A new study published in Discover Informatics by Thi Khanh Hoai Nguyen and Chaochang Chiu of Yuan Ze University in Taiwan tackles a deceptively simple question: when graph neural networks outperform traditional fraud screening, where exactly does that advantage come from? Rather than assuming that deep learning automatically dominates rule-based methods, the researchers built a controlled experimental framework to measure the marginal contribution of each component in a fraud detection pipeline. Their central finding is counterintuitive and operationally important: the value of graph learning in synthetic identity fraud detection comes from the completeness of the identity relations the model can access, not from increasingly sophisticated ways of blending graph scores with rule scores.

The study exploits a structural weakness that fraudsters cannot easily avoid. Fabricating a fully unique identity for every fraudulent profile is expensive, so offenders routinely recycle a limited pool of identifiers, reusing the same Social Security Numbers, email addresses, and telephone numbers across many customer accounts. This reuse links otherwise unrelated profiles into a web of shared identifiers, forming the empirical basis for identity-linkage analysis. The researchers modeled this structure as a heterogeneous graph with four node types—Client, SSN, Email, and Phone—connected by typed relations. When several fraudulent clients converge on a single SSN node, a detectable fraud ring appears in the topology, which is precisely the signature that linkage-based screening exploits.

The data came from a PaySim-derived identity linkage dataset containing 2,433 customer profiles, of which 433 were fraudulent, a prevalence of about 17.8 percent. Across the dataset, roughly 12 to 13 percent of clients shared each identifier type with at least one other client, and fraudulent clients shared identifiers at roughly 2.6 times the rate of normal clients. At zero noise, normal clients had an average rule score of just 0.003, with only 0.25 percent sharing at least one identifier, while fraudulent clients averaged a rule score of 4.947, with 76.44 percent sharing at least one identifier. Even at the highest tested noise level, 64 percent of fraudulent clients still shared at least one identifier, explaining why a simple counting rule is such a formidable baseline.

The evaluation was organized in two stages under a single controlled protocol. In the first stage, four model families were compared independently: transparent rule-based scoring, non-graph machine learning such as logistic regression and gradient boosting, lightweight label propagation methods, and deep graph models including graph convolutional networks and the Heterogeneous Graph Transformer, or HGT. In the second stage, representative models were anchored to the rule core in a hybrid pipeline, so the incremental value of each component could be isolated. The rule itself is elegantly simple: each client is scored by how many other clients share its SSN, email, or phone, a metric that requires no training and is fully auditable by regulators and compliance teams.

The results were revealing. The rule alone achieved an average Precision-Recall Area Under the Curve, or PR-AUC, of 0.777. Non-graph machine learning models produced results almost identical to the rule, confirming that when their input features are derived from the same linkage counts, conventional learners extract nothing new. LabelSpreading, an iterative diffusion method with no deep encoder, was the strongest standalone model at 0.807, a finding the authors emphasize should elevate propagation methods to the status of serious competitors rather than weak baselines. Among deep models, R-GCN collapsed to near-random performance due to a representational failure in which embedding variance fell to approximately 8 times 10 to the minus 11, while HGT-based models performed strongly.

The standout hybrid was Rule + HGT+SVM, which combined the rule score with a Heterogeneous Graph Transformer encoder feeding a support vector machine classifier. It achieved the best average PR-AUC of 0.835, a gain of 0.049 over the rule alone, with a positive difference in all 20 seed-noise combinations tested. The hybrid also degraded more gracefully under corruption: from 0 to 40 percent fraud-only noise, it lost 0.092 PR-AUC compared with 0.116 for the rule alone, and at the highest noise level it retained 0.768 PR-AUC against 0.704 for the rule. In fraud screening, where reviewers can only examine a limited number of flagged profiles per day, such gains in ranking quality translate directly into additional true positives surfaced within a fixed review budget.

What makes the study distinctive is its forensic dissection of why the hybrid wins. Because rule and graph scores ranked clients almost identically, the researchers tested whether two more elaborate fusion mechanisms—an adaptive per-client weight and a tie-breaking scheme—could extract further gains. Neither could. The fixed-weight fusion preserved the rule’s ordering for 100 percent of pairs with different rule scores in every run, and 99.5 percent of clients shared their exact integer rule score with another client, making ties the typical case rather than the exception. Within tied groups, every fusion formula reduced to the same graph-only ordering, which was only marginally better than a random tie order at 51.1 percent accuracy. The fusion weight, in other words, was not the source of the improvement at all.

