Bankruptcy forecasting is an important function for banks and other financial institutions for determining default risk and making informed choices. However, conventional bankruptcy forecasting models are challenged by the intricate, non-linear relations among different financial metrics. These models are often unable to identify the best features and obtain high predictive accuracy, which results in poor risk estimation. For solving these issues, sensitivity analysis is used within this paper to find out which features are most contributing to bankruptcy outcomes. The predictive performance is boosted by using sophisticated optimization methods, including Ant Colony Optimization (ACO) and Cuckoo Search Optimization (CSO), for hyperparameter and feature selection. These tools facilitate improved determination of features and the accuracy of a model. The findings indicate that Competitiveness is indeed the most significant characteristic that is ranked highest by Recursive Feature Elimination (RFE) and sensitivity analysis. Moreover, ensemble-based models, including XGBC, optimized by ACO (XGAC), achieve better performance compared to standard models, including Logistic Regression Classification (LRC). The performance is tested against two scenarios, namely, bankrupt (B) and non-bankrupt (NB), with better performance by XGAC. The conclusions are of direct applicability for real-world bankruptcy forecasting, which will help banks enhance risk-handling capabilities. Further research ought to focus on combining time-series data for further model enhancement.
The number of edge devices in large-scale edge systems is rapidly increasing, while these devices operate under constraints in processing capacity, memory, and network bandwidth. Efficient resource utilization and data management are therefore important for edge query processing. Join operations are among the most computationally and resource-intensive database operations, particularly when large volumes of data must be scanned. Column Imprint-Hash Join (CI-HJ), a state-of-the-art approach for edge query processing, addresses this challenge through data-aware equi-height binning to reduce unnecessary cacheline accesses during hash joins. However, its binning strategy does not consider query workload characteristics and may therefore provide limited benefit when frequently accessed value ranges are not aligned with the data distribution. This paper presents Workload-Aware Column Imprint-Hash Join (WACI-HJ), which incorporates query workload characteristics into Column Imprint construction to improve workload-aware data skipping. WACI-HJ further employs ARIMA-based workload prediction to anticipate query access patterns and organize the data structure before subsequent queries arrive. The framework consists of two phases: (1) WACI-HJ Generation, comprising Pre-processing, Prediction, and Blocking and Hashing modules for workload-aware bin construction, and (2) Query Processing and Resource Utilization, which processes incoming queries using the currently available WACI-HJ structure and monitors CPU, RAM, and I/O utilization. Evaluation using benchmark and real-world smart-transportation datasets demonstrates a 54% reduction in Percentage of Cachelines Read (PCR) and a 10% improvement in Query Execution Time (QET) compared with CI-HJ. The proposed approach is also evaluated under different data scales and skewness levels to examine its scalability and robustness. Resource-utilization analysis shows reductions of 1%, 38%, and 49% in CPU, RAM, and I/O utilization, respectively. These results indicate that workload-aware Column Imprint organization can reduce unnecessary data access and improve low-latency query processing in resource-constrained edge environments. The approach is particularly applicable to smart-transportation workloads, including traffic analysis, congestion management, and route-related query processing, and can be extended to other edge-computing applications requiring efficient query processing.
Multivariate time series prediction is a key technology in finance, transportation and energy. However, the multi-variable coupling, multi-scale features and implicit high-order relationships make accurate prediction challenging. Existing deep learning models, especially graph neural networks, are limited to binary edge connections and cannot adequately model the complex high-order interactions across multiple scales. To address this, we propose a multi-scale dual hypergraph convolutional attention network (MSADHGCANet). The model employs a dual hypergraph generation module to construct dynamic and static hypergraphs, and a dual hypergraph learning module to capture intra-scale and cross-scale high-order interactions. Experiments on six public datasets demonstrate that our model consistently outperforms state-of-the-art baselines, and ablation studies confirm the effectiveness of each core component.
