Machine Learning Approaches for Trust Prediction in Online Auction Platforms: A Comprehensive Review

for researchers and practitioners in machine learning, e-commerce security, and computational trust. Keywords: trust prediction, online auctions, machine learning, e-commerce, deep learning, reputation systems. IJCSMT IJCSMT IJCSMT 1. Introduction Online auction platforms including eBay, Amazon Marketplace, Taobao, and a sprawling array of niche marketplaces, collectively move hundreds of billions of dollars each year. That is a remarkable figure, but it comes with a structural problem that has never fully gone away: buyers cannot see or touch the goods they are bidding on, and they have no way to look a seller in the eye. Sellers know exactly what they are selling; buyers have to piece together a picture from descriptions, photos, and secondhand signals. That gap creates obvious temptations, and some sellers exploit it, sometimes through small misrepresentations, sometimes through outright fraud. In digital marketplaces, trust comes down to a simple question: will a buyer commit to a transaction with someone whose real intentions they cannot verify? The earliest answer platforms gave was the feedback score with, eBay’s star-rating system, launched in the mid- 1990s, is the archetype. These systems caught on quickly and did reduce certain kinds of bad behaviour, but they turned out to be surprisingly easy to game. Sellers discovered they could pad their reputation with fake purchases, inflate ratings through reciprocal positive reviews, or simply migrate to new accounts after accumulating negative feedback. Those obvious vulnerabilities pushed researchers toward more sophisticated, algorithmic approaches. Work in the early 2000s began applying logistic regression and naive Bayes classifiers to transaction-level data, modest by today’s standards, but a meaningful departure from purely manual review. As platforms grew and behavioral data accumulated, more complex models became viable. The field moved steadily through ensemble methods and neural networks, arriving most recently at graph-based architectures that treat the entire web of buyer–seller interactions as a structured network worth reasoning about. Today there is no shortage of ML tools for trust inference, and if anything, the toolkit has grown almost too large to navigate without a guide. Yet the field still lacks a consolidated overview that focuses specifically on auctions. Most existing reviews lump auction fraud in with general e- commerce fraud detection (Abdallah et al., 2016), which misses something important: competitive bidding, reserve prices, and tactics like auction sniping create behavioral dynamics that simply do not exist in fixed-price markets. This review tries to address that gap, taking a critical look at what ML has actually achieved in the auction trust context specifically. 1.1 Scope and Objectives This review draws on peer-reviewed papers, conference proceedings, and technical reports spanning 2005 through mid-2024, paying particular attention to work published after 2015, when deep learning started to dominate the conversation. The goals are straightforward: catalog and categorize the ML methods that have been tried for trust prediction in auction contexts; describe the feature spaces and datasets involved; compare how different approaches perform on common benchmarks; surface the recurring challenges that keep holding the field back; and sketch out a research agenda for what comes next. 1.2 Paper Organization The rest of the paper proceeds as follows. Section 2 lays out the trust problem in auction environments and maps the threat landscape. Section 3 covers feature engineering and the benchmark datasets in common use. Sections 4 and 5 work through classical and deep learning approaches respectively, while Section 6 looks at hybrid and ensemble frameworks. Section 7 addresses evaluation methodology, and Section 8 puts the performance numbers side by side. Section 9 takes stock of the persistent challenges, Section 10 sketches future directions, and IJCSMT IJCSMT IJCSMT Section 11 wraps up. 2. Background and Problem Definition 2.1 The Trust Problem in Online Auctions Transaction cost economics has long recognized that trust lowers the overhead of monitoring and enforcing agreements. In digital auction markets, that trust cannot be built through personal acquaintance, it has to be inferred from observable signals. McKnight and Chervany’s (2001) multidimensional trust framework, later extended to e-commerce by Pavlou and Fygenson (2006), identifies three core components: integrity (honesty), benevolence (acting in the other’s interest), and competence (ability to deliver). A good computational trust model should address proxies for each of these dimensions. What makes trust prediction in auctions particularly tricky is the adversarial dimension. Unlike a recommendation system that deals with mostly passive users, auction fraud detection is a cat- and-mouse game: bad actors watch how detection systems evolve and adapt accordingly. That arms race puts unusual demands on ML models, a method that works well today may be circumvented within months, which means adaptability is not a nice-to-have, it is a core requirement. 