Intelligent Data Analysis: An International Journal

Subspace clustering (SC) approximates high-dimensional data as a combination of low-dimensional subspaces, which is suitable for high-dimensional data analysis across various domains including image segmentation and face recognition. Existing SC methods typically obtain the global structure representation solely through the self-representation of the samples, thereby neglecting the intrinsic local connections among the samples. Moreover, due to their inherent framework design, obtaining additional a priori information in unsupervised scenarios presents a significant challenge. To address these limitations, this paper proposes a new method, named Sample-Dependent Subspace Clustering with Elastic Structure Consistency Constraints (SDSC). Firstly, we introduce a new Elastic Structure Consistency Constraints (ESCC) strategy to measure global and local structures elastically. Benefiting from this strategy, SDSC can flexibly explore the structural information within the samples to obtain a comprehensive data representation. By employing the joint regularization term, SDSC can learn effective cluster assignment information directly from the constrained structured data representation, and the cluster assignment information and representation coefficient matrix are smoothly integrated into a unified framework and learn in a mutually reinforcing manner. This learning approach contributes to comprehensive and high-quality clustering results, enhancing the robustness and utility of SDSC. Extensive experiments on several real-world benchmarks and synthetic datasets demonstrate the feasibility and effectiveness of SDSC.

Recent deep model pruning methods predominantly focus on large-scale datasets and typically require finetuning before deployment. However, in real-world applications, pruning is often necessary for scenarios with fewer classification categories, where finetuning must be avoided to preserve the model’s generalization ability. To address these challenges, we introduce a novel pruning method called Cluster-based Redundancy Elimination (CRE). Specifically, CRE represents each convolutional kernel as a point in a high-dimensional space. A distance-based strategy is then used to compute a clustering radius for each convolutional layer. Based on these radii, core point filters are selected for pruning, as they represent redundant information that can be captured by neighboring filters in the high-dimensional space. This approach eliminates the need for finetuning, thus preserving the generalization of deep models. Extensive experiments on five benchmark datasets with limited classification categories, across multiple model architectures, demonstrate the effectiveness of our method and its superiority over several state-of-the-art pruning techniques.

Spatial co-location pattern mining aims to uncover associations among spatial features, enabling users to discover correlation knowledge from spatial datasets. However, as spatial datasets grow, traditional frameworks for mining co-location patterns produce an overwhelming number of redundant results, which complicates further analysis. This paper focuses on extracting worthy co-location patterns, which are concise summaries of prevalent co-location patterns. We introduce two similarity measures—feature-based similarity and distribution-based similarity—to evaluate redundancy between co-location patterns from both feature and instance perspectives. Using these measures, we propose a novel approach called the Worthy Co-location Patterns Mining algorithm (WCPM) to condense prevalent co-location patterns. Initially, we employ a clique-based method to discover prevalent co-location patterns and categorize them into Maximal Co-location Patterns (MCPs) and Non-Maximal Co-location Patterns (NMCPs). Subsequently, we cluster the MCPs to extract the feature-similar MCPs, and based on distribution similarity, identify the worthy MCPs from the clustering results. Finally, we design a top-down algorithm to mine Worthy Non-Maximal Co-location Patterns (WNMCPs). Experiments on both synthetic and real datasets demonstrate that WCPM outperforms similar state-of-the-art approaches in terms of compression power and running time.

Micro-expression is difficult to recognize due to short duration and subtle action range, but it contains rich and real psychological information, which has important research value in criminal investigation, teaching and other fields. In response to issues like limited facial expression dynamics, suboptimal feature extraction, and susceptibility to overfitting, we proposed a micro-expression recognition method based on optical flow and multi-task convolutional neural network (OFMT-Net). It capitalizes on optical flow data from onset to apex frames as input. Feature extraction is conducted through a shared-parameter network, funnelling outputs into a dual-tower network designed for emotional and Action Unit (AU) recognition. This network incorporates a self-attention mechanism for effective classification, driven by a dual weighted loss function. The method fully extracts the relevant information contained in the facial action unit, and uses the implicit data enhancement advantages of the multi-task framework to improve the recognition accuracy and reduce the sample dependence problems. Cross-validation results on the joint dataset demonstrate that the model achieves an accuracy rate of 79.89%, an unweighted average recall rate of 75.05%, and an unweighted F1 score of 75.08%, surpassing many mainstream models. The related code is publicly available at https://github.com/WenyuanLi001/OFMT-Net