Compared with homogeneous networks, heterogeneous networks are more suitable for depicting most complex systems in reality, such as social networks, transport networks, and biological networks. Heterogeneous networks typically contain rich semantic information and comprehensive structural features, which in turn make it particularly challenging to represent them effectively. The primary concern in heterogeneous network representation is to obtain low-dimensional node vectors while preserving the original nontrivial information. So far, although tremendous efforts have been paid on this issue, challenges remain. In this paper, we propose a Multi-view Heterogeneous Network Representation (MHNR) approach to capture valuable information from node-view, semantic-view and higher order-view. Specifically, we employ an attention-based cross-modal fusion strategy to effectively learn the multi-modal attributes of nodes. Subsequently, the deep semantic feature extraction techniques are integrated with a dual-layer attention-based semantic aggregation module to achieve fine-grained semantic fusion. Finally, a higher-order relation-aware structural subgraph construction method is designed to capture long-range dependencies and complex interaction patterns in heterogeneous networks. Experimental results demonstrate that MHNR model achieves the average improvements of 4.55% in Macro-F1/Micro-F1 for classification and 4.95% in NMI/ARI for clustering, outperforming the selected baseline methods.
Software defect prediction (SDP) is an effective means to assist software testing for resource allocation. Especially, it is recommended that association classification rule for SDP considers both comprehensibility and performance. However, due to class imbalance of software defect data, there is a preference to discover more non-defective class rules than defective ones in rule generation. Classical associative classification rule algorithms usually are assumed that metrics are of equal importance, which may lead to the elimination of useful rules. On the other hand, they adopt discretization methods to transform the continuous metric data into several disjoint hard intervals, ignoring the variation of interval boundary values leading to boundary sharpening. This promotes the emergence of fuzzy weighted associative classification. However, customized membership functions based on domain knowledge often fail to achieve self-adjustment of different data intervals. Therefore, we propose an innovative SDP model based on fuzzy clustering weighted class association rule mining (FCWCAR), which is considered to address the class imbalance, feature importance, and boundary sharpening problems simultaneously. The fuzzy clustering approach based on objective function can be employed to achieve smooth transition between fuzzy intervals, the multi-fuzzy weighted supports-based framework can be implemented to solve the class imbalance, and the Pearson correlation-based approach can be utilized to assign different weights to metrics. In addition, the rule ranking, pruning and prediction stages are optimized according to fuzzy weighted support. We conduct experiments on 27 open-source datasets and results show the validity and feasibility of FCWCAR in terms of Balance, G−mean, MCC, and F−measure.
This paper introduces a classification-based recommendation system that outputs purchase probabilities rather than just recommending items based on traditional methods. Unlike traditional methods that focus on rating prediction or implicit feedback, this system provides a probabilistic ranking that is directly used for personalized marketing strategies, such as targeted price discounts. By leveraging large-scale datasets (1.2 billion records) and six machine learning (ML) classifiers, the proposed system is designed to be scalable to e-commerce environments. Additionally, our research compares the performance of the classification-based system against classical collaborative filtering techniques using the Surprise package. Furthermore, the use of different data formats including JSON-based order data highlights the system’s ability to handle heterogeneous and unstructured data sources. The recommender system is updated as the users, items and orders update dynamically. The high volume, velocity and variety of data position our probabilistic approach within the framework of big data. Two strategies are foreseen to handle big data: data sampling or applying XGBoost and LightGBM with parallelism, GPU acceleration and out-of-core training. The proposal of a dynamic pricing adjustment based on predicted purchase probabilities (e.g., discounts for items with probability between 0.35 and 0.49) bridges the gap between predictive modeling and actionable business decisions.
RNN-based and Transformer-based methods respectively face challenges in long-term dependency and computational efficiency. The complex data structure, the information decay inherent in recurrent units, the high computational complexity of the attention mechanism, and the error accumulation caused by point-wise iterations are the reasons behind the limitations of these methods in time series forecasting. To address these problems, a hybrid forecasting architecture is proposed, which integrates adaptive time series decomposition, gate-guided attention network, and segmentation prediction mechanism, namely SegGANet. Adaptive time series decomposition can mine the trend, seasonality, and residual components in time series, reducing mutual interference among components. Gate-guided attention network incorporates gated recurrent unit, self-attention, and multi-head attention mechanism, which can effectively capture long-term dependencies and enhance overall forecasting accuracy. Segmentation prediction mechanism replaces point-wise iterations with segment-wise iterations, significantly improving forecasting efficiency. Extensive experiments conducted on multiple public benchmark datasets demonstrate that SegGANet significantly outperforms mainstream RNN-based and Transformer-based methods, achieving higher accuracy and efficiency.