The real answer came from a relation-level ablation. Removing SSN, Email, or Phone individually from the graph before training consistently reduced HGT+SVM performance across all tested seed-noise configurations, with SSN removal producing the largest average degradation. Crucially, a density-matched control—randomly deleting the same number of edges while keeping all three relation types—performed worse than structural relation removal, with nominal p-values of 0.0007 or lower in the descriptive 20-pair comparison. This means a graph missing one relation entirely but retaining the others in full is a better input to the encoder than a graph that keeps all relations but thinned. What matters is having some relations represented completely rather than all relations represented but degraded, a finding that reframes how hybrid fraud systems should be designed.

The authors are candid about the limits of their work. The dataset is small and synthetic, the noise protocols are simpler than real adversarial behavior, and with only four independent seeds the conservative seed-level statistical tests cannot reach conventional significance thresholds, so the findings are best read as directionally consistent evidence rather than confirmatory proof. A further negative result stands out: in an inductive setting where new clients are attached to the graph only at inference, the two-layer HGT architecture collapsed to random ranking, meaning the hybrid’s advantage is currently validated only for retrospective, transductive batch screening of an existing customer book. Even so, the practical message is clear. Transparent rule-based identity-linkage scoring remains the necessary, auditable backbone of a defensible fraud pipeline, while heterogeneous graph models serve as a complementary refinement layer whose value flows from relation-aware representation—not from fancier score fusion.

Subject of Research: Graph-based machine learning methods for detecting synthetic identity fraud in payment systems

Article Title: Rule anchored graph learning improves synthetic identity fraud detection through multi relation structure rather than score level fusion

Article References: Nguyen, T. K. H., & Chiu, C. (2026). Rule anchored graph learning improves synthetic identity fraud detection through multi relation structure rather than score level fusion. Discover Informatics, 1(1), Article 22. https://doi.org/10.1007/s44564-026-00024-z

Image Credits: AI Generated

DOI: 10.1007/s44564-026-00024-z

Keywords: synthetic identity fraud, graph neural networks, heterogeneous graph transformer, identity linkage, fraud detection, rule-based scoring, score fusion, label propagation, PR-AUC, noise robustness, machine learning, financial crime

The study, published in the International Journal of Machine Learning and Cybernetics by Liuhang Zhao, Qianwen Huang, and Huaiguang Liu, introduces a model called GAM-DRCN, short for a Gramian Angular Field and attention mechanism-based multiscale denoising residual convolutional neural network. The name is a mouthful, but each element addresses a specific and stubborn problem in the field of machinery health monitoring. Traditional convolutional neural network models for bearing monitoring typically rest on an idealized assumption: that the vibration signals fed into them are relatively clean, well-behaved recordings in which the diagnostic features stand out clearly against the background. In the real world of a working ship, that assumption collapses almost immediately. Engine room machinery, hull vibrations, wave loading, and propeller turbulence all conspire to bury the subtle signatures of bearing condition inside a dense fog of noise, and the features that do survive tend to follow nonlinear distributions that simple models struggle to untangle.

The first clever trick in the new architecture lies in how it represents the raw data. Vibration signals from a thrust bearing arrive as one-dimensional time series, essentially long strings of amplitude measurements sampled thousands of times per second. Convolutional neural networks, which have revolutionized image recognition, perform best when given two-dimensional inputs that preserve spatial relationships. The researchers therefore employed a mathematical transformation known as the Gramian Angular Difference Field, or GADF, which converts a one-dimensional time series into a two-dimensional feature map. The technique works by rescaling the signal values to angular coordinates and then computing the trigonometric relationships between every pair of points in the series. The result is a matrix, and therefore an image, in which the temporal correlations of the original signal are encoded as geometric patterns. Periodic structures, transient shocks, and drifting trends each paint distinctive textures onto this canvas, allowing a vision-oriented network to detect patterns that would be nearly invisible in the raw waveform.