With the rapid advancements in industrial big data, the Internet of Things, and sensor acquisition technologies, the similarity measurement of multivariate time series has emerged as a pivotal research area in data mining and machine learning. To enhance the accuracy and efficacy of multivariate time series similarity measurement, this paper proposes a sliding window approach based on Transformer. Specifically, each dimension of the multivariate time series is processed through sliding windows and input into a Transformer for feature extraction. By using multiple window sizes, the method simultaneously captures localized temporal segment features and identifies local patterns within the time series. Encoded window features for each sample are combined to form a comprehensive feature sequence that represents the global characteristics of the entire time series. These global features are then used to compute the final similarity measure through Dynamic Time Warping (DTW). This approach effectively captures both local and global features of multivariate time series, significantly improving similarity measurement precision. The effectiveness of the proposed method is validated through 1-Nearest Neighbor (1NN) classification experiments, demonstrating superior accuracy and enhanced performance in similarity measurement. The experiments showed that ten of the sixteen datasets had the best performance in terms of classification accuracy.

Due to the scarcity of annotated real-world data for specific categories, audio-visual generalized zero-shot learning (GZSL) has attracted significant attention. GZSL aims to classify novel classes absent during training while ensuring stable performance on seen classes. However, most existing methods operate implicitly, often neglecting the effective utilization of temporal, spatial, and semantic consistency. To address these challenges, we propose the Temporal Spatial Semantic Fusion network (TSSF). Specifically, we explore both audio and visual modalities using a multi-branch, multi-grained structure comprising a temporal global extraction module, a spatial local refinement module, and a multi-grained fusion module. The temporal global extraction module employs a Transformer-based Spiking Neural Network to extract explicit temporal representations and capture global dependencies. Simultaneously, the spatial local refinement module focuses on spatial information and local details using a window attention mechanism. Furthermore, the temporal and spatial features are hierarchically fused in the multi-grained fusion module, which incorporates both temporal and spatial attention for semantic enrichment. To explore multi-modal interactions, we enhance audio and visual features through cross-modal attention, followed by multi-modal alignment with text embeddings. Experiments on three benchmark audio-visual datasets validate the superiority of our method over state-of-the-art approaches. Notably, TSSF achieves significant improvements of 30.34% and 6.96% in HM and ZSL metrics on the VGG-GZSL dataset.

This paper proposes a unique approach to enhance image captioning by leveraging an Asynchronous Dual Attention (ADA) mechanism within a Vision Transformer (ViT) based framework. Traditional deep-learning models for image captioning often struggle with multimodal interactions and capturing local-to-global visual contexts, including both prominent and subtle features. To address this, the proposed model integrates global self-attention (ViT-B/16) with a Joint Calibration Module during image encoding to enhance the quality of visual embeddings and combines dynamic step-wise attention (Bahdanau) with a Gated Recurrent Unit (GRU) during decoding. This forms an ADA pipeline that decouples visual and linguistic pathways, allowing adaptive refinement of visual features and more precise alignment with linguistic context. Unlike synchronous attention models, ADA enables dynamic image region selection and improved spatial reasoning through enhanced multimodal interaction, leading to more contextually coherent and informative captions for complex visual scenes. The proposed approach demonstrates consistent improvement over state-of-the-art methods on benchmark datasets, achieving CIDEr scores of 0.946 and 1.364 and SPICE scores of 0.188 and 0.248 for Flickr 30k and MSCOCO datasets, respectively. Additionally, the framework incorporates Google’s text-to-speech synthesis to generate audio captions, enhancing accessibility for visually impaired users.

Temporal Knowledge Graph (TKG) representation learning aims to project entities and relations from high-dimensional spaces into low-dimensional ones while preserving dynamic relational characteristics. However, many existing methods primarily focus on single-time-stamp Knowledge Graphs, neglecting the importance of time in capturing the evolving relationships within TKGs. To address this limitation, we introduce TAR-TKG (Temporal-Aware Representation for Temporal Knowledge Graph), a novel framework that consists of three core modules. The first module, the Temporal Dynamics Steering Module, enhances dynamic temporal features by employing a multi-time-awareness network to capture changes between time stamps, thereby improving the understanding of temporal data evolution. The second module, the Cross-Time Domain Gating Module, applies cross-time domain graph convolution and adaptive gating to learn relationships between different time stamps, facilitating the integration of information across multiple time spans to improve the accuracy of temporal reasoning. The third module, the Temporal Adaptive Relation Perception Module, combines temporal embeddings, causal reasoning, and multi-modal relation fusion to enhance the model’s ability to perceive temporal relationships, particularly in managing causal dependencies and complex time-based interactions. Experimental results demonstrate that TAR-TKG outperforms existing baseline methods on three real-world datasets, proving its effectiveness in capturing dynamic relationship evolution and improving temporal reasoning within TKGs.