We introduce Multi-Level Differentiated Optimization for Differential Privacy (ML-DO-DP), a practical mechanism designed to improve the privacy–utility trade-off under a fixed global privacy budget ε. ML-DO-DP performs feature-wise clipped sensitivity estimation, distributes the global budget across features, applies exponential time-decay weights to discount stale records, and optionally applies a Kalman-based post-processing smoother that preserves the DP guarantee. We benchmark ML-DO-DP against classical mechanisms (Laplace, Gaussian) and modern baselines (Personalized DP, DP-SGD) on three public binary classification datasets (Breast Cancer, Pima Diabetes, Heart Disease). At ε=1.0, ML-DO-DP attains AUCs of 0.92, 0.78, and 0.86 respectively, an average improvement of ≈0.03 over Laplace, and reduces membership-inference attack success by ≈4.5 percentage points on average. The algorithm runs in O(n·d) time when implemented column-wise, demonstrates millisecond-level per-sample runtime in our experiments, and scales approximately linearly with dataset size. These results indicate that sensitivity-aware, temporally adaptive budget allocation combined with lightweight smoothing can materially improve practical differential-privacy deployments.
Traffic flow prediction is challenging due to dynamic and complex spatio-temporal correlations. However, existing methods still suffer from two key limitations. First, many methods typically capture the dynamic spatio-temporal dependencies of traffic flows using static presets or adaptively learned spatial maps, which can only capture shared patterns over the entire time period, resulting in poor performance. In addition, most of the methods consider the absolute error between ground truth and prediction separately and independently at each time step, and still lack the ability to model the dynamic road network structure in a specific spatio-temporal context. To this end, this paper proposes an Adversarial Learning and Joint Spatio-Temporal Embedding based traffic flow prediction model (AL-JSTE), which combines Dynamic Graph Convolutional Networks (DAGCN) with Generative Adversarial Networks (GAN) for traffic flow prediction. Specifically, in the temporal dimension, a Temporal Decoupling Layer (TDL) is constructed to reveal heterogeneous traffic patterns using frequency domain analysis; in the spatial dimension, a Dynamic Adaptive Graph Generation (DAGG) is constructed to succinctly and efficiently model dynamic spatial dependencies in an adaptive manner modeling; Then, Adversarial Dynamic Trend Alignment (ADTA) is designed, which contains two discriminators to maintain the global properties of the predicted time series to further enhance the predictive capability of the model. Finally, extensive experiments on six benchmark traffic flow datasets are conducted in this paper, which fully demonstrate the excellent prediction performance of the proposed AL-JSTE model.
To achieve accurate and efficient classification of e-commerce traffic big data, a proposed method based on the Clara algorithm is used. The approach involves analyzing the components of e-commerce traffic big data in relation to the transaction process. The precision and time consumption of classifying and mining e-commerce traffic big data are significantly influenced by the effectiveness of data clustering. The Wavelet threshold function is utilized to de-noise and smoothen e-commerce traffic big data. This enhances the quality of traffic data. Subsequently, the Clara algorithm is employed to mine and cluster the de-noised traffic data. Finally, the support vector machine is used to classify the clustered data. The experimental results show that the method proposed in this paper performs clustering and classification mining on e-commerce traffic big data. The clustering results show no data overlap between categories, and the average accuracy of classification mining reaches 95.6%, with a maximum value of 97.9%. At the same time, the maximum single classification mining time is only 14.5 minutes, significantly improving the accuracy and efficiency of e-commerce traffic big data classification mining.