Converting signals to images, however, is only half the battle, because noise contaminates those images just as thoroughly as it contaminates the original data. This is where the second pillar of the architecture comes into play: a multiscale denoising module, abbreviated MSS in the paper. The module operates on the principle that useful diagnostic information in vibration signals lives at multiple characteristic scales simultaneously. A developing thermal fault in a bearing might manifest as slow modulations of the vibration envelope, as changes in mid-frequency resonance bands, and as alterations in high-frequency micro-impacts, all at once. A network that examines the signal through only a single filter size risks missing whichever scale carries the strongest clue. The multiscale denoising module therefore applies parallel convolutional pathways of different receptive field sizes, extracting features at several granularities while actively suppressing the noise components that pervade each band. Crucially, the researchers built this capability on a residual backbone, the now-classic deep learning design in which identity shortcuts allow information to bypass layers, enabling very deep networks to train stably without suffering from vanishing gradients.

Extracting multiscale features creates its own dilemma, though: which of the many extracted features actually matter for the task at hand? To resolve this, the team incorporated a convolutional block attention module, known widely in the computer vision community as CBAM. Attention mechanisms allow a neural network to learn where to look, dynamically weighting channels and spatial locations according to their relevance to the classification objective. In this application, CBAM effectively acts as an intelligent filter that integrates the multiscale feature maps, amplifying the channels that carry genuine thermal state information and dampening those dominated by noise or irrelevant machinery activity. The combination is elegant in its division of labor: the GADF transformation makes the signal visible to vision-based learning, the multiscale denoising module cleans and enriches the representation across scales, and the attention module decides which of those enriched features deserve to influence the final prediction of the bearing’s oil temperature state.

To test whether this elaborate pipeline actually delivered, the researchers subjected GAM-DRCN to extensive experimental validation under two distinct operational conditions of a marine thrust bearing. They evaluated the model using the standard metrics of classification science: precision, recall, and the F1 score, which balances the two. They also performed robustness analysis by deliberately contaminating the test signals with synthetic noise at controlled signal-to-noise ratios, and they used t-SNE visualization, a dimensionality reduction technique, to inspect how cleanly the network separated different thermal states in its learned feature space. The results were striking. Under the two operating conditions, the model achieved optimal accuracies of 99.02 percent and 99.13 percent, outperforming a suite of baseline models drawn from the existing literature. In the visualization analysis, the learned representations formed well-separated clusters corresponding to different thermal states, providing visual confirmation that the network had discovered physically meaningful structure in the data rather than merely memorizing training examples.

The most impressive numbers, however, emerged from the noise stress tests. When the researchers degraded the signals to a signal-to-noise ratio of minus six decibels, a condition in which the noise power is nearly four times the signal power and the vibration data sounds, to human ears, like pure static, GAM-DRCN still delivered average accuracy rates of 86.41 percent and 90.35 percent across the two operating conditions. For context, many conventional diagnostic networks that perform well on clean laboratory data fall to near-chance performance under such severe contamination. This robustness is precisely the property that matters for real deployment, because a monitoring system that only works in a quiet engine room is a monitoring system that fails exactly when it is needed most, during the rough, loud, and unpredictable conditions of actual seagoing operation.

The team also conducted ablation experiments, systematically removing individual components of the architecture to measure each one’s contribution. These experiments demonstrated the respective impact of both the multiscale denoising module and the convolutional block attention module on overall performance, confirming that the gains were not simply a byproduct of adding more layers or parameters. Each module earned its place in the design. This kind of component-level validation is increasingly important in a field crowded with architectures whose complexity sometimes outpaces their justification, and it lends the reported results a credibility that raw accuracy figures alone cannot provide.

The broader significance of this work extends well beyond a single bearing type. Predictive maintenance powered by machine learning has become one of the most economically consequential applications of artificial intelligence in heavy industry, with documented use cases spanning automotive manufacturing, wind turbines, gearboxes, diesel generators, and rotating machinery of every description. Failures of marine propulsion systems are particularly costly, since a disabled thrust bearing can strand a vessel at sea, and historical failure analyses of submarine thrust bearings have highlighted how design and operational factors combine to produce dangerous thermal conditions. By inferring oil temperature, a direct indicator of friction, load distribution, and lubrication adequacy, from vibration data alone, the new approach reduces the dependence on dedicated temperature sensors embedded in the bearing bush, which are themselves vulnerable to failure and difficult to replace in service. Prior research has explored fiber Bragg grating sensing and other direct measurement techniques for bearing bush temperature, but the ability to derive thermal state information indirectly from ubiquitous vibration sensors offers a complementary and often more practical pathway.