Model-Agnostic Meta-Learning (MAML) has proven to be effective in various learning environments. However, it faces challenges with domain adaptation because it depends on gradient-based optimization, which does not explicitly integrate prior knowledge from related tasks. This limitation results in slow adaptation to new domains and suboptimal performance when Signiant domain shifts occur. Utilizing transfer learning, which skillfully incorporates domain-specific knowledge to boost generalization and adaptability, effectively resolves the challenges faced by current MAML-based techniques across various environments. This study explores the use of transfer learning to create a strong and flexible model that can effectively detect occurrences of data loss or leakage within credit cardholder information datasets. The model is trained on a source domain and ne-tuned on a target domain relevant to data loss detection by leveraging transfer learning. The effectiveness of the Transfer Learning-Based Data Loss Detection on MAML is evaluated through learning iteration versus mean squared error plots. The proposed system also surpasses the existing few-shot learning-based MAML. These plots provide insights into the model's convergence, adaptability, and performance. The abstract highlights the signicance of transfer learning in enhancing the efficiency and accuracy of data loss detection systems, particularly when utilizing the MAML is evaluated through learning iteration versus mean squared error plots. The proposed system also surpasses the existing few-shot learning-based MAML. The findings contribute to the expanding knowledge of transfer learning applications in cybersecurity and data protection. Experiments conducted on the IEEE-CIS Fraud Detection dataset demonstrate that our approach achieves an accuracy of 92.3% and a notable reduction in MSE by 15% compared to standard MAML, underscoring its effectiveness and robustness across various environments.

Early hospital admission prediction at the triage stage is an important and challenging task for emergency departments (EDs), aimed at effectively managing and utilizing limited medical resources for critical patients. A retrospective study was conducted at MacKay Memorial Hospital (MMH) from 2011 to 2018, including 1,061,760 records of valid patients, using logistic regression (LR), eXtreme Gradient Boosting (XGBoost), Word2Vec, and bidirectional encoder representations from transformers (BERT). The chief complaints (CCs) and limited structured variables collected at triage are considered predictor variables. The results show that XGBoost achieves better prediction than LR with patient structured variables and better prediction than Word2Vec with patient CCs in terms of AUC and F-measure. We further propose the novel concept of generating expanded CCs as BERT input by integrating the original CCs with selected structured variables using XGBoost to predict the probability of patient admission. Among the structured variables, triage category, mode of arrival, age, arrival time, and fever status are the most important. This study demonstrates BERT's (in particular, BERT-ROS with 5 variables) superior prediction capability compared to other models by considering only patient CCs or expanded CCs in terms of AUC and F-measure. Moreover, given the low admission rates in Taiwan's EDs, this study employs imbalanced data processing to show that the proposed method enhances the predictive capability of hospitalization. These experimental results provide a reference model with associated variables for developing a hospital admission tool at triage, identifying the risk of stratification of critical patients.

In this cognitive era, vast amount of data are accumulated every day. Analysing such unstructured information and obtaining insights will be challenging. To address this, Large language models have been developed to support analysis of extensive data corpora. However, it tends to cause hallucination due to a lack of proper knowledge sources. If the analysis has to be performed with respect to the health care domain or finance domain, the challenge is raised because of the lack of domain specificity. COVID-19 sentiment analysis is one of the complex responsibilities of the government since it needs to know the opinions of people to take necessary measures. This paper presents COVID-19 retrieval augmented and fine-tuning (RAFT), a novel framework that includes the analysis of COVID-19 vaccine tweets through retrieval augmented-based approaches. This integrated domain-specific knowledge through a retrieval-augmented generation-based approach with external knowledge sources. We employed a transformer-based semantic approach in embedding generation via vector database. Furthermore, this framework exhibited generalizability when integrated with domain knowledge. It uses parameter efficient fine tuning with quantization to use a large language model with a reduced number of parameters, which will allow a model to be used in low-resource-constrained devices. This framework achieved an accuracy of 0.886 on the Twitter dataset containing tweets specific to Indian region and 0.912 on the Twitter dataset with tweets from Global region.