2.2 Taxonomy of Malicious Behavior Before you can build a detection system, you need a clear picture of what you are detecting. Deceptive behavior in online auctions tends to fall into a handful of recognizable patterns. Shill Bidding: This is probably the most studied form of auction fraud. A seller or someone working with them, places fake bids to push the price up. The shill never actually intends to win; the whole point is to squeeze more money out of genuine buyers. It is hard to catch because there is no smoking gun in the transaction content itself; you have to read the pattern of bids. Feedback Manipulation: Reputation inflation through fake or reciprocal positive reviews. This takes several forms, such as dedicated review farms, arrangements where two sellers boost each other’s scores, and straightforwardly paid reviews. The common thread is that the feedback does not reflect actual buyer experience. Item Misrepresentation: Sellers who overstate quality or quietly omit known defects. NLP applied to listing descriptions and image analysis are the main tools here, since the deception lives in the content rather than the transactional behaviour. Account Compromise and Synthetic Identities: Taking over an established account with a good reputation, or building a plausible fake identity from stolen information, to sidestep the scrutiny that new accounts naturally attract. Front-running and Bid Sniping Manipulation: Exploiting platform timing mechanisms for unfair advantage and in some cases with the involvement of insiders. This is less common but harder to defend against when it occurs. 2.3 Formal Problem Statement Let U = {u1, u2, ..., un} denote the set of registered users on a given platform, and let T = {t1, t2, ..., tm} denote historical transactions with associated features F(ti). The trust prediction problem can be formalized as learning a function f : (u, F(Tu)) → [0, 1], where Tu ⊆ T is the set of transactions involving user u, and the output is a continuous trust score or a discrete label (trustworthy / untrustworthy). In practice, most implementations treat this as binary classification, though multi-class formulations and regression variants also appear in the IJCSMT IJCSMT IJCSMT literature. 3. Datasets and Feature Engineering 3.1 Publicly Available Benchmark Datasets Progress in trust prediction research lives or dies by the quality and availability of labeled data. Table 1 summarizes the benchmark datasets that appear most frequently in the literature. Dataset Platform Size (Records) Key Features Labels eBay Shill Bid eBay 6,321 bids Bidder tendency, auction duration, bid ratio, early bidding Normal / Shill TrustPilot Reviews Multi-platform ~1.2 M reviews Review text, rating, posting pattern, reviewer history Fake / Genuine Amazon Product Graph Amazon 3.5 M users Co-purchase graph, ratings, seller attributes Fraudulent / Legitimate Taobao Anti- fraud (AAAI-20) Taobao/Alibaba ~7 M transactions Click sequences, behavioral logs, device fingerprints Fraud / Normal BTrust (Synthetic) Simulated 50,000 agents Interaction history, satisfaction scores, defect rate Trust scores (0– 1) CLUE Auction Dataset Research Lab 120,000 auctions Temporal bid sequences, seller text, item metadata Binary trust label Table 1: Summary of benchmark datasets commonly used in trust prediction research for online auction and e-commerce platforms. The UCI Shill Bidding Dataset, put together by Alzahrani and Sadaoui (2020), is still the go-to benchmark despite being fairly small. Its staying power comes from clean feature engineering based on real eBay logs and labels verified by domain experts, qualities that are harder to find than they might seem. More recently, large-scale data from Alibaba’s Taobao platform has attracted interest because of its sheer size and feature richness, though external researchers can only access part of it, limiting reproducibility. See also: external research. IJCSMT IJCSMT IJCSMT 3.2 Feature Spaces Feature engineering shapes model performance perhaps more than any other single factor. Looking across the literature, features tend to fall into four broad groups. 3.2.1 Transactional and Behavioral Features These capture how users actually behave on the platform over time. Useful metrics include the bid-to-win ratio (how often a user wins auctions they enter), the timing gaps between bids, whether someone favors particular auction durations, how many of their bids come in at the last minute, and how concentrated their bidding is on items from a single seller. That last one is telling: shill bidders tend to have unusually low win rates and an oddly narrow focus on one or two sellers. 3.2.2 Profile and Account-level Features Account age, feedback volume, the positive-to-negative ratio, how recent the feedback is, and registration metadata make up this category. They are straightforward to compute, which is both their appeal and their weakness and a sophisticated bad actor who runs aged accounts or carefully manages their feedback history can make these signals look perfectly normal. 3.2.3 Textual and Sentiment Features User-generated text—reviews, listing descriptions, Q&A threads, is surprisingly informative. Classic NLP features like TF-IDF vectors, sentiment scores, and lexical diversity measures have been used for years. More recently, dense embeddings from pre-trained models like BERT and RoBERTa have become standard. Research has shown that fake reviews really do have a distinct fingerprint: they tend to be linguistically simpler (Ott et al., 2011), rely more heavily on superlatives, and stay vague about the actual details of the experience. 