The fifth-generation (5G) wireless communication standards have significant potential for implementation using non-orthogonal multiple access (NOMA). Successive interference cancellation (SIC) is an essential receiver-side technique used in both downlink and uplink communication in typical NOMA detection. The old SIC approach limited performance due to the difficult multipath channel setup and the errors that arose. One advantage of the deep-learning approach is that it is a highly effective tool that circumvents the limits of traditional detection techniques. Thus, 5G communication based on Power-Domain Non-Orthogonal Multiple Access (PD-NOMA) with an optimized deep19net estimator is used to enhance spectral efficiency. At first, the User Equipment data are collected and processed for transmission. The transmitter base station consists of Quadrature Amplitude Modulation (QAM), superposition coding, a cyclic prefix, an Inverse Discrete Wavelet Transform (IDWT), and a serial-to-parallel converter. The inverse process is performed to obtain the processed signals at the receiver base station. The superposition-coded signals are detected using the Gold Rush Optimization-Deep19 Neural Networks (GRO-DNN19NEt)- based SIC to accurately recover the superpositioned signals. These signals are then demodulated using QAM. For evaluation of the model, the performance metrics such as Fairness, Average Sum Rate, Average Energy Efficiency, Throughput, Error, F1_score, Recall, Precision and Accuracy are considered, whose values are 0.68 (300 users), 48b/s/Hz (300 users), 239 (b/J/Hz) (300 users), 56 Mbps (300 users), 5.70%, 92.44%, 91.8%, 93.1 and 94.3. The proposed approach outperforms the existing model based on the performance metrics obtained. As a result, the proposed approach, which is based on PD-NOMA and uses an optimized deep19net estimator, performs better by increasing the model's spectral efficiency.
Data publishing of the intelligent logistics big data platform is an important part of data management and sharing on this platform which regularly publishes logistics related data, mainly logistics order data, to the public. Although sensitive identifiers such as logistics order numbers, names, phone numbers, etc. in logistics order data have been desensitized using private desensitization technologies such as generalization, now a more effective privacy protection still has not been provided to the location data in the logistic order datasets. Attackers can combine the location data, use data mining, machine learning, and other methods to infer the sensitive information of the shipper and recipient, resulting in the leakage of sensitive information. Therefore, we propose a differential privacy protection scheme for shipper and receiver location data in logistics orders. Our approach first use the non-uniform Quadtree method to partition the location data space, and use the hierarchical progressive privacy budget to allocate for each node. Then, an L-staircase data perturbation algorithm based on a staircase noise mechanism and a epsilon-LLCDP privacy protection algorithm suitable for publishing logistics order position data were designed. Finally, the privacy and time complexity of the epsilon-LLCDP algorithm were theoretically analyzed. The experimental results indicate that the location data of the logistics order data in the intelligent logistics big data platform processed by epsilon-LLCDP not only retains its usability, but also avoids the leakage of the location privacy information related to the shipper and payee in the publishing of the logistics order data.
The present study proposes an integrated clustering framework for spatiotemporal data that considers various features, such as occupancy, spatial coordinates, and temporal usage patterns, to identify functionally similar groups of transit stops. The framework integrates a dimensionality reduction with a clustering algorithm and an optimization algorithm for hyperparameter tuning. Moran's I is used as the objective function to ensure spatial validity of the resulting clusters. The methodology is evaluated on Electronic Ticketing Machine (ETM) data from a public bus network, clustering 176 bus stops into 8 distinct groups, each reflecting unique temporal and directional usage patterns. Results show that Uniform Manifold Approximation and Projection (UMAP) with Whale Optimization Algorithm (WOA) and Affinity Propagation effectively capture complex spatial patterns, while UMAP with WOA and HDBSCAN performs better under sparse or irregular data conditions. In terms of key performance indicators (KPI), UMAP with WOA performs best with respect to structural parameters and spatial consistency, whereas UMAP with Bayesian Optimization (BO) demonstrates superior computational performance with a marginal drop in the KPIs. This reflects a clear trade-off between clustering quality and runtime performance. Compared to traditional clustering methods, deep learning-based approaches (e.g., Variational Autoencoders), and optimization using Bayesian techniques, the proposed framework consistently demonstrates improved spatial coherence and interpretability. These findings highlight the potential of the proposed framework for scalable analysis of spatiotemporal networks.
Joint Video Moment Retrieval (MR) and Highlight Detection (HD) are essential for understanding the evergrowing volume of video content. Existing Transformer-based methods often struggle with misaligned features in complex video-text scenarios and quadratic computational costs. To overcome these bottlenecks, we propose MomentMamba, a unified and efficient framework built upon Selective State Space Models (Mamba). First, we design a multimodal alignment module that integrates Optimal Transport (OT)-based local alignment with a global contrastive strategy, enabling precise and robust semantic alignment between video and text at both token and modality levels. Second, we propose a novel Mamba-based highly-efficient architecture that comprises a Collaborative Mamba module to efficiently capture both modality-specific information and cross-modal interactions, and a Mamba decoder that enables joint MR and HD prediction, which effectively avoids high computational costs. Extensive experiments on QVHighlights, TVSum, and Charades-STA benchmarks demonstrate that MomentMamba consistently outperforms state-of-the-art methods in both MR and HD tasks, while at the same time reducing the theoretical complexity from quadratic O(L2 d) to linear O(L d) and achieving a 9% reduction in the average training time per iteration.
