In the automotive industry, quality assurance is crucial for ensuring safety and performance. An integral part of quality assurance is Noise, Vibration and Harshness (NVH) testing in transmission production. NVH testing utilizes sensors such as accelerometers, microphones, and encoders to capture acoustic and dynamic characteristics under simulated road conditions. Traditionally, these characteristics are evaluated against manually defined thresholds set by NVH engineers—a process that is labor intensive, costly, and complex. Although advances in machine learning (ML) offer the potential to improve efficiency in NVH testing, challenges such as severe class imbalance, outliers, and limited access to production data hinder its adoption. To address these challenges, we propose Modified Z-Score Cyclic Exponential Averaging of Minority Data ( mZ - cEAMD ), a noise-robust over-sampling method that integrates exponential averaging with outlier removal to generate balanced and representative samples, while preserving the fundamental characteristics of the original data distribution. This is complemented by a feature extraction approach that integrates spectral order tracking, which captures gear meshing frequencies relative to input rotation speed, with the Short-Time Fourier Transform (STFT), which represents time-frequency characteristics. Together, these features contain critical discriminative information for transmission classification and are used to assess the effectiveness of the proposed over-sampling method. Comparative experiments against established over-sampling methods demonstrate that the proposed mZ - cEAMD approach outperforms alternative techniques in addressing class imbalance in datasets containing outliers. This study advances ML-based NVH testing, thereby fostering the digital transformation of automotive quality assurance.
Aspect-based sentiment analysis (ABSA) is a specialized form of sentiment analysis that provides more detailed and granular insights. Due to the increased availability of customer feedback data on various products, services, and other aspects, aspect-based sentiment analysis has gained significant attention over the past few years. The advancement in this field motivated different researchers to develop different methods, datasets, and tasks to better analyze ABSA. To conduct this comprehensive survey, more than 140 research articles published from 2014 to 2024 were collected from reputable academic journals’ databases. This survey is focused on an in-depth explanation of different tasks, trends, methods, and datasets involved in the field of ABSA. Besides, this survey outlines the evolution of ABSA and its applications. At the end, this paper also highlights findings, associated challenges, and future scope related to this field.
The integration of deep learning models into healthcare, finance, and government sectors has significantly advanced the capabilities of artificial intelligence (AI). However, these deep learning models require substantial computational power, leading many organizations to rely on cloud-based services for model training. This shift has raised critical privacy concerns because sensitive data used in training may be exposed to privacy attacks. This survey systematically examines the state of privacy attacks against deep learning and the effectiveness of current privacy-preserving methodologies. Using a PRISMA-style review, we analyze papers from the recent five-year literature and present (1) a unified taxonomy of attack families (membership inference, model inversion, model extraction, gradient leakage, property/attribute inference, poisoning/backdoors, side-channels, and LLM-specific leakage); (2) a taxonomy of defenses spanning differential privacy, federated learning (and robust aggregation), cryptographic protocols (HE/SMPC), anonymization sanitization, PATE/shuffle models, and hybrid stacks; (3) recommended evaluation metrics for each attack family; and (4) a reproducibility checklist and an attack × defense protection matrix that synthesizes the literature’s qualitative findings. We complement these contributions with three case studies analysis of current landscape of privacy-preserving defenses against privacy attacks in centralized image classification, federated learning, and LLMs, which highlight practical trade-offs between privacy, utility, and system cost. Our analysis finds that noise-based methods such as DP remain the practical baseline but offer only partial protection; cryptographic approaches provide stronger theoretical guarantees at substantially higher cost; and LLM/multimodal leakage remains an urgent, under-benchmarked gap.
One of the primary challenges in training deep neural networks is the requirement for a robust and diverse data source. To address this limitation, data augmentation techniques have emerged as a promising solution, enabling the expansion of training datasets without requiring additional data collection. This study investigates the effectiveness of code refactoring-based augmentation and the role of augmented data volume in improving code smell classification in deep learning systems. We conducted a comparative analysis of rule and large language model-based augmentation with a loss penalty, an algorithmic data balancing approach for neural networks. Additionally, a novel weighting strategy was developed to optimize the evaluation of augmentation volume distribution, reducing the computational overhead while maintaining analytical precision. We used pre-trained BERT, CodeBERT, and GraphCodeBERT models to generate code embeddings and evaluated their performance for model training. Our findings demonstrate that code refactoring-based augmentation using GraphCodeBERT and large language model enhances model performance, particularly in addressing class imbalances. The impact of data volume varies depending on the size of the classes, with the greatest improvement observed during the initial augmentation increments applied to underrepresented classes. Conversely, the influence of the loss penalty, BERT, and CodeBERT-based augmentation on overall performance was found to be negligible or, in some cases, detrimental. These results emphasize the potential of code refactoring-based augmentation to drive the development of more efficient data augmentation strategies, ultimately enabling better performance in code smell classification tasks and other deep learning applications.