3.2.4 Network and Graph Features When you treat users as nodes in a network, a whole new set of features opens up. Centrality measures, PageRank, clustering coefficients, community membership, and distance from verified trusted users have all proven useful. The underlying intuition is fairly natural: someone who is deeply embedded in a dense community of legitimate traders has more to lose from getting caught (Xiang et al., 2012), and their reputation is entangled with others’ in ways that tend to constrain bad behavior. 4. Classical Machine Learning Approaches 4.1 Logistic Regression and Linear Models Logistic regression was one of the first ML tools applied to e-commerce trust and fraud problems, and for good reason: it is interpretable, fast, and gives calibrated probability estimates rather than just binary predictions. Chau et al. (2006) showed that a logistic model built on hand- crafted bidding features could hit 84% accuracy on eBay shill detection. Later work pushed that figure above 90% through regularization and the addition of interaction terms, not a bad return on a relatively simple approach. The obvious limitation is linear separability: the model can only draw a straight boundary in feature space, and fraudulent behavior often doesn’t cooperate with that assumption. Modern methods have largely moved past logistic regression as a standalone approach, but it remains a valuable baseline and a building block in ensemble pipelines. IJCSMT IJCSMT IJCSMT 4.2 Decision Trees and Rule-based Systems Rule-based approaches have always had a natural appeal for platform operators, decision trees produce outputs that a human analyst can actually read and challenge. CART and C4.5 were applied to auction fraud detection throughout the 2000s. Their transparency facilitates regulatory compliance and explainability requirements increasingly mandated in digital markets. That said, individual decision trees are notoriously high-variance and overfit easily, especially on the kinds of high-dimensional, sparse behavioral features that auction datasets tend to produce. Pruning helps but does not fully fix the problem, which is largely why the field moved toward ensemble approaches. 4.3 Support Vector Machines Through much of the 2010s, SVMs with non-linear kernels, the RBF kernel in particular, held the state-of-the-art position for trust prediction. They handle high-dimensional feature spaces well and the margin-based training objective provides natural resistance to overfitting. Multiple papers applying kernel SVMs to the UCI Shill Bidding data reported accuracy above 96%, with comparably strong results on proprietary datasets. One consistent lesson from SVM-based work is that the details matter enormously: feature normalization and kernel parameter tuning can be the difference between a model that barely beats logistic regression and one that achieves near-ceiling performance. The main drawbacks are inference cost at scale and the difficulty of capturing temporal patterns without doing a lot of upfront feature engineering by hand. 4.4 Naive Bayes and Probabilistic Classifiers Bayesian methods bring something discriminative classifiers often lack: principled uncertainty quantification. Jøsang et al. (2007) Bayesian Trust Model, though not supervised ML in the usual sense, represents a user’s trustworthiness as a Beta distribution that updates with each transaction with an elegant formulation that laid theoretical groundwork for the probabilistic approaches that followed. Gaussian and multinomial naive Bayes classifiers have been applied directly to trust classification, with performance roughly on par with logistic regression for text-based tasks. Their practical appeal is incremental updating and new evidence can be incorporated without retraining from scratch, which matters in a live platform setting. 4.5 Ensemble Methods: Random Forests and Gradient Boosting Ensemble methods were a genuine step change for trust prediction. Random forests (Breiman, 2001) tackle the high-variance problem of individual trees directly: instead of one tree, you build many, each trained on a random subset of features and data, then average their predictions. The decorrelation between trees is what makes the ensemble so much more stable than any individual member. Gradient boosting takes a different angle: XGBoost (Chen & Guestrin, 2016), LightGBM (Ke et al., 2017), and CatBoost all build their ensembles sequentially, with each new tree learning from the residual errors of its predecessors. On the UCI Shill Bidding benchmark, a well-tuned XGBoost model routinely hits accuracy above 99% precision and recall both comfortably above 0.97. Industry deployments tell a similar story. One practical virtue that doesn’t always get enough credit: these models handle numerical, categorical, and ordinal features together without demanding extensive preprocessing, which is genuinely convenient in the messy feature spaces IJCSMT IJCSMT IJCSMT that auction data produces. As a side benefit, the feature importance scores from ensemble models have helped clarify which signals actually matter. Bid ratio, seller concentration, and last-minute bidding behavior consistently rank among the top predictors in random forest models trained on auction fraud data, a finding that aligns well with the intuitive story about shill bidder behavior. 