Review summarization performs a significant reduction of lengthy user reviews to short summaries, specifying the major opinions about the product or service. Users face difficulty analyzing large volumes of unstructured online reviews. These challenges are overcome by introducing a comprehensive approach for graph-guided thematic review summarization that enhances the analysis of multi-domain reviews with personalization, interpretability, and efficiency as its core attributes. This research introduces text-derived distilling relational spectral graph-based segmentation (TDRSGS) with the Motorbike Courier Optimization algorithm (MCO), efficiently identifies and groups semantically similar review themes for accurate summarization. The large language model, primarily GPT-4, generates concise theme-level and global summaries while preserving semantic insights and reducing redundancy. The transformer engagement predictor (TEP) tailors summaries to fit the user’s preferences, thereby providing personalised and context-aware recommendations. An exact local interpretable model-agnostic explanation increases the quality of trust and user understanding. The experimental evaluation on a variety of datasets, including Amazon, Flipkart, Yelp, TripAdvisor, IMDB, Goodreads, and RateMDs, shows that the model’s performance is excellent, getting a BERTScore of 97.3, ROUGE-L of 98.4, and accuracy of 98.7
Traditional opinion analysis methods frequently reduce online discussions to simplistic categories, such as favor, against, and none. This oversimplification fails to capture the nuances of opinions, such as the intensity of feelings or the degree to which a stance leans toward favor or against. To address this limitation, a new method called SFLBiL is proposed, which leverages fuzzy logic and BiLSTM to create more nuanced categories for opinions. Specifically, token-based fuzzy matching scores between tweet pairs are transformed into linguistic memberships via fuzzy membership functions and inference rules, which are fused with glove word embeddings and passed through a BiLSTM network that jointly captures ordinal class structure and sequential context. This allows for a more accurate understanding of people’s views, especially in the context of ordered-class classification. The performance of the proposed model is evaluated on two new datasets and the benchmark SemEval-2016 TaskA dataset. The SFLBiL model achieves consistent improvements over state-of-the-art (SOTA) models, with accuracy and F1-score improvements of 4.65
Semantic communication is the key to enable 6G communication, from sending raw data to conveying meaningful information. In contrast, current models fail to process multimodal information in a structured, context-aware manner. To tackle this problem, the proposed Multi-modal Knowledge-Graph Guided Graph Transformer (MKG-GT) combines text and image modalities into a unified heterogeneous graph equipped with a prior knowledge graph. It employs a dual-stage fusion process that captures semantics and structural relationships with Graph Attention Networks (GAN) and Graph Kernel Attention Transformers (GKAT), respectively, and a lightweight semantic compression unit to encode the representation for efficient 6G transmission. Experiments on Flickr30K achieve 0.95 accuracy, 0.94 F1, and 0.97 macro F1, while improving the performance of the current baselines. The model’s semantic pipeline inference time is 1.05ms, supporting low-latency, real-time 6G applications with pre-computed features. Overall, the results show that graph reasoning, knowledge augmentation and compact semantic encoding are a strong method for next-generation multimodal communication.
Audiograms are used to diagnose auditory diseases and are depicted in a standardized manner using mostly identical threshold signs. This study proposes a novel method for detecting and standardizing audiometric thresholds using artificial intelligence, specifically DETR and YOLO models. We introduce a size-fixed rectangular frame and size-fixed bounding-box strategy tailored to the standardized characteristics of audiogram symbols, while comparing performance between YOLOv5 and DETR. The method comprises two main phases: training and detection. In the training phase, a dataset of audiograms is prepared by extracting frame rectangles from original audiogram images and resizing them to 416 × 416 pixels. Air-conduction and bone-conduction threshold signs are annotated using a fixed bounding-box size with the labelImg software. A DETR/YOLO model is then trained on this annotated dataset, leveraging transfer learning from a pre-trained model to enhance efficiency and accuracy. In the detection phase, the trained DETR/YOLO model is applied to original audiogram images. Frame rectangles are extracted and resized to 416 × 416 pixels before being subjected to inference. The error detection rate was as low as 0.3
Neural architecture search (NAS) has recently gained significant attention in the field of AutoML, driven by advances in deep learning techniques. Although many existing NAS methods prioritize classification performance, they require a small architecture to operate effectively in memory-constrained environments. Because model size is a crucial factor in deep learning, NAS methods that balance model size and classification performance are essential. However, owing to their large architectural search spaces, conventional NAS methods tend to converge on large architectures and consume substantial computational resources. To address this issue, this study proposes a novel NAS method that introduces constraints on model size to reduce search time and computational cost. Specifically, our approach involves adjustments to the learning process to improve efficiency, rearranging algorithmic sequences, and integrating a technique called “change alpha’s priority”. Experimental results confirm that the proposed method can identify search for smaller network architectures than the original unconstrained NAS method while maintaining high classification accuracy. The proposed method can search for neural network architectures suitable for various devices and applications.
