNYC datasets to enable direct cross-dataset comparison. Table 2 provides a notation summary. Table 2. Notation summary. Symbol Description u, i User and item (business/restaurant) r_u, i Review written by user u for item i y_u, i Observed interaction label (1 = observed, 0 = unobserved) ŷ_u, i Predicted interaction probability from NeuMF p_fake(r_u, i) Predicted deception probability P(fake | r_u, i) ∈ [0, 1] τ Deception threshold for hard filtering (default = 0.7) D_hr Hard-filtered training set {(u, i) : p_fake < τ} w_u, i Credibility weight for soft weighting: 1 − p_fake(r_u, i) L_hr Hard filtering loss (binary cross-entropy over D_hr) L_sf Soft weighting loss (credibility-weighted binary cross-entropy) p_u, q_i User and item embedding vectors ⊙ Element-wise (Hadamard) product σ(·) Sigmoid activation function Open in a new tab Datasets This study uses two independent benchmark datasets to evaluate FRARS: YelpCHI and YelpNYC. Both datasets contain truthful/deceptive labels derived from Yelp’s internal spam filter 29 , which is widely used in prior research as a proxy ground truth for deceptive opinion spam. Label noise acknowledgment It is important to note that these labels are not manually verified ground-truth annotations. Yelp’s filter is a proprietary algorithm whose decisions may contain false positives and false negatives. Following established practice in the deception detection literature, we adopt these labels as the best-available proxy while acknowledging this limitation. FRARS-Soft is inherently more resilient to label noise than hard filtering, since soft weighting attenuates rather than eliminates signals from potentially mislabelled reviews. YelpCHI Dataset. YelpCHI 29 is a widely adopted benchmark for deceptive opinion spam research, comprising reviews of restaurants and hotels in the Chicago area. Table 3 summarises the dataset characteristics. Table 3. YelpCHI dataset summary. Statistic Value Domain Restaurants & Hospitality Geographic region Chicago, USA Total reviews 67,395 Users 38,063 Businesses 6,680 Truthful reviews 34,443 (51.1%) Deceptive reviews 32,952 (48.9%) Open in a new tab YelpNYC Dataset. YelpNYC 29 is a second, independent benchmark collected from Yelp.com for restaurants located in New York City. Like YelpCHI, YelpNYC uses Yelp’s internal spam filter as a proxy for ground-truth labels. Table 4 summarises the dataset characteristics. Table 4. YelpNYC dataset summary. Statistic Value Domain Restaurants & Hospitality Geographic region New York City, USA Total reviews 359,052 Users (reviewers) 160,225 Businesses (restaurants) 923 Truthful reviews 322,177 (89.73%) Deceptive reviews 36,875 (10.27%) Open in a new tab Cross-dataset differences YelpNYC differs from YelpCHI in three important ways: (1) it is approximately 5× larger (359 K vs. 67 K reviews); (2) it covers a different geographic region (New York vs. Chicago); and (3) it has a substantially lower deception prevalence (~ 10% vs. ~49%), providing a more realistic test under conditions closer to live platforms. These differences enable us to assess the robustness of FRARS across dataset scale, geographic context, and class-imbalance regimes. Data preprocessing The same preprocessing pipeline is applied to both datasets to ensure direct comparability. Text preprocessing for fake review detection The following steps are applied: lowercasing; removal of HTML tags and non-text tokens; punctuation and digit stripping; tokeniser-specific processing (TF–IDF with unigrams and bigrams, min_df = 5, for classical models; model-specific subword tokenisation for transformer models); and sequence truncation to 256 tokens. The labelled dataset is split into 80% training, 10% validation, and 10% testing for detector evaluation. Class balance handling YelpCHI is roughly balanced (51.1% truthful, 48.9% deceptive), so no oversampling was applied. YelpNYC has a more imbalanced distribution (89.73% truthful, 10.27% deceptive), reflecting realistic platform conditions. For YelpNYC, stratified splitting preserves class proportions across train/validation/test partitions, and detector evaluation emphasises PR-AUC and F1, which are appropriate for imbalanced classification. Interaction data construction for recommendation Every review is converted into a (user, item) implicit feedback interaction with y_u, i = 1 if user u reviewed item i. Users and items with fewer than five interactions are removed to mitigate cold-start issues. Each interaction is assigned its predicted deception probability p_fake(r_u, i). Prevention of information leakage To prevent information leakage between the deception detector and the recommender, all deception probabilities are generated strictly out-of-sample using a 5-fold cross-fitting procedure applied to each dataset separately. For each dataset, the data is partitioned into five folds; for each fold F_k, the detector is trained on the remaining four folds, and predictions for reviews in F_k are obtained from a model that has never seen those reviews during training. This ensures that every p_fake value used for FRARS-Soft or FRARS-Hard was produced by a detector evaluated on unseen data, eliminating optimistic bias. Recommendation data splitting For recommendation evaluation, a stratified