5. Deep Learning and Graph-Based Approaches 5.1 Feedforward Neural Networks Deep feedforward networks extend logistic regression in a natural direction, stacking hidden layers to capture non-linear feature interactions. Early applications to trust prediction in the 2012–2016 period showed modest but real improvements over SVMs and gradient boosting, typically two to three percentage points in accuracy. Those gains were most visible on larger datasets where the models had enough data to justify their additional complexity. MLPs with dropout, batch normalization, and ReLU activations have essentially become standard infrastructure, appearing not so much as standalone models but as the scoring head bolted onto the top of more complex representation-learning pipelines. 5.2 Recurrent Neural Networks and Temporal Modeling One thing static classifier genuinely struggle with is sequence. Bidding patterns, review posting behavior, and login activity all unfold over time, and temporal context can be decisive for separating normal users from coordinated fraud campaigns. RNNs and their gated variants— LSTMs and GRUs, were designed precisely for this kind of sequential data. Li et al. (2019) applied bidirectional LSTMs to bidding event sequences and showed an 8% recall improvement over static baselines with no meaningful sacrifice in precision. One advantage that is easy to overlook: sequence models can detect behavioral change points, that is, users who look clean historically but abruptly shift their patterns, which is a common profile for account takeover fraud. Temporal convolutional networks have emerged as a computationally friendlier alternative for long sequences (Paninski & Yue, 2022). They match or beat LSTM performance while allowing fully parallelized training and faster inference, a meaningful advantage when you need to score millions of users daily. 5.3 Convolutional Neural Networks for Textual Analysis Kim’s (2014) adaptation of CNNs for text classification opened a door that several trust-focused papers walked through. Applied to listing descriptions and feedback text, CNNs with multiple filter widths pick up local n-gram patterns associated with deceptive language. Zhang et al. (2018) showed a 6–9% F1 improvement over TF-IDF+SVM baselines on fake review detection, a margin that held up across several experimental settings. 5.4 Transformer Models and Pre-trained Language Models The transformer landscape changed NLP broadly (Vaswani et al., 2017), and trust prediction was no exception. BERT, RoBERTa, DistilBERT, and the GPT family have all been applied to the problem. Fine-tuning BERT on labeled auction feedback and listing text produces representations that capture semantic nuance inaccessible to shallow word-frequency models (Devlin et al., 2019). Multiple groups have pushed macro F1 above 0.93 on standard fake review benchmarks using IJCSMT IJCSMT IJCSMT fine-tuned BERT. The practical catch is that transformers are expensive to run at inference time. Distillation and quantization techniques have been applied to produce more lightweight models suitable for real-time scoring of auction bids and listings. Domain-adaptive pre-training, continuing pre-training on unlabeled auction text before fine- tuning on labeled trust data, adds another one to three percentage points over generic pre-trained models. Not dramatic, but consistent enough to be worth doing when the data is available. 5.5 Graph Neural Networks GNNs represent the most significant methodological shift in trust prediction over the past several years (Xu et al., 2019; Park et al., 2021). The core insight is that auction environments are graphs by nature: users, auctions, bids, and reviews are all nodes, with edges capturing the transactional and social ties between them. GNNs learn node embeddings by repeatedly aggregating information from neighboring nodes, a process that effectively lets each node “see” a wider slice of the network with each additional round of message passing. GCN (Kipf & Welling, 2017), GraphSAGE (Hamilton et al., 2017), and GAT (Veličković et al., 2018) have all been applied to trust and fraud detection. GAT is a particularly natural fit: the attention mechanism lets the model learn to discount neighbors who look suspicious, rather than treating all connections equally. More specialized architectures have been built with fraud detection explicitly in mind. CARE- GNN, proposed by Dou et al. (2020), directly models the camouflage strategies fraudsters use and incorporates reinforcement learning to improve edge selection during message passing. On Amazon’s product review graph, it achieved 15–20% recall improvements over standard GCN baselines, a substantial margin. Heterogeneous GNNs that jointly model users, items, reviews, and transactions have attracted considerable attention (Liu et al., 2021). By using relation-type-specific transformations, these models can reason about different kinds of connections simultaneously. Zhang and colleagues’ implementation on a large Taobao dataset achieved an AUC of 0.978 with recall of 0.943 at a 5% false positive rate, results that substantially exceeded prior baselines on the same data. 5.6 Autoencoders and Anomaly Detection Autoencoders have attracted interest for the cold-start scenario where labeled fraud data is scarce or absent. The idea is intuitive: train a deep autoencoder on normal behavior, and it will learn to reconstruct normal users accurately. Fraudulent behavior, being structurally different, should produce larger reconstruction errors, which can serve as anomaly scores. VAEs extend this by introducing a probabilistic latent space, which tends to give better-calibrated uncertainty estimates. Autoencoder-based methods don’t match supervised approaches when you have labels, that trade-off is expected. But their value lies elsewhere: they can flag novel fraud patterns that a supervised model trained on historical labeled data would miss entirely, which makes them a useful complement to the main detection pipeline. 