Smart contracts represent a cornerstone of blockchain technology, enabling key applications in cryptocurrencies and decentralized finance (DeFi). Nevertheless, the inherent complexity in developing smart contracts frequently introduces severe security vulnerabilities. Compounding this issue, smart contracts become immutable upon deployment, exposing users and participants to substantial financial losses and risks. Consequently, vulnerability detection and security auditing emerge as imperative tasks. Existing deep learning-based vulnerability detection approaches are constrained to fixed-length input sequences, offering no effective mechanism for handling the variable-length opcode sequences derived from transaction executions. This limitation often results in data truncation, which compromises detection accuracy by discarding critical information. To overcome this challenge, a novel neural network model is proposed that accommodates variable-length input sequences. Opcode sequences are acquired through transaction replay as the dataset. Variable-length opcode sequences are processed using the ChordMixer architecture to preserve complete sequence information and extract global features. Additionally, a retention mechanism is integrated into the network to identify salient local features. This combined emphasis on global and local sequence information facilitates more precise security analysis and vulnerability prediction in smart contracts. Extensive experimental validation confirms the model’s efficacy, attaining 93.5
Analysis of high-dimensional Internet of Things (IoT) network data for intrusion detection is an important task, as the data generated by heterogeneous IoT devices are unbalanced, diverse, and context-dependent in nature. Feature selection (FS) plays a major role in selecting the most informative, highly relevant, and non-redundant features from high-dimensional IoT intrusion datasets. In this paper, we discuss a novel ensemble FS method, EF ^2 RB, that incorporates the innovative concept of Ranking with Boosting. The method is evaluated on high-dimensional IoT intrusion datasets using baseline machine learning (ML) models and found that the proposed EF ^2 RB achieves very high detection accuracy on the selected feature subset. Moreover, we tested the proposed EF ^2 RB method in an IoT intrusion detection system (IDS) and observed that the IDS achieved consistently high performance (95–100 https://github.com/2ez4n0dyX/EF2RB.git .
Out-of-distribution (OOD) detection flags test inputs that depart from the data used to train a model. For structured tabular problems with regression or survival outcomes, existing methods remain limited because many OOD detectors are designed for classification or unstructured data. We introduce a tree based method that uses the rule structure of a supervised forest to determine the predictive subspace, the model-aware reference neighborhood, and the final OOD score in a single construction. The score compares each test input with forest selected reference cases only on prediction relevant variables, reducing signal dilution from nuisance coordinates. Across synthetic and real data benchmarks, the method performs especially well for subtle targeted feature shifts and changes in dependence. An esophageal cancer survival study further shows how OOD scores can reveal lymphadenectomy related shifts relevant to surgical guidelines.