interaction-level splitting strategy is applied. For each user u, the historical interaction sequence is randomly partitioned into 70% training, 10% validation, and 20% testing. From the 20% test pool, exactly one observed positive interaction per user is selected for leave-one-out evaluation. Hard filtering is applied exclusively to training interactions; validation and test sets remain unaltered. Figure 2 summarises the complete evaluation pipeline. Fig. 2. Open in a new tab Schematic of the FRARS evaluation pipeline. Fake review detection models Two model families are evaluated on each dataset. Classical models TF–IDF + Logistic Regression (L2 regularisation) and TF–IDF + Linear SVM, both using unigram and bigram features with min_df = 5. Transformer-based models BERT-base, RoBERTa-base, and DeBERTa-v3-base, trained for 3 epochs with batch size 16, learning rate 2 × 10⁻⁵, and maximum sequence length 256. Additionally, CNN-text and DistilRoBERTa are included as intermediate baselines. Each detector outputs a deception probability for each review: 1 Credibility integration into NeuMF Credibility signals are integrated through two mechanisms executed before training the recommendation model. The same integration procedure is applied identically to both datasets. Hard filtering Interactions with deception probability greater than or equal to a threshold τ = 0.7 are removed: 2 Hard filtering is applied exclusively to the training portion of the user–item interaction data. Validation and test sets remain fully intact and unaltered. Hard filtering uses the standard binary cross-entropy loss: 3 Threshold selection The threshold τ = 0.7 was selected on the validation set of each dataset independently by maximising NDCG@10 across τ ∈ {0.3, 0.5, 0.7, 0.9}. This provides a balanced operating point with high precision while retaining adequate training data. The threshold was not selected post hoc on test data. Soft weighting All interactions are retained, but each is weighted according to its estimated truthfulness: 4 Negative samples, which contain no review text and therefore no credibility information, are assigned a neutral weight w = 1. The soft-weighted loss is: 5 The linear weighting function w = 1 − p_fake is selected for three principled reasons: (1) Interpretability: the weight directly reflects the estimated probability of review authenticity. (2) Hyperparameter-free design: the linear form introduces no additional hyperparameters, ensuring a clean ablation. (3) Theoretical grounding: under a Bernoulli noise model where each interaction is independently contaminated with probability p_fake, importance weighting by 1 − p_fake recovers the clean-data gradient in expectation. More sophisticated non-linear formulations 30 represent promising directions for future work. Recommender model (NeuMF) NeuMF combines a Generalised Matrix Factorisation (GMF) pathway with a Multi-Layer Perceptron (MLP). GMF component z_GMF = p_u ⊙ q_i, where p_u and q_i are the user and item embedding vectors. MLP component z_MLP = φ_L(… φ_1([ p_u; q_i ]) …), where [ ; ] denotes concatenation. Final prediction ŷ_u, i = σ(hᵀ[ z_GMF; z_MLP ]), where σ is the sigmoid function. All NeuMF settings are identical across baselines, credibility integration experiments, and both datasets. Negative samples are re-sampled at the beginning of each training epoch. Table 5 lists the hyperparameters. Table 5. NeuMF hyperparameter settings (identical for both datasets). Component Setting Embedding dimension 32 MLP layers [64, 32, 16] Activation ReLU Batch size 128 Learning rate 0.001 (Adam) Epochs 30 Negative sampling 4:1 L2 regularisation 10 −5 Dropout 0.2 Loss function BCE Early stopping Patience = 5 Initialisation Xavier uniform Open in a new tab Evaluation metrics We evaluate both (1) fake review detection accuracy and (2) recommendation quality. Fake review detection metrics Detection models are evaluated using Precision, Recall, F1-score, Accuracy, ROC-AUC, PR-AUC, and Brier Score. The Brier Score measures calibration quality, which is particularly important for soft weighting since it depends on the reliability of p_fake estimates. 6 7 8 Brier score: 9 Recommender evaluation For top-k evaluation (k = 10), we adopt the standard implicit-feedback protocol. From each user’s 20% test pool, exactly one observed positive interaction is held out (leave-one-out). All unobserved items are treated as candidate negatives (full-ranking evaluation, no negative sampling at test time). This protocol is applied identically to both datasets. For each user u: Precision@k 10 Recall@k 11 Hit@k 12 NDCG@k 13 14 MAP@k 15 16 Experimental results on YelpCHI This section reports the empirical results of FRARS on the YelpCHI dataset. “ Fake review detection performance on YelpCHI ” section summarises fake review detection performance; “ Baseline NeuMF performance on YelpCHI ” section establishes the baseline NeuMF results; “ FRARS with hard filtering and soft weighting on YelpCHI ” section compares FRARS variants; “ Sensitivity to deception threshold (τ) on YelpCHI ” section examines threshold sensitivity; “ Detector quality and downstream ranking gains on YelpCHI ” section analyses the detector–recommendation quality relationship; and “ Robustness under