Document understanding in Indian scripts is challenging due to the non-linear grapheme composition, including conjunct consonants and spatially displaced vowel modifiers, which create complex two-dimensional character structures. To address these challenges, we introduce the IndicScript dataset, a newly curated and annotated dataset comprising Hindi, Bangla, and Odia document images for character segmentation and word-level information extraction. We further propose IndicCharGrid, a grapheme-aware 2D character grid representation that encodes each character at its precise spatial location, preserving both structural dependencies and semantic information. The grid representation is processed using a multi-scale Feature Pyramid Network (FPN) within an encoder–decoder architecture, enabling accurate character segmentation and entity extraction. We evaluate IndicCharGrid against baselines, including LayoutLMv3 (fine-tuned), Donut (OCR-free transformer), and a Transformer-based OCR pipeline. Experimental results shown in Fig. 1 demonstrate that IndicCharGrid consistently outperforms these approaches across all three languages, achieving F1-scores of 99.49
Gradual patterns translate co-variations of the numerical attributes of transactional databases. They play a crucial role in many real-world applications where there is a large amount of digital data to manage. This type of patterns has attracted attention of the data mining community, and several algorithms have been designed to extract frequent gradual patterns from transactional databases. The algorithms for extracting frequent gradual patterns in large databases are CPU and memory intensive, which poses the problem of improving their performance. This paper proposes a novel approach to improve the performance of frequent gradual pattern mining algorithms. It relies on the design of a gradual support inference graph to avoid redundancies in the calculations of gradual supports and to bypass the calculation and storage of the adjacency matrices of certain patterns. The exploitation of said graph in the extraction algorithms leads to a significant improvement in CPU and memory consumption. Experimental results on transactional databases of different natures confirm the effectiveness of the proposed approach.
Thermal imaging offers a promising solution for gender classification, surpassing traditional visible-light approaches by effectively handling fluctuating lighting conditions, shadows and facial occlusions. In this work, Thermal Squeeze and Excitation ResNet (TH-SE-ResNet), an innovative Convolutional Neural Network (CNN) architecture is proposed. TH-SE-ResNet model integrates Squeeze-and-Excitation (SE) blocks to enhance the model’s ability to prioritize relevant features and employs a channel input adapter to address varying input formats of different thermal dataset. This model is used to classify gender from thermal facial images. This study tackles significant obstacles in thermal image analysis, such as inconsistencies in dataset channel configurations, imbalanced class distributions and the challenge of ensuring model robustness across diverse thermal imaging scenarios. The performance of proposed model is compared with several prominent CNN models like AlexNet, VGG, InceptionV3, ResNet50 and EfficientNet across two thermal face datasets: the Tufts University Thermal Face Dataset and the Charlotte-ThermalFace Dataset. Experimental results shows that TH-SE-ResNet demonstrates highly competitive performance achieving a remarkable 97 https://github.com/adityabhattad2021/evaluating-deep-architectures-for-thermal-gender-detection .
In recent years, deep generative models for multimodal data have gained significant attention. Among these, multimodal variational autoencoders (VAEs) have emerged as a promising approach, aiming to capture a shared latent representation by integrating information across different modalities through their inference models. A primary challenge for multimodal VAEs is accurately inferring representations from arbitrary subsets of modalities after learning a multimodal inference model. Naively, this would require training 2^M different inference networks ( M is # of modalities) to handle every possible combination of modalities, which is infeasible for a large number of modalities. Mixture-based models address this challenge by requiring only as many inference models as there are modalities, aggregating unimodal inferences to perform multimodal inference. However, when modalities are missing, these models suffer from information loss, particularly of modality-specific information, leading to deteriorated inference performance. Alternatively, alignment-based multimodal VAEs aim to align unimodal inference models with a multimodal inference model by minimizing the Kullback–Leibler (KL) divergence between them. Yet, the multimodal amortized inference, which is alignment source in these models inherently suffers from amortization gaps, preventing it from perfectly approximating the true inference and compromising the accuracy of unimodal inference. To address both issues, we introduce an iterative amortized inference mechanism within the multimodal VAE framework, termed multimodal iterative amortized inference. By iteratively refining the multimodal inference using all modalities, this method overcomes the information loss due to missing modalities in mixture-based models and minimizes the amortization gap in alignment-based models. Furthermore, by aligning the unimodal inference to approximate this refined multimodal posterior, we obtain unimodal inferences that effectively incorporate multimodal information while requiring only unimodal inputs at inference time. Experimental results on two benchmark datasets demonstrate that the proposed method improves the performance of the inference itself, suggested by higher linear classification accuracy and cosine similarity, and that the learned representations effectively capture the distributions of other modalities, as indicated by lower Fréchet Inception Distance (FID) scores in cross-modal generation. This indicates that the proposed approach significantly enhances the inferred representations from unimodal inputs.