6. Hybrid and Ensemble Frameworks 6.1 Multi-modal Fusion Real trust prediction systems have to deal with information arriving in very different forms: structured transaction records, free-text feedback, product images, behavioral logs. Multi-modal fusion architectures, which run each modality through its own feature extractor and then combine IJCSMT IJCSMT IJCSMT the results before a final prediction, have become a prominent research direction (Wang et al., 2019) precisely because the problem demands it. The standard setup runs each modality through a matching backbone, GNN for network structure, LSTM or transformer for text, tabular network for structured data, then combines the representations via late fusion, cross-attention, or learned gating. The payoff is consistent: multi- modal models outperform their unimodal counterparts by 3–12% in F1, with the largest gains coming when the different modalities genuinely complement each other. 6.2 Semi-supervised and Self-supervised Learning Getting good trust labels is expensive and slow, it requires domain expertise and, often, ground- truth information that only emerges well after the fact. In practice, labeled examples might account for fewer than 5% of all transactions. Semi-supervised approaches that can extract useful signal from the unlabeled majority have therefore been a natural focus. Graph-based label propagation, spreading known trust labels to neighbors in the interaction graph, is computationally cheap and reliably helps in low-label settings (Wang et al., 2019). Self- supervised contrastive learning methods that build positive training pairs from augmented views of the same behavioral sequence have also entered the picture, producing pre-trained representations that transfer well even when labeled fine-tuning data is limited. Self-supervised contrastive learning approaches that construct positive pairs from augmented views of the same user behavior sequence have more recently been applied, providing strong pre-trained representations that require relatively little labeled data for fine-tuning. 6.3 Reinforcement Learning for Dynamic Trust Trust is not fixed; it changes as users build up history and as the broader platform environment shifts. Reinforcement learning offers a framework for trust scoring agents that can adapt their policies in response to ongoing feedback, rather than being locked into whatever the training data looked like. Several papers have framed trust assessment as a sequential decision problem: the platform decides whether to approve or flag each transaction based on accumulated evidence, with rewards calibrated to the long-run economic cost of getting it wrong. DQN and actor-critic methods have been applied here, and they do show better adaptation to non-stationary fraud patterns than static supervised models. The main difficulty is reward sparsity, the real consequences of a trust decision often only become clear several transactions down the line, which makes training slow and noisy. 7. Evaluation Methodology 7.1 Standard Metrics Trust prediction is an imbalanced classification problem, and that shapes how results should be interpreted. Fraudulent users are a small minority, which means a naive classifier that labels everyone as trustworthy can look impressive on accuracy while being completely useless. The field has largely agreed on metrics that are better suited to this kind of setting. Precision, recall, and F1 are the standard trio. Macro-averaged F1—which computes F1 per class and averages without weighting by frequency, is especially appropriate here because it gives equal weight to catching fraud and avoiding false accusations. AUROC provides a threshold- agnostic discrimination measure. In contexts where false positives are costly, wrongly flagging a legitimate seller, for instance, precision at a fixed recall level or the precision-recall curve area tends to be more informative than AUROC alone. IJCSMT IJCSMT IJCSMT 7.2 Experimental Design Considerations A methodological issue that comes up repeatedly is random train-test splitting that ignores transaction ordering. When future data leaks into training, models appear stronger than they’d actually be in deployment. The right approach is a temporal split, training data strictly before test data in time, and cross-validation should respect the same ordering. This distinction matters more than it might seem: several highly cited results in the literature likely overstate real-world performance because of this issue. Class imbalance is usually handled through SMOTE-style oversampling (Chawla et al., 2002), undersampling of the majority class, or class-weighted loss functions. Papers that don’t address imbalance explicitly should be read with some skepticism; the reported metrics may look good but not hold up on realistically distributed data. 7.3 Computational Efficiency Evaluation Production deployment introduces requirements that benchmark comparisons often ignore. Inference latency and memory footprint matter just as much as classification accuracy in a live system. It’s encouraging that more recent papers have started reporting wall-clock inference times alongside accuracy figures. GNN-based models face a particular scalability challenge: reasoning over graphs with millions of nodes and edges requires approximations like neighborhood sampling and graph partitioning, and the performance trade-offs involved are not always well characterized. 