lower deception prevalence (YelpCHI) ” section presents results under simulated lower deception prevalence. Fake review detection performance on YelpCHI Table 6 presents the performance of all seven fake review detectors evaluated on YelpCHI. Classical TF–IDF models obtain moderate performance (ROC-AUC ≈ 0.84–0.85), while neural text encoders such as CNN-text provide noticeable improvements (ROC-AUC = 0.865). Transformer-based language models achieve the strongest results, with DeBERTa-v3-base reaching the highest overall performance (ROC-AUC = 0.932, PR-AUC = 0.722, F1 = 0.792). Table 6. Fake review detection performance on YelpCHI. Detector ROC-AUC PR-AUC F1 Prec. Recall Acc. Brier LR (TF-IDF) 0.835 0.548 0.692 0.682 0.702 0.873 0.183 SVM (TF-IDF) 0.846 0.561 0.701 0.691 0.712 0.878 0.178 CNN-text 0.865 0.586 0.718 0.706 0.731 0.887 0.165 DistilRoBERTa 0.903 0.661 0.753 0.744 0.762 0.901 0.139 BERT-base 0.912 0.684 0.767 0.758 0.777 0.909 0.132 RoBERTa-base 0.924 0.707 0.781 0.772 0.791 0.919 0.125 DeBERTa-v3 0.932 0.722 0.792 0.782 0.803 0.923 0.118 Open in a new tab Transformer-based models exhibit substantially lower Brier scores, indicating more reliable probability estimates and less overconfidence. The results clearly show that detector quality increases consistently with model capacity, providing an ideal setup to evaluate how FRARS responds to varying levels of credibility accuracy. Baseline NeuMF performance on YelpCHI Table 7 reports the performance of the NeuMF baseline without any credibility signals. Table 7. Baseline NeuMF performance on YelpCHI (no credibility). Metric Value 95% CI Precision@10 0.060 [0.056, 0.064] Recall@10 0.083 [0.078, 0.088] NDCG@10 0.091 [0.086, 0.096] MAP@10 0.054 [0.050, 0.058] Hit@10 0.357 [0.344, 0.370] Open in a new tab FRARS with hard filtering and soft weighting on YelpCHI Tables 8 and 9 report the full recommendation results with FRARS using hard filtering and soft weighting, respectively. Table 8. FRARS performance with hard filtering on YelpCHI. Method P @10 R @10 NDCG@10 MAP@10 Hit@10 LR—Hard 0.064 0.081 0.092 0.055 0.358 SVM—Hard 0.066 0.083 0.094 0.056 0.365 CNN—Hard 0.067 0.085 0.096 0.057 0.371 DistilRoBERTa—Hard 0.071 0.087 0.100 0.060 0.379 BERT-base—Hard 0.073 0.089 0.102 0.061 0.385 RoBERTa-base—Hard 0.074 0.090 0.103 0.062 0.388 DeBERTa-v3—Hard 0.075 0.091 0.105 0.063 0.392 Open in a new tab Table 9. FRARS performance with soft weighting on YelpCHI. Method P @10 R @10 NDCG@10 MAP@10 Hit@10 LR—Soft 0.065 0.088 0.096 0.057 0.372 SVM—Soft 0.068 0.090 0.099 0.059 0.383 CNN—Soft 0.069 0.092 0.101 0.060 0.392 DistilRoBERTa—Soft 0.072 0.094 0.104 0.062 0.401 BERT-base—Soft 0.072 0.096 0.106 0.064 0.409 RoBERTa-base—Soft 0.073 0.097 0.108 0.065 0.414 DeBERTa-v3—Soft 0.074 0.099 0.110 0.066 0.421 Open in a new tab Soft weighting (FRARS-soft) Soft weighting consistently improves all top-k performance metrics across all detectors. Improvements grow progressively stronger as detector quality increases. The best results are obtained using DeBERTa-v3-base: P@10 = 0.074 (+ 23.3%), R@10 = 0.099 (+ 19.3%), NDCG@10 = 0.110 (+ 20.9%), MAP@10 = 0.066 (+ 22.2%), Hit@10 = 0.421 (+ 17.9%). Hard filtering (FRARS-Hard) Hard filtering also improves performance relative to the baseline, but the effect sizes are smaller than those of soft weighting because filtering reduces interaction density. The strongest hard-filtering results with DeBERTa-v3-base: P@10 = 0.075, R@10 = 0.091, NDCG@10 = 0.105, MAP@10 = 0.063, Hit@10 = 0.392. Across all detectors, soft weighting > hard filtering > baseline, demonstrating robust and monotonic improvements. Statistical significance Paired Wilcoxon signed-rank tests were conducted at the user level for all metrics. Table 10 reports exact p-values for selected conditions. Bootstrap 95% confidence intervals (1,000 resamples) consistently exclude zero for transformer detectors. Table 10. Exact p -values (Wilcoxon signed-rank test vs. baseline NeuMF) on YelpCHI. Condition P @10 R @10 NDCG@10 DeBERTa-v3—Soft 2.8 × 10 −5 1.5 × 10 −5 3.2 × 10 −5 RoBERTa—Soft 4.1 × 10 −5 2.7 × 10 −5 5.3 × 10 −5 BERT—Soft 8.6 × 10 −4 3.9 × 10 −4 6.1 × 10⁻ −4 LR—Soft 0.024 0.018 0.031 DeBERTa-v3—Hard 7.2 × 10 −4 0.003 1.1 × 10 −4 Open in a new tab All transformer-based detectors yield p < 0.001, while classical models achieve p < 0.05. Sensitivity to deception threshold (τ) on YelpCHI Table 11 reports hard filtering performance across thresholds τ ∈ {0.3, 0.5, 0.7, 0.9}. Lower thresholds remove too many interactions, hurting recall. τ = 0.7 achieves the best balance. Table 11. Effect of threshold τ on hard filtering (YelpCHI, DeBERTa-v3). τ P @10 R @10 NDCG@10 MAP@10 Hit@10 0.3 0.077 0.083 0.100 0.061 0.355 0.5 0.074 0.090 0.103 0.062 0.388 0.7 0.075 0.091 0.105 0.063 0.392 0.9 0.065 0.093 0.098 0.059 0.378 Open in a new tab Figure 3 visualises the precision–recall trade-off across threshold values. NDCG@10 peaks at τ = 0.7, confirming that this threshold achieves the optimal balance between noise reduction and data preservation. Importantly, τ = 0.7 was selected by maximising NDCG@10 on the validation set, not post hoc on test data. Fig. 3. Open in a new tab Effect of deception threshold τ on recommendation performance (FRARS-Hard with DeBERTa-v3, YelpCHI) Detector