Fairness is a critical concern in machine learning, as biases in prediction outcomes with respect to sensitive attributes such as gender and race have raised ethical and societal issues in real-world applications. To address this, fair representation learning aims to extract essential information from data as representations that are useful but independent of sensitive attributes. However, existing methods often suffer from limited predictive performance when their learned representations are applied to downstream tasks. In this study, we propose predictive FairDisCo (PdFairDisCo), an extension of FairDisCo to learn fair representations with variational autoencoders. PdFairDisCo enhances predictive performance by incorporating contrastive losses into the objective function. In addition, we introduce two oversampling methods for PdFairDisCo to mitigate bias by balancing the proportion of the sensitive attribute in training data. We demonstrate the effectiveness of PdFairDisCo through experiments. The experiments also show that the oversampling methods can further improve the performance of fairness.
In this paper, we propose a hybrid approach that combines Small Language Model (SLM)-based interpretation with machine learning (ML)-based prediction to analyze stress levels and related factors from step-count data. While several datasets exist for predicting mental health conditions from sensor data, most do not explicitly address the underlying factors associated with stress. To explore this issue, we collect step-count data from 30 nurses, together with stress assessments (QIDS: Quick Inventory of Depressive Symptomatology) and stress factor ratings based on six questionnaire items measured on a 4-point Likert scale, collected over 8 days within 4 weeks. We evaluate the proposed approach through two tasks. The first task examines how intermediate textual interpretations relate to stress presence estimation. Under our baseline experimental settings, BERT (Bidirectional Encoder Representations from Transformers) with intermediate stress interpretations achieved the highest accuracy (0.74), compared with BERT using raw step-count representations (step count: 0.63, distance: 0.59) and a prompt-based approach. The second task evaluates the association between intermediate interpretations and stress factor ranking. In this setting, BERT with intermediate stress interpretations achieved a ranking accuracy of 0.60, compared to 0.56 when using step-count sequences without interpretation. Higher correlations were observed for work-related stress factors such as “workplace relationships,” “busy work,” “heavy work responsibilities,” and “lack of time off.” Overall, these results suggest that intermediate textual representations derived from step-count data can be useful for stress analysis under baseline conditions, while avoiding causal claims about stress determinants.
Our study explores the conditions and incentives for government officials in procuring entities (orderers) to initiate bid-rigging and the measures to deter it. First, we used gaming simulations with human agents to investigate the factors that enhance bid-rigging in public procurement. Next, we utilized agent-based modeling and simulation to replicate and verify the gaming simulation. Finally, we used the verified agent-based model and simulations to explore deterrence measures for government official-initiated bid-rigging. We specified two critical factors from the gaming simulation: the type of orderer players and the sense of compliance with social rules. We demonstrated that a strong orderer, unaffected by the surrounding players, could considerably enhance bid-rigging norms. In comparison, a weak orderer affected by the surrounding players was less effective at this enhancement. Next, we replicated the results of the gaming simulations in an agent-based simulation. The results of the agent-based model and simulations revealed that the chance of auction failure was more critical than weakness in the sense of compliance with social rules when a weak orderer intended to establish bid-rigging norms. Finally, we implemented the two existing significant measures into the agent-based models and simulations. In instances where the quality problem of procured goods was not critical, flexible setting of reserve prices was practical; conversely, in cases where bidders encountered challenges in accurately estimating reserve prices or effectively executing contracts, it became imperative to increase both the surcharge and the detection rate for bid-rigging.