Correctly and reliably identifying AI-text reuses and human text reuses is extremely important for the preservation of authenticity (or lack thereof, and the associated potential for misinformation), the preservation of the integrity of language, and the mitigation of misinformation. This is particularly critical because current AI text detection tools, which identify and classify a single uncontextualized text, are limited in addressing these challenges. By operating at the phrasal level, detection frameworks can identify nuanced AI-generated transformations that sentence or document-level methods often overlook, thereby ensuring precise differentiation between human-authored and AI-reused text while maintaining the integrity of the language.For the first time, and in Urdu, we focus on text generation at the phrasal level, and for this purpose, we have generated a benchmark corpus of 9,908 aligned phrase pairs, divided equally between human text rewrites (UTRD-Phr-23) and Generative Pre-trained Transformer 4 (GPT-4-Mini) generations. This benchmark captures a wide variety of paraphrasing and intrinsic and extrinsic linguistic phenomena, with a lexical overlap of 52.81% and a transformation ratio of 0.717, covering both Overlap Rewriting (OR) and Non-Overlap Rewriting (NOR) phenomena.We test ten models, ranging from basic overlap models to state-of-the-art models including multilingual sentence embeddings (i.e., LaBSE, MiniLM) and generative models (i.e., GPT-2, Large Language Model Meta AI (LLaMA)). The proposed Feature Fusion model, which combines transformer and generative models within a Random Forest classifier, achieves 97.32% accuracy and F1 score for Urdu, surpassing all previously established benchmarks and matching or exceeding English-language model performance.This research provides a scalable approach for distinguishing between human and AI-generated text reuse content across multiple domains, including academic writing, digital journalism, social media content moderation, semantic reuse detection, authorship attribution, and content authenticity verification, supported by both AI detection and the benchmark corpus.
The development of 5G technology in the telecom sector has a significant impact on people's quality of life and the expansion of the national economy. Extracting useful features from large, complex consumption data remains difficult, despite the fact that a number of recommender systems for telecommunication packages have been proposed recently. In this paper, a recommendation system based on Generative Pre-trained Transformer-based feature selection and a GRU-SGO-based classification approach is developed. Reviews of various network services from websites are considered input. Pre-processing is performed in various stages, such as tokenization, case conversion, spell correction, stop-word removal, stemming, and lemmatization. Generative pre-trained transformer-3 (GPT-3) is used as a feature extraction algorithm to convert raw data into numerical features. Following feature extraction, a feature selection procedure selects the most pertinent features from the original set while discarding redundant, irrelevant, or distracting ones. Neighborhood Component Analysis (NCA) is employed as a feature selection method to select appropriate features. The selected features are then fed into a hybrid Gated Recurrent Unit-Squid Game Optimization (GRU-SGO) model for the final prediction across five classes based on customer reviews. SGO is utilized in the GRU for optimal selection of hidden neurons. The proposed system achieves 96% accuracy, 91% precision, and 97% specificity. Thus, the proposed GPT-3 and optimization-based feature selection algorithm is the best choice for an effective recommendation system.