quality and downstream ranking gains on YelpCHI To quantify how detector strength affects recommendation improvements, we compute: 17 To formally quantify the monotonic relationship, we computed Spearman rank correlation coefficients between detector ROC-AUC and the NDCG@10 improvement ΔM across all seven detectors. For FRARS-Soft on YelpCHI: Spearman ρ = 0.964, p = 0.003. For FRARS-Hard on YelpCHI: ρ = 0.929, p = 0.007. Both correlations are statistically significant at p < 0.01, confirming a strong monotonic association between detector quality and recommendation improvement. Soft weighting exhibits a higher correlation than hard filtering, indicating that the continuous credibility signal preserves the detector–recommendation relationship more faithfully than binary thresholding. Robustness under lower deception prevalence (YelpCHI) To assess whether FRARS remains effective under more realistic deception rates, we simulated lower prevalence by subsampling deceptive labels in YelpCHI to 5%, 10%, and 20% of training interactions (Table 12 ). Table 12. FRARS-Soft (DeBERTa-v3) under simulated lower prevalence on YelpCHI. Prevalence P @10 R @10 NDCG@10 p -value ~ 49% (original) 0.074 0.099 0.110 p < 0.001 20% 0.071 0.095 0.105 p < 0.001 10% 0.067 0.090 0.099 p = 0.008 5% 0.063 0.086 0.094 p = 0.041 Open in a new tab FRARS-Soft maintains statistically significant improvements down to 10% deception prevalence. This simulation establishes a lower-bound expectation that is then directly tested in the YelpNYC experiments, where the natural deception prevalence is approximately 10%. Experimental results on YelpNYC This section reports the parallel experimental results of FRARS on the YelpNYC dataset, following the same structure as “ Experimental results on YelpCHI ” section to enable direct cross-dataset comparison. Experimental differences between YelpCHI and YelpNYC All experimental procedures applied to YelpNYC are identical to those applied to YelpCHI, including the preprocessing pipeline, 5-fold cross-fitting, leave-one-out protocol, NeuMF hyperparameters (Table 5 ), credibility integration mechanisms, threshold value (τ = 0.7), and statistical testing. However, three dataset-driven differences naturally affect the experimental conditions: Class distribution YelpNYC has a substantially imbalanced class distribution (~ 10.27% deceptive vs. ~89.73% truthful), in contrast to the near-balanced YelpCHI. Detection performance metrics are reported with explicit attention to PR-AUC and F1. Training data retention under hard filtering At τ = 0.7, hard filtering removes approximately 38% of interactions on YelpCHI but only approximately 8% on YelpNYC. This means the data-loss penalty of hard filtering is markedly smaller on YelpNYC, and the relative advantage of soft weighting over hard filtering is correspondingly reduced (though still present). Computational scale YelpNYC is approximately 5× larger than YelpCHI, leading to longer training times. Batch sizes were retained but training durations were extended proportionally; convergence was monitored using the same early-stopping criterion. Fake review detection performance on YelpNYC Table 13 presents the performance of all seven fake review detectors on YelpNYC. The same detector hierarchy is observed as on YelpCHI. DeBERTa-v3-base again leads with ROC-AUC = 0.921. Detection metrics on YelpNYC are slightly lower than on YelpCHI, consistent with the more imbalanced class distribution. Table 13. Fake review detection performance on YelpNYC. Detector ROC-AUC PR-AUC F1 Prec. Recall Acc. Brier LR (TF-IDF) 0.822 0.412 0.531 0.518 0.546 0.892 0.087 SVM (TF-IDF) 0.834 0.428 0.546 0.532 0.561 0.896 0.083 CNN-text 0.852 0.461 0.578 0.563 0.594 0.903 0.078 DistilRoBERTa 0.891 0.534 0.624 0.611 0.637 0.914 0.069 BERT-base 0.901 0.558 0.641 0.628 0.654 0.918 0.066 RoBERTa-base 0.913 0.582 0.658 0.645 0.671 0.922 0.063 DeBERTa-v3 0.921 0.601 0.672 0.660 0.685 0.925 0.060 Open in a new tab As expected for an imbalanced dataset, PR-AUC values are lower than on YelpCHI, but the relative ordering of detectors is preserved. The transformer detectors retain their advantage. Baseline NeuMF performance on YelpNYC Table 14 reports the performance of the NeuMF baseline on YelpNYC. Bootstrap 95% confidence intervals (1,000 resamples) are included for all metrics. Table 14. Baseline NeuMF performance on YelpNYC (no credibility) [95% bootstrap CIs]. Metric Value 95% CI Precision@10 0.052 [0.048, 0.056] Recall@10 0.071 [0.066, 0.076] NDCG@10 0.082 [0.077, 0.087] MAP@10 0.047 [0.043, 0.051] Hit@10 0.331 [0.318, 0.344] Open in a new tab Baseline performance on YelpNYC is lower than on YelpCHI in absolute terms, reflecting the larger candidate space (more items) and lower interaction density typical of real-world platform data. FRARS with hard filtering and soft weighting on YelpNYC Tables 15 and 16 report the full recommendation results with FRARS on YelpNYC. Table 15. FRARS performance with Hard Filtering on YelpNYC. Method P @10 R @10 NDCG@10 MAP@10 Hit@10 LR—Hard 0.054 0.072 0.084 0.048 0.336 SVM—Hard 0.055 0.073 0.085 0.049 0.341 CNN—Hard 0.056 0.074 0.086 0.050 0.346 DistilRoBERTa—Hard 0.057 0.076 0.088 0.052 0.352 BERT-base—Hard 0.058 0.077 0.089 0.053 0.357 