Artificial Intelligence, particularly machine learning (ML) algorithms, plays a crucial role in detecting cyberattacks, including anomalies and intrusions. However, machine learning models trained on imbalanced cybersecurity datasets often struggle to accurately detect minority data instances and potential threats, thereby weakening overall system security. Despite extensive research, a persistent challenge is the inadequate explanation for model predictions concerning minority data classes. This study aims to address these limitations by developing a generative AI-based approach to manage minority classes in anomaly detection, incorporating concept drift handling and explainability analysis. We introduce an over-sampling technique, CGGReaT, designed to enhance the presence of minority classes in the anomaly detection domain. Leveraging Large Language Models (LLMs) as a hybrid approach, we use pre-trained transformer-based LLM DistilGPT-2 for generating synthetic tabular data. Extensive experiments on two publicly available benchmark datasets, UNSW NB15 and CIC-IDS2017, underscore the efficacy of our proposed approach. We employed concept drift detection and adaptation techniques to maintain reliable and sustainable ML performance. To enhance interpretability, eXplainable Artificial Intelligence (XAI) methods, including SHAP and LIME, are employed to quantify feature contributions to model outputs. Extensive experiments reveal that testing ML algorithms on datasets balanced with synthetic samples generated by cGGReaT boosts the prediction accuracy on the UNSW NB15 and CIC-IDS2017 datasets, compared to classifiers tested on imbalanced datasets.
Cyberbullying, a pervasive issue within online social networks (OSNs), poses a significant threat to the well-being of children and adolescents. As online social media usage continues to rise, with platforms such as Facebook, Instagram, and Twitter integrating deeply into daily life, the incidence and impact of cyberbullying have become critical areas of concern. This paper explores the evolution of OSNs and their role in facilitating cyberbullying, highlighting the psychological, emotional, and social repercussions on young users. Through a comprehensive review of current literature and statistical data, we examine the various manifestations of cyberbullying, including harassment, racism, and other forms of online abuse. The paper also evaluates existing solutions to detect and mitigate cyberbullying, ranging from general preventative measures to advanced machine learning and artificial intelligence-based techniques. By identifying the limitations of current approaches and suggesting avenues for future research, this study aims to develop more effective strategies to combat cyberbullying and protect future generations from its detrimental effects.
Nowadays, the Internet of Medical Things (IoMT) has received more attention in the healthcare management sector and healthcare services. The integration of healthcare devices and the Internet of Things (IoT) is called IoMT and makes it available for numerous healthcare activities to take place, such as real-time diagnosis, remote patient monitoring, and real-time medicine prescriptions, amongst other things. However, the security and privacy of the data, garnered by the IoMT devices, are primary problems while storing or transmitting it in the cloud. Implementing effective protocols and guaranteeing compatibility among the IoMT systems is significant to rectify this limitation. So, in this research, an innovative IoMT-based disease prediction model is presented. Initially, the requisite medical images are garnered from the standard datasets, and encryption is performed in the collected images by using the Hybrid Advanced Encryption standard with Rossler Hyperchaotic (HAERH). Further, the key needed for encryption is created using the Enhanced Pine cone Optimization (EPO). Subsequently, the encrypted images are decrypted to perform the disease prediction task using the Parameter Optimized ShuffleNet (PO-SNet) technique, where several parameters are tuned by the EPO to enhance the prediction performance. Finally, the designed PO-SNet model gives the disease-predicted outcome. Finally, the experiments are done for the developed work over traditional methods. The results show that the designed EPO-PO-SNet’s accuracy is 94.61
Unstructured scene understanding in real-world environments remains difficult, such as how to understand unstructured paths in field environments using resource-constrained systems with extremely limited computational, memory, and energy resources. Traditional methods are too focused on the specific goal of detection or segmentation, and seriously neglect the sustainability of energy consumption, cost and complexity. In this study, we present a sustainable method to understand unstructured paths and reconstruct them in 3D space and plan walkable routes through a low-cost monocular camera. Edges lines are detected. Contour candidate curves are extracted by the relative integrity of lines. Based on geometric orientation continuity and coverage, candidates are refined to get approximate contours, which are regarded as walkable regions. By analyzing the relative geometric constraints between different walkable contours, unstructured paths can be understood and reconstructed in 3D space and walkable routes can be planned. This method requires no prior training, and no calibration of the internal parameters of monocular camera, which has the sustainability of low power consumption and low cost. We provide a comprehensive evaluation of the method in terms of multiple dimensions such as energy consumption, cost and interaction ratio. The results show that the method not only can understand unstructured paths, but also its sustainability of low power, low cost, and lightweight are more suitable for a resource-constrained system in field environments.
This study introduces a novel adaptive multi-agent system (MAS) for dynamic missing value imputation in time series data, essential for accurate forecasting. The MAS employs a hybrid approach, intelligently selecting and combining techniques like regression, interpolation, and rule-based methods based on missing data patterns. Its extensible architecture allows seamless integration of new imputation methods. The system adapts dynamically, using regression for large gaps and statistical methods for smaller ones, with concurrent, restartable agents enhancing efficiency. A knowledge-capture mechanism enables continuous learning by storing imputed data and context for future use. The MAS was extensively evaluated on a real-world human mobility dataset with over 100 million records from Hiroshima, Japan, serving as a preprocessing step in an AutoML platform for time series forecasting. Tests on vehicle movement data and Airline passenger data demonstrated its effectiveness, with low evaluation metrics confirming its accuracy and efficiency. The proposed MAS offers a robust, scalable solution for missing value imputation in diverse time series applications.