Fish species classification plays a vital role in biodiversity assessment, sustainable fisheries, marine ecosystem monitoring, and environmental engineering applications. Despite major advances in deep learning, identifying visually similar fish species remains challenging due to underwater factors such as lighting variation, occlusion, cluttered backgrounds, and subtle morphological differences. This paper introduces a hybrid deep learning architecture named Graph Convolutional Vision Transformer (GCvT), which integrates three complementary components: CNNs for local texture extraction, Vision Transformers (ViTs) for capturing global dependencies, and graph convolutional networks (GCNs) for modeling relational structures among image regions. In GCvT, each image is divided into fixed-size patches. CNN layers extract spatial and texture features, which are refined by ViT blocks through self-attention mechanisms to model global semantics. A dynamic graph is then constructed where patches act as nodes, and edges are determined by feature similarity and spatial proximity. The GCN propagates contextual information across nodes, enhancing patch-level representations through neighborhood aggregation. This multi-level fusion enables GCvT to jointly exploit local, global, and structural cues for robust feature learning. A global aggregation layer followed by a fully connected classifier performs the final prediction. Experiments on multiple benchmark underwater fish datasets demonstrate that GCvT achieves accuracy above 99
The federated learning (FL) on edge networks is rapidly spreading because it enables training the models without data privacy violations, as it trains the model on many computers at a time. Nonetheless, the training of the several models in these settings commonly referred to as multi-model federated learning (MMFL) comes with tremendous complexities especially in resource allocation, energy use, and communication effectiveness. In the current paper, a new hybrid optimization system is presented that aims at addressing the issues of MMFL in edge computing. The suggested framework combines mixed-integer programming (MIP) with decomposition solutions to optimize the choice of participants and the learning schedules distribution, and, in this case, the considerations of energy, latency, and fairness are lowering in a coordinated approach. It aims at reducing the total operational expenses, which are the communication overheads, the computational loads and the energy spending and meeting the draining fairness and model precision requirements. Through the hybrid decomposition method, the optimization problem is decomposed into small sub-problems, which can be solved in an iterative manner, not only accelerates calculations, but also makes federated learning scaling possible in large networks. The proposed framework is demonstrated, via a large-scale simulation, to outperform classical algorithms, e.g., greedy scheduling and random sample selection, by consuming much less energy (30
Multi-view clustering aims to integrate complementary information from diverse data sources to achieve more comprehensive and consensus representations. While existing multi-view subspace clustering methods have made considerable advances in learning and fusing subspaces, most approaches directly apply clustering on the fused representation, leaving the underlying structural properties of the subspace underexplored. To this end, we propose an adaptive structure-guided framework for multi-view subspace clustering (ASG-MVSC). The key idea is to construct density-aware structure information in the learned subspace to capture global dependencies for clustering optimization. Specifically, ASG-MVSC incorporates: (1) a density-aware structure learning strategy that redefines K-nearest neighbors and density functions in the subspace to capture robust global relations, and (2) an adaptive fusion mechanism that evaluates the consistency between view-specific structure information and the global consensus to dynamically adjust fusion weights. Extensive experiments on eight benchmark datasets demonstrate that ASG-MVSC outperforms state-of-the-art methods in clustering accuracy and robustness. In addition, ablation experiments, statistical significance tests, convergence analyses, and parameter analyses validate the effectiveness and stability of the proposed framework.
Financial constraints, which limit firms’ access to external resources and impede growth, remain a critical challenge in decision-making. These constraints are often exacerbated by information asymmetry among stakeholders. To address this issue, this study leverages social media news for its strong capacity to disseminate information on inter-firm relationships (e.g., cooperation, competition). Text mining (TM) is used to extract firms’ relationship networks, while social network analysis (SNA) derives graph-theoretic attributes (e.g., betweenness centrality). Drawing on ensemble learning, the study shifts from a model-centric to a data-centric perspective by refining and augmenting training data to enhance AI forecasting performance. Results show that firms with advantageous network positions can alleviate financial constraints, and forecasting accuracy improves through data refinement. These findings support information-access theory and ensemble learning, offering guidance for managers and investors in resource allocation and strategy adjustment.
Big data classification categorizes unstructured or structured data depending on file contents, types, and other metadata. The issues faced by big data classification are poor generalization, inaccurate predictions, security vulnerabilities, and a lack of big data scientists and data specialists. Hence, an effective fractional hunter osprey optimization-based parallel convolutional neural network (FHOO_PCNN) is proposed for big data classification. The process begins with collecting the input big data and partitioning it using deep embedded clustering (DEC) for efficient structuring. Then, big data classification is implemented using the MapReduce framework, involving mapper and reducer stages. Data normalization using logarithmic scaling is executed in the mapper stage to standardize the dataset. Then, normalized data is passed to feature weighting utilizing the proposed hunter osprey optimization (HOO), which incorporates the honey badger algorithm (HBA) and osprey optimization algorithm (OOA). The reducer phase merges the weighted features, after which classification is achieved using a PCNN optimized through FHOO, formulated by integrating fractional calculus and HOO. Moreover, FHOO_PCNN yields 91.765
Knowledge graph completion has attracted significant attention due to predicting missing links in knowledge graphs. Recent advancements in embedding-based methods have demonstrated excellent performance in 1-to-1 relation prediction. However, these methods encounter difficulties in modeling complex relations, including 1-to-N, N-to-1, and N-to-N, resulting in less effective of complex relation prediction. To address the limitations of semantic representation caused by relational complexity in embedding-based methods, we propose the joint entity and relation embedding network (JERENet) for multi-relational knowledge graph completion. Specifically, JERENet can effectively capture complex relations by leveraging multi-relational features among entities and neighborhood structure information. JERENet also introduces adversarial training through the construction of adversarial samples. Our proposed method can effectively capture complex relations by leveraging multi-relational features among entities and neighborhood structure information. Extensive experiments on public datasets WN18RR and FB15k-237 reveal that JERENet shows superior or at least comparable performance to the state-of-the-art baseline methods in link prediction.