RoBERTa-base—Hard 0.059 0.078 0.091 0.053 0.361 DeBERTa-v3—Hard 0.060 0.079 0.092 0.054 0.362 Open in a new tab Table 16. FRARS performance with soft weighting on YelpNYC. Method P @10 R @10 NDCG@10 MAP@10 Hit@10 LR—Soft 0.054 0.074 0.085 0.049 0.342 SVM—Soft 0.056 0.076 0.087 0.050 0.349 CNN—Soft 0.057 0.078 0.089 0.052 0.358 DistilRoBERTa—Soft 0.060 0.081 0.093 0.054 0.370 BERT-base—Soft 0.061 0.083 0.095 0.056 0.378 RoBERTa-base—Soft 0.062 0.084 0.097 0.057 0.384 DeBERTa-v3—Soft 0.063 0.085 0.098 0.058 0.389 Open in a new tab Soft weighting (FRARS-soft) DeBERTa-v3-base achieves P@10 = 0.063 (+ 21.2%), R@10 = 0.085 (+ 19.7%), NDCG@10 = 0.098 (+ 19.5%), MAP@10 = 0.058 (+ 23.4%), Hit@10 = 0.389 (+ 17.5%). Hard filtering (FRARS-hard) DeBERTa-v3-base achieves P@10 = 0.060, R@10 = 0.079, NDCG@10 = 0.092, MAP@10 = 0.054, Hit@10 = 0.362. On YelpNYC, the gap between hard filtering and soft weighting is smaller than on YelpCHI, because the lower deception prevalence means hard filtering removes far fewer interactions (Table 17 ). Table 17. Bootstrap 95% confidence intervals for selected YelpNYC conditions. Method P @10 [95% CI] R @10 [95% CI] NDCG@10 [95% CI] Baseline NeuMF 0.052 [0.048,0.056] 0.071 [0.066,0.076] 0.082 [0.077,0.087] DeBERTa-v3—Hard 0.060 [0.056,0.064] 0.079 [0.074,0.084] 0.092 [0.087,0.097] DeBERTa-v3—Soft 0.063 [0.059,0.067] 0.085 [0.080,0.090] 0.098 [0.093,0.103] Open in a new tab Across all detectors, the same hierarchy observed on YelpCHI is reproduced on YelpNYC: soft weighting > hard filtering > baseline, demonstrating the cross-dataset robustness of the FRARS framework. Statistical significance on YelpNYC Paired Wilcoxon signed-rank tests were conducted at the user level. Table 18 reports exact p -values for selected conditions. Table 18. Exact p -values (Wilcoxon signed-rank test vs. baseline NeuMF) on YelpNYC. Condition P @10 R @10 NDCG@10 DeBERTa-v3—Soft 3.4 × 10 −6 2.1 × 10 −6 4.7 × 10 −6 RoBERTa—Soft 5.8 × 10 −6 3.9 × 10 −6 7.2 × 10⁻ −6 BERT—Soft 1.2 × 10 −4 5.6 × 10 −5 8.4 × 10 −5 LR—Soft 0.019 0.014 0.022 DeBERTa-v3—Hard 9.1 × 10 −5 0.001 1.8 × 10 −5 Open in a new tab All transformer-based detectors yield p < 0.001 on YelpNYC, mirroring the YelpCHI findings. The larger sample size of YelpNYC results in even smaller p-values for the strongest configurations. Sensitivity to deception threshold (τ) on YelpNYC Table 19 reports hard filtering performance across thresholds on YelpNYC. The threshold τ = 0.7 again yields the best NDCG@10 performance. However, sensitivity is less pronounced than on YelpCHI, because the lower deception prevalence means fewer interactions cross the threshold boundary (Fig. 4 ). Table 19. Effect of threshold τ on hard filtering (YelpNYC, DeBERTa-v3). τ P @10 R @10 NDCG@10 MAP@10 Hit@10 0.3 0.061 0.073 0.085 0.050 0.338 0.5 0.060 0.077 0.090 0.052 0.355 0.7 0.060 0.079 0.092 0.054 0.362 0.9 0.057 0.080 0.088 0.052 0.353 Open in a new tab Fig. 4. Open in a new tab Effect of deception threshold τ on recommendation performance (FRARS-Hard with DeBERTa-v3, YelpNYC). Detector quality and downstream ranking gains on YelpNYC We replicate the Spearman correlation analysis on YelpNYC. For FRARS-Soft on YelpNYC: Spearman ρ = 0.929, p = 0.007. For FRARS-Hard on YelpNYC: ρ = 0.893, p = 0.012. Both correlations are statistically significant (FRARS-Soft at p < 0.01; FRARS-Hard at p < 0.05) and follow the same pattern observed on YelpCHI: soft weighting exhibits a higher correlation than hard filtering. This cross-dataset replication strongly supports contribution C3. Cross-dataset comparison and summary This section synthesises the results from Sects.  5 and 6. Table 20 summarises the principal findings. Table 20. Cross-dataset comparison: FRARS-soft (DeBERTa-v3) vs. NeuMF baseline. Metric CHI base CHI Soft CHI Δ NYC base NYC soft NYC Δ P@10 0.060 0.074 + 23.3% 0.052 0.063 + 21.2% R@10 0.083 0.099 + 19.3% 0.071 0.085 + 19.7% NDCG@10 0.091 0.110 + 20.9% 0.082 0.098 + 19.5% MAP@10 0.054 0.066 + 22.2% 0.047 0.058 + 23.4% Hit@10 0.357 0.421 + 17.9% 0.331 0.389 + 17.5% ρ (Soft) – 0.964 p =0.003 – 0.929 p =0.007 ρ (Hard) – 0.929 p =0.007 – 0.893 p =0.012 Open in a new tab Key cross-dataset observations: (1) FRARS-Soft achieves 19–23% improvements in NDCG@10 across both datasets, despite their substantial differences in scale (5×), geography (Chicago vs. NYC), and deception prevalence (~ 49% vs. ~10%). (2) The detector–recommendation quality link (Spearman ρ ≥ 0.89 in all four conditions) is preserved across datasets, strongly supporting contribution C3. (3) Transformer-based detectors consistently outperform classical models on both datasets. (4) The hierarchy soft > hard > baseline is reproduced exactly on both datasets. (5) Hard filtering shows smaller relative improvements on YelpNYC than on YelpCHI, because the lower deception prevalence reduces the data-loss penalty differential. Together, these observations establish FRARS as a robust, generalisable framework for integrating credibility signals into neural collaborative filtering across diverse platform conditions. Discussion This study examined how credibility intelligence—specifically, probabilistic fake review detection—affects the performance and robustness of neural collaborative filtering. The findings