In recent years, transformer-based architectures have achieved remarkable success in time-series forecasting, primarily owing to their ability to effectively capture long-range dependencies through self-attention mechanisms. These models surpass traditional forecasting approaches by offering flexible contextual modeling as well as highly parallel computation, making them particularly suitable for complex sequential tasks such as stock price prediction. Nevertheless, standard Transformer-based techniques often encounter difficulties in modeling fine-grained local variations and nonlinearities that characterize volatile financial data. Moreover, their reliance on positional encodings and purely attention-based structures can lead to inadequate representation of abrupt shifts and localized temporal dynamics, which are crucial for accurate short-term forecasting. To address these challenges, we propose ARBFformer, a novel hybrid model that integrates an attention-augmented radial basis function (RBF) neural network within an inversed transformer-based framework. The RBF-based encoding component captures local nonlinear fluctuations with high precision, while the attention mechanism adaptively reweights these radial responses to enhance feature discrimination. This localized representation is subsequently fused with an inversed Transformer encoder that models global temporal dependencies in a reversed generative manner, improving stability and long-horizon prediction. Furthermore, advanced pooling strategies are incorporated to optimize feature aggregation as well as reduce redundancy across temporal dimensions. Extensive experiments conducted on multiple real-world stock price datasets demonstrate that the proposed ARBFformer consistently outperforms recent state-of-the-art transformer-based baselines in terms of forecasting accuracy, robustness, and generalization. These results highlight the effectiveness of combining local nonlinear modeling with global contextual learning for robust financial time-series forecasting.
Colorectal cancer is a solid tumor whose molecular subtypes which is sufficient for precision medicine in order to improve treatments. This leads to a reduction in classification accuracy and clinical reliability of prediction models on high-dimensional gene expression data. To address this issue, this study proposes a multi-level clustering framework for predicting the subtype of colorectal cancer using gene expression profile. The methodology integrates structural and relational clustering with a Modified Plasma Formation Optimization (MPFO) algorithm for effective outlier removal. The dimensionality reduced by using the Dynamic Swarm Optimization (DESO) based feature selection technique that preserves informative genes. The Kaggle dataset is used in classification using ensemble models from AlexNet, VGGNet, Inception and ResNet. The findings indicate that the proposed MPFO+DESO+ResNet model achieves the best performance with an accuracy of 96.589%, outperforming CNN by 5.54% and Random Forest by 14.29%. Evaluation of all clusters shows 9% accuracy improvement for all types. Confirming the proposed model for enhancing predictive accuracy robustness as well as computational efficiency, the proposed framework is shown to be an effective mechanism for colorectal cancer molecular subtype prediction and precision medicine.
The accurate prediction of gas emission rates is crucial for mine ventilation and gas disaster prevention. However, traditional methods often fail to capture the complex nonlinear relationships among multidimensional features, and data scarcity further limits model performance. This paper proposes TDG-Net (Temporal and DualGranularity feature extraction Network), a gas emission prediction framework combining CopulaGAN-based data augmentation with a dual-branch spatiotemporal feature extraction architecture. CopulaGAN explicitly models the marginal distributions and correlation structures of tabular features, generating high-quality synthetic samples that alleviate the small-sample problem while avoiding the mode collapse of standard GANs. The dualbranch architecture consists of a temporal convolutional attention branch that captures dynamic temporal dependencies via multi-head attention and dilated causal convolutions, and a Dual-Granularity Feature Interaction Module (DGFIM) that models inter-feature coupling at both fine-grained pairwise and coarse-grained cross-group levels, with fine-grained and coarse-grained representations integrated through a dimension-wise gating mechanism within DGFIM. The outputs of the temporal and spatial branches are concatenated and compressed through a two-layer fully-connected network, and then input to XGBoost, selected as the optimal predictor from 28 candidate algorithms. Experiments on a dataset of 322 samples from seven Chinese coal-producing provinces demonstrate that TDG-Net achieves R2=0.941, MSE=0.554, and MAE=0.491, outperforming all 28 baseline models and three state-of-the-art literature methods, confirming its effectiveness for gas emission prediction under small-sample conditions.
The Optimization of decision support systems represents a more important research problem today, especially in terms of increasing complexities of workloads in big data environments. In NoSQL data warehouses, column family schema organization is a critical factor in query performance and scalability. Current research work primarily focuses on static schema definition methods or simple clustering algorithms such as CLARANS, making them less adaptable to varying workloads and access patterns. In this paper, an adaptive and automatic column family schema design technique for NoSQL decision support systems is presented. The proposed method exploits attribute co-occurrence patterns extracted from analytical workloads to dynamically group attributes into column families. To assess the robustness and effectiveness of the approach, several clustering algorithms including K-Means, CLARANS, BIRCH, DBSCAN, and Mean-Shift are systematically evaluated and compared. Experimental validation will be performed on TPC-DS benchmark using a distributed Hadoop environment with the HBase NoSQL database, to show the benefit of our adaptive schema design over static schema designs in terms of analytical query performance.