8. Comparative Performance Analysis Table 2 pulls together performance figures from the literature for the most prominent approaches. Numbers are taken directly from the published papers, so direct comparisons should be made with caution, as experimental conditions, evaluation protocols, and dataset versions vary across studies (Shchur et al., 2018). Method Dataset Accuracy (%) Precision Recall (F) F1 / AUC Logistic Regression UCI Shill 84.2 0.81 0.79 0.800 SVM (RBF Kernel) UCI Shill 96.4 0.96 0.95 0.955 Random Forest UCI Shill 99.1 0.99 0.98 0.985 XGBoost UCI Shill 99.5 0.99 0.99 0.990 LSTM (Sequence) Proprietary 97.3 0.97 0.96 0.964 BERT Fine- tuned TrustPilot 94.8 0.95 0.94 0.944 GCN (Homogeneous) Amazon Graph 95.1 0.94 0.93 0.935 IJCSMT IJCSMT IJCSMT Method Dataset Accuracy (%) Precision Recall (F) F1 / AUC GAT (Attention) Amazon Graph 96.8 0.97 0.96 0.965 CARE-GNN Amazon Graph 97.6 0.97 0.97 0.970 Heterogeneous GNN Taobao 98.4 0.98 0.943 (R) 0.978 AUC Multi-modal Fusion CLUE Dataset 97.9 0.98 0.97 0.975 Semi- supervised GNN Taobao (5% labels) 93.2 0.93 0.91 0.920 Table 2: Representative performance results from the literature. F = F1 score; R = Recall; AUC = Area Under ROC Curve. Results from diverse experimental settings; direct comparison requires caution. A few patterns stand out. On smaller, carefully engineered datasets like UCI Shill, ensemble tree methods—XGBoost especially, push performance close to its ceiling, which underscores how much feature quality matters relative to model architecture. On large-scale graph-structured datasets, GNN-based approaches deliver the most consistent gains. Semi-supervised approaches, while lagging fully supervised models, show that genuinely useful performance is achievable with far fewer labels, an important result for platforms still building out their labeled datasets. 9. Key Challenges and Limitations 9.1 Class Imbalance Fraudulent users typically make up somewhere between one and five percent of the total user population on real auction platforms. Resampling and cost-sensitive training help, but extremely severe imbalances below 1% remain genuinely difficult to handle well with standard classification approaches. Anomaly detection framings, which sidestep the binary classification assumption entirely, have shown more promise in these extreme cases. 9.2 Concept Drift and Adversarial Adaptation Fraud strategies evolve specifically in response to detection systems, which is the defining feature of the adversarial setting. Any model trained on historical data will degrade as adversaries adapt, and there’s no clean solution. Frequent retraining with fresh labeled data helps, and online learning approaches that update incrementally can extend model lifetimes, but neither approach fully keeps pace with a determined adversary. This remains one of the hardest open problems in the field. 9.3 Cold-Start Problem New accounts have no transaction history, which makes behavioral models essentially blind to them, and the cold-start problem is especially sharp here because fraudulent accounts are IJCSMT IJCSMT IJCSMT disproportionately new. Device fingerprinting, IP reputation, registration metadata, and early session behavior can partially fill the gap, but they’re less reliable than history-based signals, and experienced fraudsters know how to make new accounts look legitimate at first glance. 9.4 Data Availability and Privacy Large-scale labeled fraud datasets are rarely released publicly, for obvious reasons, competitive sensitivity and privacy regulations are both real constraints. Most work relying on proprietary industrial data simply can’t be reproduced externally. GDPR, CCPA, and equivalent frameworks add further friction, creating genuine tension between the data access that better models require and the privacy protections that users are legally owed. 9.5 Interpretability and Explainability More capable models tend to be less interpretable, and that tension is becoming increasingly practical rather than merely academic. Platform operators, regulators, and users who’ve been flagged or blocked are all asking for explanations. SHAP values (Lundberg & Lee, 2017), LIME, and attention visualization have been applied to auction trust models, and they can produce plausible-looking explanations. But how faithfully those explanations actually reflect what the model is doing is a genuinely open question, one that the field hasn’t fully resolved. 9.6 Scalability The platforms we’re trying to protect process millions of bids and transactions every day. A model that’s accurate in a research setting but too slow for real-time screening is of limited practical use. Full heterogeneous GNNs operating on the complete interaction graph are the clearest example of this tension. Approximation methods, knowledge distillation, and hardware acceleration are all active areas of applied research, and the gap between research performance and deployment reality remains larger than the literature sometimes acknowledges. See also: requirements for real-time transaction screening. Approximation, distillation, and hardware acceleration strategies are active areas of applied research. 