The Granular-ball Rough Set (GBRS) model is a promising approach that combines classical rough sets with neighborhood rough sets. However, GBRS suffers from the issue of high randomness in its granular-ball generation process, leading to instability and inconsistent partitioning results. Additionally, it does not take into account the correlation between conditional attributes and the decision attribute, which can affect the accuracy and effectiveness of attribute reduction. To address these issues, we propose the Integrated Weighted Granular-ball Rough Set (IWGBRS) model, which integrates multiple granular-ball splitting strategies to achieve more stable and reliable partitions. The IWGBRS model introduces a novel granular-ball generation algorithm that employs a combination of various clustering methods for initial data partitioning. The IWGBRS model reduces the randomness in the granular-ball generation process and ensures more consistent and stable partitioning results. Furthermore, to address the issue of not considering the correlation between conditional attributes and the decision attribute in GBRS, the IWGBRS model incorporates a weighting mechanism for conditional attributes. This mechanism enables the model to more accurately capture the contribution of each attribute to the decision attribute. A greedy searching algorithm is then used to select a subset of conditional attributes that exhibit both strong correlation and high dependency with the decision attribute, effectively reducing dimensionality and enhancing the overall performance of the attribute reduction process. Experimental results on sixteen datasets from the UCI machine learning repository demonstrate that the IWGBRS-based attribute reduction algorithm outperforms other popular methods. It offers improved classification accuracy and more effective attribute reduction with reduced randomness.
Imbalanced regression refers to regression problems where the target variable has a non-uniform distribution. Classical approaches to address this issue include weighted regression and resampling techniques. While weighted regression modifies the learning algorithm, resampling methods adjust the data distribution, making them more flexible and commonly used. However, resampling methods are sensitive to noise, particularly in overlapping regions between rare and frequent instances, which can lead to the generation of misleading synthetic samples. In this context, noise refers to rare instances that resemble non-rare ones in feature space or non-rare instances that exhibit rare-like features, making resampling less effective. In this study, we propose RA-HyRL designed to address the noise problem in imbalanced regression. RA-HyRL is a hybrid method that performs both oversampling and undersampling. It leverages boosting weights to identify noisy and informative samples, and selects instances to resample based on their relevance. The number of neighbors used for synthetic sample generation is also automatically determined for each instance. We evaluate the proposed method against six existing resampling techniques on 30 datasets using 10 regression models. Statistical tests demonstrate that RA-HyRL outperforms the alternatives at rareness thresholds of 0.5 and 0.8, according to the SERA metric. The source code is available at: https://github.com/fatihsaglam/ImbRegSamp.
Social media has become a primary source of information, but its rapid growth has facilitated the spread of misinformation. Existing methods often rely on static graph neural networks (GNNs) to analyze news propagation patterns, overlooking temporal dynamics and semantic nuances. To address this, we propose a Temporal Graph Convolution Network with a Hierarchical Co-Attention (TGCN-HCA) model for effective fake news detection. TGCN-HCA leverages a dynamic graph representation to capture the evolving nature of news propagation. By incorporating temporal information, we can accurately model the sequential spread of news and identify patterns associated with fake news. To enhance semantic understanding, we employ GPT-2 to extract deep contextual features from news content. These features, combined with structural information from the graph, provide a rich representation of news propagation. A hierarchical co-attention mechanism further refines feature representation by jointly considering multiple perspectives: the importance of nodes within the graph and the relevance of semantic features to the task of fake news detection. By integrating these components, TGCN-HCA effectively captures the complex interplay between temporal dynamics, structural information, and semantic content. Extensive experiments on Twitter15, Twitter16, and Weibo datasets demonstrate the superior performance of TGCN-HCA over state-of-the-art methods. Our model achieves significant accuracy gains of 5.37
Transformer architectures dominate contemporary machine learning, yet face critical limitations in security applications: vulnerability to adversarial attacks, lack of calibrated uncertainty estimates, and difficulty distinguishing confident predictions from uncertain cases requiring human review. We introduce stochastic probably approximately correct (PAC) Bayesian transformers that convert deterministic attention into probabilistic variants via variational inference, unifying uncertainty quantification, and adversarial robustness within a single framework. Our approach replaces fixed attention weights with learned variational distributions and propagates uncertainty through Monte Carlo (MC) sampling, creating moving targets that degrade adversarial effectiveness. We derive joint PAC-Bayesian bounds showing that parameter stochasticity improves both calibration and robustness, with complexity scaling as O( √(KL( ρ ||π)/n)) , where KL denotes the Kullback–Leibler divergence between the learned posterior ρ and prior π , and n is the sample size. Across network intrusion detection, toxic content detection, and fake news identification, we achieve 96.8 ± 0.8% accuracy with the expected calibration error (ECE) of 0.043 ± 0.006 , and maintain 88.3 ± 1.5% robust accuracy under multiple attack strategies. Active learning guided by uncertainty reduces labeling requirements by 68% , reaching 95% of full-data performance with only 35% of labels.