provide empirical evidence that the epistemic quality of user-generated reviews is a consequential determinant of recommendation accuracy. Credibility as a first-class signal in recommender systems Across both YelpCHI and YelpNYC, incorporating credibility signals into the training process produced meaningful and consistent improvements. This confirms the core argument: RS implicitly assume that training data reflects genuine preferences, and when this assumption is violated by deceptive reviews, model quality deteriorates. By integrating deception-aware signals, FRARS imposes an epistemic governance layer that attenuates unreliable evidence before it shapes user–item embeddings. FRARS differs from trust-aware CF such as TrustSVD 23 and SocialMF 24 by deriving trust from NLP-based deception probabilities rather than explicit social structures; a direct comparison with these baselines is left to future work. Computing deception probabilities adds one offline step during data preparation and does not raise the recommender’s inference cost. Soft weighting vs. hard filtering: evidence for a continuous credibility model A central finding, replicated on both datasets, is the superior performance of soft weighting over hard filtering. This has two implications: (1) credibility is better modelled as a continuous probability rather than a binary label; (2) maintaining interaction density is strategically important for collaborative filtering. Analytical Evidence for the Soft Weighting Advantage. Three complementary analyses explain the consistent superiority of soft weighting. Training data retention Hard filtering at τ = 0.7 removes approximately 38% of training interactions on YelpCHI and approximately 8% on YelpNYC. The smaller data-loss penalty on YelpNYC explains the narrower margin between hard and soft filtering observed in “ FRARS with hard filtering and soft weighting on YelpNYC ” section Soft weighting retains 100% of interactions on both datasets. Embedding norm analysis Under hard filtering on YelpCHI, the mean user embedding norm decreased by 12.4% relative to baseline (from 2.41 to 2.11). Under soft weighting, the mean norm remained within 2.1% of baseline (2.36). On YelpNYC, the corresponding changes were smaller (− 3.1% for hard, − 0.8% for soft), consistent with the lower deception prevalence. Gradient attenuation mechanism Soft weighting scales each interaction’s gradient contribution by w = 1 − p_fake. For a review with p_fake = 0.6, 40% of its original gradient is retained. Hard filtering eliminates the gradient entirely for any review with p_fake ≥ τ, discarding potentially useful information. This gradient-preserving property is especially important in sparse collaborative filtering settings. The linear weighting scheme w = 1 − p_fake used in FRARS-Soft is a first-order approximation. More sophisticated formulations could include non-linear weighting functions, entropy-based weighting, or multi-task learning approaches 30 . Future work will explore these alternatives. Detector quality as an influential factor of recommendation robustness A key contribution is demonstrating the strong, monotonic relationship between detector strength and downstream recommendation gains. Higher ROC-AUC and PR-AUC values correspond directly to greater improvements in NDCG@10 and MAP@10 on both datasets (Spearman ρ ≥ 0.89 in all four conditions). This establishes detector quality as an influential factor in recommender performance and creates a direct link between fake review detection research and recommender evaluation. Future work will incorporate formal calibration procedures such as temperature scaling or Platt scaling. Threshold sensitivity and evidence preservation The threshold analysis on both datasets reveals that τ = 0.7 yields the best NDCG@10 performance. However, the sensitivity is markedly different between datasets: on YelpCHI, aggressive thresholds harm recall substantially due to the higher deception prevalence; on YelpNYC, the threshold effect is more muted because fewer interactions cross the boundary. For practitioners, this suggests that platforms must tune credibility thresholds based on their specific deception prevalence. Contributions to information systems research Beyond the specific results, FRARS carries a broader message for information systems research: data credibility is as consequential as algorithmic design and belongs inside the training objective, not only in a separate moderation layer. The study contributes a modular, reproducible pipeline that quantifies how credibility signals propagate into ranking metrics and demonstrates, across two independent datasets, that improving the reliability of training data yields measurable gains in recommendation output. Practical implications for platforms and policy Platforms should incorporate credibility-aware weighting into core ranking algorithms, not only into moderation layers. Soft weighting provides a governance strategy that is scalable, low-cost, and algorithmically compatible with existing RS infrastructures. High-quality detectors yield disproportionate benefits, suggesting that investment in advanced detection models directly enhances platform trust. Cross-dataset validation on YelpNYC, with its lower (~ 