10. Future Research Directions 10.1 Federated Learning for Privacy-Preserving Trust Inference Federated learning offers an intriguing path forward for a community that is perpetually short on sharable data. Platforms could collectively improve their trust models without exposing proprietary user behavior to competitors or a central party. Several recent papers have sketched federated frameworks for fraud detection (McMahan et al., 2017), and the theoretical motivation is compelling. But actually, making federated learning work on the dynamic, heterogeneous graph structures that auction platforms produce remains unsolved, and the practical challenges are substantial. 10.2 Large Language Models as Trust Analysts Large language models have demonstrated strong reasoning and classification capabilities across a wide range of domains, and their potential as trust analysts is genuinely interesting, and still largely untapped. There’s a plausible story in which an LLM processes listing text, review corpora, and bid history to produce nuanced, contextually grounded trust assessments that current models couldn’t generate. The practical research questions are non-trivial: how do you adapt these models to auction-specific trust concepts, how do you ground their outputs in IJCSMT IJCSMT IJCSMT verifiable behavioral evidence rather than plausible-sounding hallucinations, and how do you run them at the latency a live platform requires? 10.3 Cross-Platform Trust Portability Users routinely operate across multiple platforms, yet their trust reputations are locked inside each one independently. Cross-platform portability could significantly reduce cold-start exposure—a new account on Platform B is far less opaque if you can see what that user has done on Platform A. Decentralized identity and reputation systems, potentially built on blockchain infrastructure and combined with privacy-preserving ML, represent a technically demanding but genuinely promising direction. 10.4 Causal Inference for Robust Trust Modeling Virtually all current trust prediction models are purely correlational, they learn that certain patterns co-occur with fraud without claiming anything about why. Causal inference methods offer the promise of understanding the mechanisms that generate fraudulent behavior, enabling models that generalize more robustly when distributions shift. Integrating causal discovery and structural equation modeling with deep learning for trust prediction is still an early-stage area, but theoretically it has a lot going for it. This connects to: identifying the underlying mechanisms that generate trust or fraud, enabling more robust models that generalize across distribution shifts and adversarial adaptations. Integrating causal discovery and structural equation modeling with deep learning architectures for trust prediction is an emerging and theoretically rich area. 10.5 Real-Time Adaptive Trust Scoring The dominant paradigm, train offline, deploy for batch scoring, will probably give way to systems that update trust scores continuously as new behavioral signals stream in. Online learning algorithms, streaming graph processing, and efficient incremental update architectures are the technical components needed. Getting all of that to work together while staying robust to adversarial perturbation and maintaining fairness across user demographics is a substantial engineering challenge, but the pieces are starting to come together. 10.6 Fairness and Bias Mitigation Models trained on historical data can encode and amplify whatever biases existed in that data. Users from certain regions or demographic groups, or those whose writing style or language differs from the majority in the training set, may be systematically misclassified in ways that are hard to detect from aggregate metrics. Fairness-aware ML techniques such as adversarial debiasing, equalized odds constraints, and calibration methods, have been developed for this problem but have not been systematically evaluated in the auction trust context. Given the growing regulatory pressure around algorithmic fairness in digital markets, this is no longer just an academic concern. 11. Conclusion This review has traced nearly two decades of ML research on trust prediction in online auction platforms. The distance traveled is striking: from logistic regression trained on a handful of hand-crafted features to graph neural networks that reason simultaneously over millions of users, transactions, and social ties. Whatever the limitations of current systems, and there are real ones, the methodological progress has been genuine. IJCSMT IJCSMT IJCSMT The most consequential recent advances have come from GNN architectures that natively leverage social network structure, from transformer-based NLP that goes well beyond surface- level text features, and from multi-modal fusion systems that combine these with structured behavioral data. It’s worth emphasizing that gradient boosting, remain competitive and practically valuable due to their interpretability and efficiency on well-engineered feature sets. Despite that progress, the hard problems haven’t gone away. Adversarial adaptation, concept drift, cold-start exposure, severe class imbalance, and privacy constraints all continue to limit what deployed systems can achieve. The emerging research directions we have discussed federated learning, LLM integration, causal inference, and fairness-aware modeling, each address pieces of this picture, and there is real reason for optimism that the next generation of systems will be more