Existing emerging topics (ETs) prediction studies primarily focus on characterizing topic-related features and constructing emerging attribute indices. However, the former often overlook the differential influences of academic entities on topics, and the latter rarely incorporate the innovation of knowledge. Therefore, this study proposes an ETs prediction framework that integrates structural and content semantics in dynamic heterogeneous networks to address these limitations. First, we construct dynamic heterogeneous networks comprising four entity types (i.e., papers, authors, topics, and venues) and their relations in real-world scientific systems. Second, a topic-aware mechanism is proposed to enhance topic-related feature characterization. It extracts co-evolutionary features while accounting for the differential influences of academic entities on the topic. Third, to better quantify topics’ emerging degrees, we propose a novel emerging index characterizing two aspects: the popularity score of topic and relative topic novelty. This index captures the innovation of knowledge by projecting content semantics from scientific texts. Last, we predict the emerging degree for each topic using its intrinsic features and co-evolutionary features. Empirical studies on the brain neoplasms and cardiovascular abnormalities datasets confirm the effectiveness of the framework. Overall, this study hopes to enrich the methodological foundations and practical guidance of ETs prediction.
Rating prediction is an important task in review mining. Aspect category sentiment analysis (ACSA) and overall rating prediction (ORP) are closely related tasks, yet existing joint models are typically limited to text-only inputs, while existing multimodal review models usually address only a single task. To bridge this gap, we propose MMRP, an end-to-end multimodal multitask framework for jointly modeling ACSA and ORP from image-text reviews. The model integrates textual and visual information through a cross-modal Transformer and further introduces modality-aware learnable positional encoding and a gated multi-head cross-modal attention mechanism to better handle the structure of multimodal reviews, where text forms token sequences and images appear as variable-sized sets. We evaluate MMRP on the ZOL mobile review dataset and an additional cross-domain multimodal benchmark, ViMACSA. Extensive experiments, including comparative evaluation, ablation analysis, sensitivity analysis, and case studies, show that MMRP consistently achieves strong performance and exhibits good robustness, cross-domain applicability, and practical deployment potential. These results demonstrate the effectiveness of jointly leveraging multimodal information for aspect-level and overall rating prediction in user reviews.
Aspect-based sentiment analysis (ABSA) facilitates fine-grained opinion analysis by integrating sentiment polarity with precise aspects in textual reviews. Despite tremendous improvements, contemporary ABSA methodologies continue to struggle with issues such as precise multi-word aspect extraction, sensitivity to noisy data, inadequate domain generalization and fragmented approach to aspect extraction and sentiment classification. In order to tackle the above-mentioned constraints, this investigation proposes a novel multilevel fast point graph transformer network (MFP-GTN) that incorporates aspect extraction and sentiment classification in to an individual framework. The design combines a joint neural conditional random field (CRF) for accurate multi-word aspect recognition with a multilevel graph-based rapid point transformer which records both long-range contextual dependencies and local sentiment cues. A dual-phase metaheuristic combining hybrid giant trevally white shark optimization (HGTWSO) and enhanced salp swarm optimization (ESSO) is implemented for optimum feature selection and model tuning, in order to lower redundancy, enhance convergence, and improve generalization. Detailed trails across multiple benchmark datasets, including Flipkart cell phone reviews, restaurant reviews, financial sentiment datasets and consumer car ratings, demonstrate the effectiveness of MFP-GTN model. The model attains accuracies of 99.36