In the landscape of viral genomic studies and knowledge representation solutions, there exists an emerging need for an ontology designed specifically to represent the genomic sequence data of the COVID-19 pandemic. Existing ontologies often lack the specificity and comprehensiveness required to adequately capture the various facets of COVID-19 genomics. To address this gap, we present the development of the COVID-19 Virus Genomics Ontology (VGO). VGO is designed to facilitate the utilization and publication of COVID-19 genomic sequence data, offering a comprehensive solution for researchers and healthcare professionals. By integrating data from the Global Initiative on Sharing All Influenza Data (GISAID), VGO enables efficient querying and visualization of genomic information, enhancing accessibility and usability. To construct VGO, we employed a combination of two established methodologies, the Yet Another Methodology for Ontology Development (YAMO), and Networked Ontology (NeOn) methodology. Utilizing the Web Ontology Language (OWL), VGO provides a robust representation of COVID-19 genomic sequence data, filling a critical gap in existing ontologies. To ensure reliability and usability, VGO underwent a comprehensive evaluation process, benchmarking it against ten prominent COVID-19-related biomedical ontologies using a multidimensional framework that includes structural, functional, usability, and quantitative dimensions. Structural quality was assessed through an extensive pitfall analysis using the OOPS! framework, identifying critical modeling issues across peer ontologies, with VGO exhibiting no significant pitfalls–indicating high modeling rigor and syntactic correctness. Logical consistency and structural soundness were validated through OntoDebug and the Pellet Reasoner, confirming the absence of inconsistencies and reinforcing internal coherence. Quantitative evaluation using OntoMetrics highlighted VGO’s balanced design, demonstrating moderate attribute and relation richness, with notable instance-level coverage–a feature absent in most comparator ontologies. Despite a smaller axiom base, VGO maintains robust schema expressivity and a clear TBox-ABox integration, enabling both knowledge representation and data-level reasoning. Finally, a SPARQL-based competency question assessment confirmed the ontology’s functional completeness and real-world applicability, mapping successfully to its structure.
The timely processing and handling of invoicing papers linked to COVID-19 has become essential in light of the epidemic. The differences between the text and their complex meanings remain the major issue of the conventional approaches to managing the documents. To overcome these issues, this research proposed an Attacker Grouping Optimized Selective Kernel-based Deep Convolutional Neural Network (AG optimized SK-DCNN) model and optimize the invoice processing workflow. The selective kernel network is utilized to obtain accurate data from invoice documents that are unstructured, whereas the Google Optical Character Recognition (Google-OCR) makes it possible to extract text from images that may be the photographs or the scanned images, with advanced technology created by Google. Text annotation with Named Entity Recognition (NER) necessitates detection along with categorizing structures of the available text that aids in achieving the efficient extraction of the data. The feature extraction and classification duties are ensured by the AG-optimized SK-DCNN, where the AGO utilizes the aggregating and foraging characteristics to enhance the system performance and reduce the error rate occurrence. The AG-optimized SK-DCNN approach achieves 94.7
Given the remarkable achievements in image generation using diffusion models, the research community has shown increasing interest in extending these models to video generation. Recent diffusion models for video generation have predominantly utilized attention layers to extract temporal features. However, attention layers are limited by their computational cost, which increases quadratically with sequence length. This limitation poses significant challenges when generating longer video sequences using diffusion models. To overcome these challenges, we propose to leverage state-space models (SSMs) as temporal feature extractors. SSMs (e.g., Mamba) have recently garnered attention as promising alternatives owing to their linear-time memory and time consumption relative to the sequence length. Employing SSMs to capture temporal dependencies in video generation enables significantly higher generative performance at the same computational cost (e.g., memory usage, inference time) compared to attention-based methods, particularly for long-term sequences. For various model sizes, we comprehensively evaluated multiple long-term video datasets: MineRL Navigate, GQN-Mazes, and CARLA-Town01. For 256-frame video sequences, SSM-based models incur lower computational cost to achieve the same Fréchet Video Distance as attention-based models. Furthermore, the ablation study shows that when using SSMs for temporal modeling, incorporating bidirectionality and selective scans enhances video generation performance. Our code is available at https://anonymous.4open.science/r/SSM-Meets-Video-Diffusion-Models-067D/README.md .