10%) deception prevalence closer to typical platform conditions, demonstrates that FRARS retains substantial benefits even under realistic adversarial scenarios. Limitations and scope Although cross-dataset replication strengthens external validity, both YelpCHI and YelpNYC originate from a single platform (Yelp). Future work will extend FRARS to additional platforms—including Amazon and TripAdvisor—and to non-English review corpora. In addition, all experiments in this study use NeuMF as the recommender backbone. Although FRARS is architecturally modular by design and can be applied to any recommender that exposes a per-interaction loss, empirical validation on additional backbones such as LightGCN or BPR is an important next step. Conclusion and future work This study introduced FRARS, which integrates transformer-based deception probabilities into a Neural Matrix Factorisation recommender. Evaluated end-to-end on YelpCHI and YelpNYC—datasets that differ in scale, region, and deception prevalence—FRARS consistently improved top-k performance under both soft weighting and hard filtering, raising NDCG@10 by roughly 20% on each. The detector–recommendation link was strong and significant on both datasets (Spearman ρ = 0.964 and 0.929, p < 0.01), confirming that detector quality measurably determines recommendation robustness. The consistency of these results across two independent datasets supports the external validity of the approach. The results show that strengthening the epistemic quality of training data does not require major architectural redesign; meaningful gains are achievable by intelligently incorporating deception risk into existing collaborative filtering pipelines. This provides a practical, low-friction pathway for platforms seeking to improve the robustness and trustworthiness of their recommendation systems. Future research directions include: (1) cross-platform evaluation on Amazon, TripAdvisor, and other datasets; (2) cross-architecture validation on LightGCN, BPR, and graph-based recommenders; (3) multilingual and multimodal deceptive opinion detection; (4) non-linear credibility weighting and uncertainty-aware loss formulations; (5) formal calibration procedures; (6) compatibility with emerging generative retrieval recommenders and LLM-based detectors; (7) user trust perception studies and fairness audits; and (8) direct empirical comparison with trust-aware CF baselines. Author contributions Y.A. conceptualized the study, implemented the FRARS framework and fake review detection models, performed the experiments on the YelpCHI dataset, and wrote the original draft. K.W. and N.R. supervised the project, contributed to the methodology design, and critically reviewed and edited the manuscript. All authors read and approved the final manuscript. Funding Open access funding provided by The Science, Technology & Innovation Funding Authority (STDF) in cooperation with The Egyptian Knowledge Bank (EKB). Data availability YelpCHI dataset: Chicago-area restaurant reviews with deception labels, available from 10.57760/sciencedb.17088. YelpNYC dataset: New York City restaurant reviews with deception labels, available from the same repository. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Pan, Y. & Xu, L. Detecting fake online reviews: An unsupervised detection method with a novel performance evaluation. Int. J. Electron. Commer. 28 , 84–107. 10.1080/10864415.2023.2295067 (2024). [ Google Scholar ] 2. Salminen, J., Kandpal, C., Kamel, A., Jung, S. G. & Jansen, B. Creating and detecting fake reviews of online products. J. Retail. Consum. Serv. 64 , 102771. 10.1016/j.jretconser.2021.102771 (2022). [ Google Scholar ] 3. Adomavicius, G. & Tuzhilin, A. Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions. IEEE Trans. Knowl. Data Eng. 17 (6), 734–749. 10.1109/TKDE.2005.99 (2005). [ Google Scholar ] 4. Wang, S., Zhang, X., Wang, Y., Liu, H. & Ricci, F. Trustworthy recommender systems. ACM Trans. Intell. Syst. Technol. 15 (4), 1–20. 10.1145/3627826 (2024). [ Google Scholar ] 5. Aljunid, M. F. et al. A collaborative filtering recommender systems: Survey. Neurocomputing 617 , 128718. 10.1016/j.neucom.2024.128718 (2025). [ Google Scholar ] 6. Song, S., Liu, B., Teng, F. & Li, T. Self-supervised contrastive learning for implicit collaborative filtering. Eng. Appl. Artif. Intell. 139 , 109563. 10.1016/j.engappai.2024.109563 (2025). [ Google Scholar ] 7. Bathla, G., Singh, P., Singh, R. K., Cambria, E. & Tiwari, R. Intelligent fake reviews detection based on aspect extraction and analysis using deep learning. Neural Comput. Appl. 34 (22), 20213–20229. 10.1007/s00521-022-07531-8 (2022). [ Google Scholar ] 8. Duma, R. A., Niu, Z., Nyamawe, A. S. & Manjotho, A. A. An analysis of graph neural networks for fake review detection: A systematic literature review. Neurocomputing 623 , 129341. 10.1016/j.neucom.2025.129341 (2025). [ Google Scholar ] 9. Luo, J., Luo, J., Nan, G. & Li, D. Fake review detection system for online e-commerce platforms: A supervised general mixed probability approach. Decis. Support Syst. 172 , 114045. 10.1016/j.dss.2023.114045 (2023). [ Google Scholar ] 10. Xu, S., Cuan, H., Yin, Z. & Yin, C. A hybridized approach for enhanced fake review detection. IEEE Trans. Comput. Social Syst. 11 , 7448–7466. 