The research, supported by the National Natural Science Foundation of China, points toward a future in which the machinery of global shipping monitors its own health continuously and intelligently, flagging thermal distress long before it escalates into failure. The authors report no competing financial interests, and their model’s demonstrated resilience to severe noise suggests a realistic path from laboratory validation to engine room deployment. As vessels grow larger, automation deepens, and the demand for reliable maritime logistics intensifies, technologies like GAM-DRCN represent the quiet revolution underway in how humanity keeps its machines alive: not by inspecting them more often, but by teaching them, through the mathematics of angular fields, multiscale filtering, and learned attention, to speak clearly about their own condition even when the world around them is deafening.

Subject of Research: Deep learning-based oil temperature monitoring of marine thrust bearings using vibration signals

Article Title: Multiscale denoising residual convolutional neural network for oil temperature monitoring of marine thrust bearings

Article References: Zhao, L., Huang, Q., & Liu, H. (2026). Multiscale denoising residual convolutional neural network for oil temperature monitoring of marine thrust bearings. International Journal of Machine Learning and Cybernetics, 17(10), Article 482. https://doi.org/10.1007/s13042-026-03319-7

Image Credits: AI Generated

DOI: 10.1007/s13042-026-03319-7

Keywords: marine thrust bearings, oil temperature monitoring, convolutional neural network, Gramian angular difference field, attention mechanism, multiscale denoising, vibration signals, predictive maintenance, fault diagnosis, machine learning, marine engineering, noise robustness

The work, led by Petr Ptáček together with Paulina Lewandowska and Ryszard Kukulski, both of whom contributed equally, addresses one of the most pressing practical constraints in near-term quantum computing. Devices in the so-called NISQ era, a term coined by John Preskill, operate without full quantum error correction. Every circuit execution is subject to noise, queue times on shared hardware are long, and the cost of running a model scales with the number of shots required per prediction. If quantum machine learning is ever to leave the laboratory and compete with classical methods, reducing the number of circuit evaluations is arguably as important as improving raw accuracy.

The core idea behind the new classifier lies in how it reads out answers from the quantum state. Conventional variational quantum classifiers typically measure the expectation value of an observable, often a Pauli operator, on the output state produced by a parameterized ansatz circuit. This expectation value is then thresholded to assign a class label. The problem is statistical: to estimate an expectation value to a given precision, the circuit must be run repeatedly, and the number of repetitions grows quadratically with the desired precision. The Ostrava team instead draws on the concept of unambiguous state discrimination, in which measurements are designed so that outcomes are either conclusive or explicitly inconclusive, never misleading. By measuring in a way that compares computational basis strings through Hamming distance, the classifier obtains richer, more informative samples from each circuit run.

Hamming distance, the number of bit positions in which two binary strings differ, has a precedent in quantum algorithms for classification. Earlier work on quantum k-nearest-neighbor algorithms used Hamming distance as a similarity metric between encoded data points. The new approach folds that metric into a variational framework: the parameterized circuit transforms and encodes data, and the measurement stage compares the resulting bit strings against reference patterns. Classical post-processing then weighs the conclusive outcomes to produce a classification decision. Because each shot carries more decision-relevant information, far fewer shots are needed per prediction, and the ansatz’s expressivity is exploited more effectively rather than being diluted by coarse averaging.

The theoretical backing matters here. The authors substantiate their experimental results with formal evidence supporting why the approach should perform well, rather than merely reporting empirical wins. This kind of grounding is notable in a field where many proposed quantum machine learning methods have been criticized for lacking provable advantages or for suffering from trainability pathologies such as barren plateaus, the flat regions of the training landscape described by McClean and colleagues in 2018. By tying the measurement scheme to information-theoretic principles and to established discrimination theory, the team provides a rationale for both the accuracy gains and the resource savings.

The empirical testbed was a demanding and socially significant one: the Wisconsin Diagnostic Breast Cancer dataset from the UCI Machine Learning Repository, a standard benchmark in medical classification involving distinguishing malignant from benign tumors based on features derived from digitized images of fine needle aspirate samples. The choice is apt for demonstrating practical relevance, since medical decision support is exactly the kind of domain where classification errors carry real costs and where the efficiency of a model matters if it is ever to run on scarce quantum hardware.