Cracks pose a significant threat to road and building safety, making effective detection of cracks on road surfaces a focus of research both domestically and internationally. Deep learning-based methods often require extensive pixel-level annotations, posing significant labor costs. We propose a single-stage weakly supervised crack segmentation model based on multi-scale feature fusion. The model is built on a single-stage weakly supervised segmentation framework, which reduces model complexity. It utilizes a multi-scale feature fusion module (PPM) to integrate features at different scales, enhancing the ability to extract features from cracks of various sizes. The combination of the Domain Restriction Suppression (DRS) module and pixel affinity convolution is employed to optimize pseudo-pixel annotations. In addition, we propose a joint loss function to mitigate sample imbalance between crack and non-crack pixels. Compared to other two-stage weakly supervised segmentation models, our model is simpler and more effective, achieving excellent results on the Deep Crack and Crack500 datasets, surpassing most weakly supervised crack segmentation models in terms of Recall (Re), F-score (F1), and mean Intersection-over-Union (mIoU), achieving similar effects to fully supervised crack segmentation models. This demonstrates the effectiveness and robustness of our model.

Accurate solar energy prediction is critical for optimising renewable energy systems, particularly in regions with diverse climatic conditions. This study investigates the impact of humidity and temperature on solar energy prediction accuracy in Ghanaian climates. It uses a hybrid Dilated Temporal Convolutional Network and Long Short-Term Memory (DTCN-LSTM) model to capture diurnal patterns and normalised temperature, humidity, and solar irradiance data from 2010–2022. The results show exceptional predictive accuracy, with R² values exceeding 99.86 and near-zero RMSE values of less than 0.0055 kWh across all areas studied, representing 0.098 percent of the designed 560Wh of the solar energy simulated. Temperature exclusion caused the largest performance decline, while humidity exclusion had negligible beneficial effects. The DTCN-LSTM model exhibited strong generalisation, with minimal training-testing discrepancies R² ≤ 0.07%, and maintained high R² accuracy of greater than 99.87% even when both temperature and humidity were excluded, highlighting its adaptability to sparse sensor environments. The findings show that prioritising temperature data collection over humidity improves solar energy prediction accuracy, and the DTCN-LSTM robustness across diverse climates is a vital tool for enhancing grid stability and energy storage planning in variable environments. This work further advances adaptive machine learning frameworks for renewable energy systems, emphasising scalability and operational practicality in global solar forecasting applications.

Most people primarily rely on subjective opinions and action assessment from others in the process of learning badminton, which leads to biases and unreliable assessment of player performance. This paper presents a method of deep learning-based recognition and quality assessment of badminton actions to accurately identify player actions and assess player performance. In this research, we construct a video dataset of standard badminton actions for training networks and design a human pose estimation and tracking network to detect keypoints of individual players in the video dataset and track their trajectories. Furthermore, badminton action recognition is carried out based on the SlowFast network framework, and a Siamese network with it as the backbone network is proposed for automating the quality assessment of badminton actions. Experimental results demonstrate that the mean average precision (mAP) of human body pose estimation reaches 83.2%, the Multiple Object Tracking Accuracy (MOTA) of pose detection and recognition in badminton games reaches 81.4%, and the Multiple Object Tracking Precision (MOTP) reaches 90.7%. The accuracy of professional players in identifying badminton strokes is 83.08% for Top-1 and 96.89% for Top-3. Therefore, the proposed method can be effectively applied to badminton action recognition and quality assessment.

In the rapid development of artificial intelligence, multimodal large models (VLM) have led a new wave of technological progress with their revolutionary breakthroughs in the field of natural language processing (NLP). In the wave of artificial intelligence, on-device multimodal large models (On-Device VLM) are becoming the new favourites of technological innovation with their rapid development speed and broad application prospects, and the demand for on-device inference is growing. This study conducts in-depth adaptation and optimization of multimodal large models on the Neural Network Processing Unit (NPU) based on the Qualcomm platform. By adopting the QNN (Qualcomm Neural Network) framework and model compression techniques such as quantization, pruning, knowledge distillation, low-rank factorization, and Lookahead decoding, efficient inference acceleration on Qualcomm NPU is achieved, significantly improving the model’s response speed and decoding efficiency. Experimental results show that the optimized model has significant improvements in first response time and decoding speed, providing a new solution for on-device AI applications.