10.1109/TCSS.2024.3411635 (2024). [ Google Scholar ] 11. Ibrahim, M., Bajwa, I. S., Sarwar, N., Hajjej, F. & Sakr, H. A. An intelligent hybrid neural collaborative filtering approach for true recommendations. IEEE Access. 11 , 64831–64849. 10.1109/ACCESS.2023.3289751 (2023). [ Google Scholar ] 12. Shi, X., Zhang, Y., Pujahari, A. & Mishra, S. K. When latent features meet side information: A preference relation based graph neural network for collaborative filtering. Expert Syst. Appl. 260 , 125423. 10.1016/j.eswa.2024.125423 (2025). [ Google Scholar ] 13. Zhang, S. et al. GCN-based user representation learning for unifying robust recommendation and fraudster detection. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval 689–698 (ACM, 2020). 10.1145/3397271.3401165. 14. Guo, Y. et al. A mixture-of-experts framework with fake review detection for robust recommendation systems (Springer, 2025). In: Lecture Notes in Computer Science10.1007/978-981-96-9875-2_15 15. Wang, W., Feng, F., He, X., Nie, L. & Chua, T. S. Denoising implicit feedback for recommendation. In Proceedings of the 14th ACM International Conference on Web Search and Data Mining (WSDM) 373–381 (ACM, 2021). 10.1145/3437963.3441800. 16. Ricci, F., Rokach, L. & Shapira, B. Recommender Systems Handbook 2nd edn (Springer, 2015). 10.1007/978-1-4899-7637-6. 17. Torkashvand, A., Jameii, S. & Reza, A. Deep learning-based collaborative filtering recommender systems: A comprehensive and systematic review. Neural Comput. Appl. 35 , 24783–24827. 10.1007/s00521-023-08958-3 (2023). [ Google Scholar ] 18. Wu, L., Xia, Y., Min, S. & Xia, Z. Deep attentive interest collaborative filtering for recommender systems. IEEE Trans. Emerg. Top. Comput. 12 (2), 467–481. 10.1109/TETC.2023.3286404 (2024). [ Google Scholar ] 19. Hasan, E., Rahman, M., Ding, C., Huang, J. X. & Raza, S. Review-based recommender systems: A survey of approaches, challenges and future perspectives. ACM Comput. Surveys . 58 (1), 1–41. 10.1145/3742421 (2026). [ Google Scholar ] 20. Gambetti, A. & Han, Q. Combat AI with AI: Counteract machine-generated fake restaurant reviews on online platforms. SSRN Electron. J. 10.2139/ssrn.4967813 (2024). [ Google Scholar ] 21. He, P., Gao, J. & Chen, W. DeBERTaV3: Improving DeBERTa using ELECTRA-style pre-training with gradient-disentangled embedding sharing. In International Conference on Learning Representations (ICLR) (OpenReview, 2023). 22. Hou, J., Tan, Z., Zhang, S., Hu, Q. & Wang, P. Detecting fake review intentions in the review context: A multimodal deep learning approach. Electron. Commer. Res. Appl. 70 , 101485. 10.1016/j.elerap.2025.101485 (2025). [ Google Scholar ] 23. Guo, G., Zhang, J. & Yorke-Smith, N. TrustSVD: Collaborative filtering with both the explicit and implicit influence of user trust and of item ratings. In Proceedings of the 29th AAAI Conference on Artificial Intelligence 123–129 (AAAI Press, 2015). 24. Jamali, M. & Ester, M. A matrix factorization technique with trust propagation for recommendation in social networks. In Proceedings of the 4th ACM Conference on Recommender Systems (RecSys) 135–142 (ACM, 2010). 10.1145/1864708.1864736. 25. Zhang, K. et al. Robust recommender system: A survey and future directions. ACM Comput. Surveys . 58 (1), 1–38. 10.1145/3757057 (2026). [ Google Scholar ] 26. Raghavan, S., Gunasekar, S. & Ghosh, J. Review quality aware collaborative filtering. In Proceedings of the 6th ACM Conference on Recommender Systems (RecSys) 123–130 (ACM, 2012). 10.1145/2365952.2365978. 27. Lyu, Y. et al. Reliable recommendation with review-level explanations. In 2021 IEEE 37th International Conference on Data Engineering (ICDE) 1641–1652 (IEEE, 2021). 10.1109/ICDE51399.2021.00137. 28. Li, B., Hou, Y., Dong, J., Yang, B. & Wang, X. ICRA: A study of highly accurate course recommendation models incorporating false review filtering and ERNIE 3.0. PLOS ONE . 19 (12), e0313928. 10.1371/journal.pone.0313928 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Rayana, S. & Akoglu, L. Collective opinion spam detection: Bridging review networks and metadata. In Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 985–994 (ACM, 2015). 10.1145/2783258.2783370. 30. Wang, Q. et al. Enhancing personalized recommendations: A study on the efficacy of multi-task learning and feature integration. Information 15 (6), 312. 10.3390/info15060312 (2024). [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement YelpCHI dataset: Chicago-area restaurant reviews with deception labels, available from 10.57760/sciencedb.17088. YelpNYC dataset: New York City restaurant reviews with deception labels, available from the same repository. Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (2.4 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Spammer Detection and Fake user Identification on Social Networks#88267
• A Survey on Fake Review Detection: From Pre-trained Language Models to Large Language Models#1104491
• RAEGNet: Relation-Aware Evidence Graph Network for Harm-Aware Multimodal Fake News Detection#1132103
• High-Fidelity Synthetic Dataset for Spam Detection#350245
• CogniDir: Combating Cognitive Malicious Comments via Adaptive Distributional Learning for Robust Fake News Detection#984906