The headline numbers are striking. The unambiguous quantum classifier achieved an average accuracy of 90 percent on the breast cancer dataset, an improvement of 6.9 percentage points over the baseline variational quantum classifier. At the same time, it required eight times fewer circuit executions per prediction. That combination, better accuracy and an eightfold reduction in execution cost, is unusual in quantum machine learning, where improvements in one metric frequently come at the expense of the other. The savings compound across training as well: since model training involves evaluating the objective function many times over many optimization steps, cutting shots per evaluation by a factor of eight can dramatically shorten wall-clock training time and reduce access fees on cloud quantum platforms.

Noise robustness is the second major finding. When noise was injected into the simulations to emulate realistic hardware conditions, the accuracy advantage shrank from 6.9 to approximately 3.1 percentage points, but the eightfold reduction in execution cost persisted. The fact that the method degrades gracefully rather than collapsing is crucial. Many quantum algorithms that look compelling in idealized simulations lose their advantage entirely under realistic noise levels. A classifier that retains a meaningful improvement over its baseline while remaining dramatically cheaper to execute is far more plausible as a candidate for deployment on actual quantum processors, where gate errors, decoherence, and readout imperfections are unavoidable facts of life.

The authors implemented and evaluated their method using the PennyLane framework, the widely used open-source library for hybrid quantum-classical computation, and they have made both the code and the data openly available in a public GitHub repository. The optimizations were handled with classical techniques suited to noisy objective functions, including simultaneous perturbation stochastic approximation, an optimizer originally developed by Spall that estimates gradients from very few function evaluations, a natural pairing with a classifier designed to be frugal with circuit runs.

The broader significance of the study lies in what it suggests about where quantum advantage might first materialize in machine learning. Rather than waiting for large fault-tolerant machines, resource-efficient redesigns of existing algorithms could deliver practical value on today’s hardware. Related efforts in the literature have pursued shot optimization, quantum kernel methods, and data re-uploading schemes, and recent theoretical work on single-shot quantum machine learning has explored how few measurements are truly needed. The Ostrava results sit squarely in this emerging conversation, offering a concrete demonstration that smarter measurement and post-processing can unlock both accuracy and efficiency. If follow-up work confirms these gains on physical quantum processors and across additional datasets, the unambiguous classifier could become a template for building quantum machine learning models that are genuinely competitive, not just conceptually interesting. The research also underscores the value of collaboration between quantum algorithm theorists and application domain experts, a combination that will be essential as the field moves from proof-of-concept demonstrations toward tools that practitioners in medicine, materials science, and beyond can actually rely upon.

Subject of Research: A resource-efficient variational quantum classifier using Hamming distance measurements and classical post-processing

Article Title: Resource-efficient variational quantum classifier

Article References: Resource-efficient variational quantum classifier. (n.d.). https://doi.org/10.1007/s42484-026-00439-9

Image Credits: AI Generated

DOI: 10.1007/s42484-026-00439-9

Keywords: quantum machine learning, variational quantum classifier, Hamming distance, unambiguous state discrimination, NISQ era, breast cancer dataset, circuit evaluations, noise robustness, PennyLane, binary classification, variational circuits, resource efficiency

The new system, called ANDF, short for aspect-centric, noise-resilient cross-modal dynamic fusion, tackles a task known as multi-modal aspect-based sentiment analysis, or MABSA. Unlike ordinary sentiment analysis, which assigns a single positive, negative, or neutral label to an entire post, MABSA asks a more delicate question: what sentiment does the author express toward a specific aspect mentioned in the text? A restaurant review might praise the food while mocking the décor, and a photo attached to the post may show only the dining room. Getting the answer right requires the model to bind the correct words to the correct visual evidence while ignoring everything else.

That binding process, known as cross-modal alignment, is where most current models stumble. Text and images live in fundamentally different mathematical spaces, and aligning a short phrase like “the battery life” with the relevant region of a photograph is far harder than aligning two sentences. The problem is compounded by visual noise: real-world images are full of background objects, lighting artifacts, and clutter that have nothing to do with the aspect under discussion. A model asked about a phone’s screen may be distracted by the hand holding it, the table beneath it, or the coffee cup beside it.

ANDF addresses these challenges through three interlocking innovations. The first is an Aspect-Prompt Encoding strategy, which generates aspect-aware query features that guide the search for relevant visual content. Rather than treating the image as a whole and hoping the model figures out what matters, the framework uses the aspect term itself as a prompt, steering the encoding process toward fine-grained alignment between specific words and specific visual regions. This is analogous to giving a human reader a highlighter and instructions about what to look for before they open the image.

