Editorial

Dear Colleague:

Welcome to volume 30(2) of the Intelligent Data Analysis (IDA) Journal.

Dear reader, welcome to this second issue of IDA's 30th year. This issue will feature papers on both theoretical contributions and applied data analysis cases.

The first part of this issue includes several theoretical contributions. The first group of papers in this section relates to clustering either as a learning task or as a support technique. First, in their paper, Zhu et al. propose Sample-Dependent Subspace Clustering (SDSC), a new method that overcomes the limitations of existing subspace clustering techniques by modelling global and local data structures simultaneously through Elastic Structure Consistency Constraints. By combining representation learning and cluster assignment in a unified, mutually supportive framework, SDSC delivers more robust and accurate clustering, as demonstrated on benchmark and synthetic datasets. The second paper on the topic, by Huang et al., introduces Cluster-based Redundancy Elimination (CRE), a new deep model pruning technique designed for scenarios with limited classification categories where fine-tuning is not possible. By clustering convolutional kernels in high-dimensional space to identify and remove redundant filters without retraining, CRE preserves model generalisation and outperforms existing pruning methods across various benchmarks and architectures. In the third contribution on clustering, Bao et al. present the Worthy Co-location Patterns Mining (WCPM) algorithm to efficiently extract concise and meaningful co-location patterns from large spatial datasets by reducing redundancy. Using feature and distribution similarity metrics and combining clique-based discovery with clustering, WCPM outperforms existing methods in both compression efficiency and runtime on synthetic and real datasets.

The second group of theoretical papers includes four papers on deep learning using either transformers or attention mechanisms. The first of these papers, by Li et al. (a), presents OFMT-Net, a micro-expression recognition method that combines optical flow with a multi-task convolutional neural network to detect subtle facial movements from onset to apex frames. Employing dual-tower systems for emotional and Action Unit recognition with self-attention, the model achieves higher accuracy than standard methods while also reducing reliance on samples and enhancing feature extraction. The second paper in this group, by Wu et al., presents a Transformer-based sliding window technique for measuring similarity in multivariate time series, capturing both local temporal patterns and global sequence features. By merging window-encoded features with Dynamic Time Warping, the method notably enhances accuracy in similarity measurement, achieving superior 1-Nearest Neighbour classification performance on ten of the sixteen benchmark datasets. The last two papers of this selection are also interesting theoretical contributions for audio-visual analysis. First, Guo et al. propose the Temporal Spatial Semantic Fusion (TSSF) network for audio-visual generalised zero-shot learning, addressing data scarcity by explicitly modelling temporal, spatial, and semantic consistency across modalities. By combining Transformer-based temporal extraction, spatial refinement, multi-grained fusion, and cross-modal alignment with text embeddings, TSSF significantly outperforms state-of-the-art methods on multiple benchmarks. The last paper of this group, by Patra & Kisku, introduces an Asynchronous Dual Attention (ADA) mechanism within a Vision Transformer framework to enhance image captioning by separating visual and linguistic pathways and improving visual-linguistic alignment. The method surpasses current leading techniques on benchmark datasets and incorporates text-to-speech synthesis for creating accessible audio captions.

We conclude the theoretical research section with two papers. The first is a contribution on knowledge graphs by Zhao et al., who present TARTKG, a Temporal-Aware Representation framework for Temporal Knowledge Graphs, designed to capture evolving relationships across multiple timestamps. By integrating temporal dynamics, cross-time domain graph convolution, and adaptive relation perception, TARTKG outperforms existing methods in modelling dynamic relationships and improving temporal reasoning on real-world datasets. The last of the theoretical contributions is by Naveenkumar & Karthikeyan. This paper improves Model-Agnostic Meta-Learning (MAML) for detecting data loss in credit cardholder datasets by adding transfer learning to include domain-specific knowledge, boosting adaptability to new domains. Experiments on the IEEE-CIS Fraud Detection dataset show that the proposed method attains high accuracy and lowers errors compared to standard MAML, demonstrating better performance and robustness.

The second section of this issue is devoted to applied contributions and use cases. The first two of these contributions are healthcare applications. In the first, Feng et al. assess early hospital admission prediction at the emergency department triage using over one million patient records and compare machine learning and NLP models based on structured triage data and chief complaints. Results demonstrate that incorporating key structured variables into expanded chief complaints and modelling them with BERT significantly enhances admission prediction performance, especially in imbalanced data conditions. The second medical contribution, signed by Prasika & Rajan, introduces COVID-19 RAFT, a retrieval-augmented and fine-tuned framework for analysing COVID-19 vaccine tweets by integrating domain-specific knowledge with transformer-based embeddings. The approach achieves high accuracy on both regional and global Twitter datasets while enabling parameter-efficient deployment on low-resource devices.

The second group of applied papers are two industrial problems. The first paper by Shen et al. introduces a single-stage, weakly supervised crack segmentation model for concrete and other materials that employs multi-scale feature fusion to detect cracks of different sizes while minimising annotation needs and reducing model complexity. By integrating Domain Restriction Suppression, pixel affinity convolution, and a joint loss function, the model attains superior performance on benchmark datasets, matching the effectiveness of fully supervised approaches. The second industrial application is the paper by Darthe et al., which demonstrates that a hybrid Dilated Temporal Convolutional Network and Long Short-Term Memory (DTCN-LSTM) model can accurately predict solar energy in Ghana, achieving an impressive coefficient of determination and near-zero error rates by effectively capturing diurnal patterns from temperature, humidity, and irradiance data. The results highlight temperature as the most critical factor for prediction accuracy, with the model showing strong generalisation and robustness, making it a practical tool for scalable solar energy forecasting and grid management.

The final application paper presents a use case in sports science. The paper by Li et al. (b) proposes a deep learning-based method for badminton action recognition and quality assessment using human pose estimation, tracking, and a SlowFast-based Siamese network. Experimental results demonstrate high accuracy in pose detection, tracking, and action recognition, showcasing the method's effectiveness for automated performance evaluation in badminton.

We conclude this issue with an insightful contribution on support technologies and cutting-edge data analysis implementations. The paper, authored by Zhu & Lu, explores the adaptation and improvement of multimodal large models for on-device inference on Qualcomm Neural Network (QNN) using Network Processing Units (NPUs), addressing the increasing need for efficient on-device AI. By employing the QNN framework along with various model compression and acceleration techniques, the approach markedly enhances response times and decoding speeds, enabling more practical on-device multimodal applications.

With my best regards,

Dr. J.M. Peña

Editor-in-Chief