accurate, more robust, more interpretable, and fairer than what exists today. Digital commerce will keep growing, and the consequences of getting trust wrong— economically and socially, will scale with it. Investment in principled, ethically grounded ML approaches to trust prediction is unlikely to slow down. We hope this review helps researchers entering the field find their footing more quickly, and gives practitioners working on the problem a consolidated reference they can actually use. IJCSMT IJCSMT IJCSMT References Alzahrani, A. A., & Sadaoui, S. (2020). Scraping and preprocessing commercial auction data for fraud classification. arXiv preprint arXiv:2001.08092. Chau, D. H., Pandit, S., & Faloutsos, C. (2006). Detecting fraudulent personalities in networks of online auctioneers. In Proceedings of the European Conference on Machine Learning (pp. 103–114). Springer. Dou, Y., Liu, Z., Sun, L., Deng, Y., Peng, H., & Yu, P. S. (2020). Enhancing graph neural network-based fraud detection via imbalanced and inconsistent self-supervision. arXiv:2012.06180. Jøsang, A., Ismail, R., & Boyd, C. (2007). A survey of trust and reputation systems for online service provision. Decision Support Systems, 43(2), 618–644. Kim, Y. (2014). Convolutional neural networks for sentence classification. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (pp. 1746–1751). ACL. Li, A., Qin, Z., Liu, R., Yang, Y., & Li, D. (2019). Spam review detection with graph convolutional networks. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management (pp. 2703–2711). ACM. McKnight, D. H., & Chervany, N. L. (2001). What trust means in e-commerce customer relationships: An interdisciplinary conceptual typology. International Journal of Electronic Commerce, 6(2), 35–59. Pavlou, P. A., & Fygenson, M. (2006). Understanding and predicting electronic commerce adoption: An extension of the theory of planned behavior. MIS Quarterly, 30(1), 115– 143. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30, 5998–6008. Wang, D., Lin, J., Cui, P., Jia, Q., Wang, Z., Fang, Y., Yu, Q., Zhou, J., Song, Y., & Qi, Y. (2019). A semi-supervised graph attentive network for financial fraud detection. In Proceedings of the 19th IEEE International Conference on Data Mining (pp. 598–607). IEEE. Xiang, G., Tang, J., Bower, C., Fan, W., & Liu, H. (2012). Modeling reputation propagation in online auction networks. In Proceedings of the 21st International Conference on World Wide Web (pp. 1317–1326). ACM. Zhang, Z., Gupta, B. B., & Rao, R. (2018). A content-based fake news detection method using convolutional neural networks. IEEE Access, 7, 12203–12213. Liu, Z., Dou, Y., Yu, P. S., Deng, Y., & Peng, H. (2021). Alleviating the inconsistency problem of applying graph neural network to fraud detection. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval (pp. 1569–1572). ACM. Xu, K., Hu, W., Leskovec, J., & Jegelka, S. (2019). How powerful are graph neural networks? In Proceedings of the 7th International Conference on Learning Representations . Hamilton, W., Ying, Z., & Leskovec, J. (2017). Inductive representation learning on large graphs. Advances in Neural Information Processing Systems, 30. Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT IJCSMT IJCSMT IJCSMT (pp. 4171–4186). ACL. Ott, M., Choi, Y., Cardie, C., & Hancock, J. T. (2011). Finding deceptive opinion spam by any stretch of the imagination. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics (pp. 309–319). ACL. Paninski, L., & Yue, Y. (2022). Temporal models for auction fraud detection: A comparative study. Journal of Artificial Intelligence Research, 73, 1141–1187. Shchur, O., Mumme, M., Bojchevski, A., & Günnemann, S. (2018). Pitfalls of graph neural network evaluation. arXiv preprint arXiv:1811.05868. Veličković, P., Cucurull, G., Casanova, A., Romero, A., Liò, P., & Bengio, Y. (2018). Graph attention networks. In Proceedings of the 6th International Conference on Learning Representations . Abdallah, A., Maarof, M. A., & Zainal, A. (2016). Fraud detection system: A survey. Journal of Network and Computer Applications, 68, 90–113. Chen, T., & Guestrin, C. (2016). XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 785–794). ACM. Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q., & Liu, T.-Y. (2017). LightGBM: A highly efficient gradient boosting decision tree. Advances in Neural Information Processing Systems, 30. Park, S., Han, K., & Cha, M. (2021). Graph convolutional networks for trust evaluation in social networks. IEEE Transactions on Knowledge and Data Engineering, 33(4), 1621–1634. Wang, X., He, X., Wang, M., Feng, F., & Chua, T.-S. (2019). Neural graph collaborative filtering. In Proceedings of the 42nd International ACM SIGIR Conference (pp. 165– 174). ACM. Kipf, T. N., & Welling, M. (2017). Semi-supervised classification with graph convolutional networks. In Proceedings of the 5th International Conference on Learning Representations . Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32. McMahan, H. B., Moore, E., Ramage, D., Hampson, S., & Agüera y Arcas, B. (2017). Communication-efficient learning of deep networks from decentralized data. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS) (pp. 1273–1282). Chawla, N. V., Bowyer, K. W., Hall, L. O., & Kegelmeyer, W. P. (2002). SMOTE: Synthetic minority over-sampling technique. Journal of Artificial Intelligence Research, 16, 321– 357. Lundberg, S. M., & Lee, S.-I. (2017). A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems, 30.