The second innovation is an Aspect-Centric Visual Disentanglement module. Using a dynamic gating mechanism, this module separates foreground information relevant to the aspect from noisy background content. Dynamic gating works like a set of adjustable valves: learned gates open or close pathways for different visual features depending on their relevance to the aspect at hand. The output is a set of robust, aspect-aware visual prototypes, essentially cleaned-up summaries of what the image actually says about the topic in question. By decoupling signal from noise before fusion, the framework avoids the common failure mode in which irrelevant visual details contaminate the sentiment prediction.

The third component, a Prototype-Guided Fusion module, brings the pieces together. Built on a Transformer-based architecture, the same family of structures that powers modern large language models, this module aggregates features from multiple sources dynamically and complementarily. Instead of fusing text and image representations with fixed weights, the module lets each aspect determine how much to trust each source. When the text is ambiguous, the visual prototype can carry more weight; when the image is cluttered or uninformative, the text can dominate. This flexibility is what the authors mean by dynamic fusion, and it is central to the framework’s resilience.

The experimental evidence is substantial. The team evaluated ANDF on two widely used benchmarks, Twitter-2015 with 2,166 samples and Twitter-2017 with 5,818 samples, both consisting of real social media posts paired with images. On Twitter-2015, ANDF achieved the highest accuracy among all compared baselines at 79.27 percent, edging out the previous best method, AMIFN, which scored 78.69 percent, by 0.58 percentage points. On Twitter-2017, generally considered the more challenging benchmark, ANDF attained a leading F1-score of 71.76 percent, surpassing every baseline tested. The F1-score, which balances precision and recall, is particularly informative on imbalanced datasets where accuracy alone can be misleading.

Perhaps more telling than the headline numbers are the controlled perturbation experiments. The researchers deliberately degraded test images in four ways: partial occlusion, fine-grained noise, semantic mismatch between text and image, and structural disruption. ANDF remained resilient across all four conditions, suggesting that its noise-suppression mechanisms are not merely artifacts of clean benchmark data but genuine robustness properties. Ablation studies, in which individual modules are removed to measure their contribution, confirmed that each of the three core components measurably improves cross-modal semantic alignment and noise suppression. The team also ran the full experiment across nine random seeds, reporting per-seed accuracy and F1 values to demonstrate that the results are stable rather than the product of a lucky initialization.

The implications extend beyond academic benchmarks. Brands monitor social media to gauge reactions to specific product features; public health agencies track sentiment around vaccines, treatments, and health behaviors; and financial analysts mine posts for signals about consumer confidence. In all of these applications, coarse document-level sentiment is of limited value. What matters is whether users feel positively or negatively about a particular aspect, and whether the attached image supports or contradicts the text. A framework that can perform that fine-grained judgment while tolerating the visual chaos of real-world imagery could make automated opinion mining substantially more reliable.

The work, published open access in Complex & Intelligent Systems, was supported by the National Natural Science Foundation of China and several Yunnan provincial research programs. The authors, Shuwan Yang, Junjun Guo, Zhengtao Yu, and Ran Song, note that their framework’s modularity invites further refinement: the disentanglement and fusion modules could in principle be adapted to other multimodal tasks, from visual question answering to cross-modal retrieval. As multimodal content continues to dominate online communication, the ability to separate what matters from what merely appears in the frame may prove to be one of the most consequential skills an AI system can acquire. ANDF offers a concrete, tested recipe for doing exactly that, and its performance on two demanding benchmarks suggests the recipe works.

Subject of Research: A noise-resilient cross-modal dynamic fusion framework for multi-modal aspect-based sentiment analysis in text-image pairs

Article Title: An aspect-centric, noise-resilient cross-modal dynamic fusion framework for fine-grained sentiment analysis

Article References: Yang, S., Guo, J., Yu, Z., & Song, R. (2026). An aspect-centric, noise-resilient cross-modal dynamic fusion framework for fine-grained sentiment analysis. Complex & Intelligent Systems. https://doi.org/10.1007/s40747-026-02521-y

Image Credits: AI Generated

DOI: 10.1007/s40747-026-02521-y

Keywords: multi-modal aspect-based sentiment analysis, cross-modal fusion, noise robustness, aspect alignment, dynamic gating, visual disentanglement, Transformer architecture, sentiment analysis, Twitter-2015, Twitter-2017, prototype-guided fusion, machine learning