Home      About Journal      Editorial Board      Instructions      Ethics Statement      Contact Us                   中文

Judgement and Disposal of Academic Misconduct Article Copyright Transfer Agreement Proof of Confidentiality Requirements for Electronic Version More.... Chinese Association of Automation National ResearchCenter for Intelligent Computing System Institute of Intelligent Machines,Chinese Academy of Sciences More....

2026 Vol.39 Issue.8, Published 2026-08-25 Papers and Reports    Researches and Applications        Papers and Reports 665 Textual Causal Feature Selection via Contrastive Learning and Large Language Model YANG Jiahao, YANG Juntao, CAO Dayuan, XIONG Shoujiu, YU Kui Learning causal features related to a target variable from textual data is considered as an important approach to improve model interpretability and decision reliability. However, textual causal feature selection still faces two challenges. Latent features are usually embedded in unstructured text, and latent feature semantics are often mixed with sentiment information in sentence representations. Consequently, latent feature extraction results are unstable. Stable and efficient causal feature selection becomes difficult due to the dynamic changes of latent features during extraction, assignment and supplementation. To address these issues, a method for textual causal feature selection via contrastive learning and large language model(TCFS) is proposed in this paper. First, a decoupled latent feature extraction method is designed to address the instability of latent feature extraction, and iterative clustering optimization is introduced to construct semantically consistent and structurally stable latent feature clusters. Second, a dynamic causal feature selection method is proposed, and the stable identification of key causal features and the removal of redundant features are achieved by dynamically maintaining the Markov blanket associated with the target variable. Finally, unstructured text is transformed into structured numerical data through latent feature assignment, and a missing-feature discovery mechanism is introduced to analyze hard-to-explain samples and iteratively supplement the missing features, thereby continuously optimizing the latent feature set and the feature selection results. Experiments demonstrate that TCFS effectively improves the stability of latent feature extraction and the accuracy of causal feature selection in text scenarios and provides effective support for interpretable prediction and textual causal analysis. 2026 Vol. 39 (8): 665-682 [Abstract] ( 15 ) [HTML 1KB] [ PDF 983KB] ( 13 ) 683 Global Trajectory-Aware Multi-camera Multi-target Tracking FAN Yuteng, WANG Qiang, ZHEN Yihao, CHEN Xiai, MA Chi, FAN Huijie In existing global tracking methods, global trajectories are mainly treated as collections of historical target features. However, trajectory-level identity semantics are not explicitly modeled, and identity consistency constraints are not imposed. Therefore, the long-term identity consistency of targets across time and space cannot be fully exploited. To address these issues, a global trajectory-aware multi-camera multi-target tracking method termed GTAT is proposed. First, for each global trajectory, a trajectory-level representation with long-term identity semantics is aggregated by the global trajectory semantic encoding module. Then, global trajectory-level identity semantics are injected into target association features within the temporal window by the global trajectory-aware enhancement module through a cross-attention mechanism. Thus, trajectory-level contextual information can be explicitly perceived by target-level representations. Finally, a global trajectory consistency loss is designed to constrain the matching relationship between trajectory-level representations and target-level association features. Consequently, the ability of global trajectory representations to model identity consistency is improved. The identity discriminability of association features is also enhanced. Experiments demonstrate that GTAT achieves superior performance on six public multi-camera multi-target tracking datasets. 2026 Vol. 39 (8): 683-696 [Abstract] ( 16 ) [HTML 1KB] [ PDF 2395KB] ( 18 ) 697 Fuzzy Concept Lattice Model for Three-Way Knowledge Space ZHI Huilai, WU Weizhi, ZHU Daxin The core challenge of knowledge space theory is to effectively construct knowledge structures from skill data to accurately reflect the cognitive abilities of learners, and this directly affects the precision of personalized learning guidance. In existing generation approaches and concept lattice models, knowledge states are characterized only by positive attributes, and negative attributes are disregarded. Diagnostic information included in negative attributes is completely ignored. Therefore, the variety of generated knowledge states is severely restricted, and actual complex learning paths are not covered. Assessment precision and learning path planning are greatly constrained. To overcome this limitation, the three-way decision is adopted. Negative attributes are incorporated into knowledge modeling. A three-way concept lattice model based on fuzzy skill formal contexts is established for the first time. At the theoretical level, knowledge spaces and knowledge states are defined from a three-way perspective. It is rigorously proved that the extents of all three-way fuzzy concepts constitute a complete knowledge space. The correctness of the lattice structure is validated through supremum operations. At the methodological level, granular description is introduced to define skill-induced atomic granules. A decision criterion based on atomic granule combinations is proposed to identify whether a problem set forms a knowledge state. Experiments indicate that the number of knowledge states yielded by the proposed model is substantially larger than that of the conventional models that omit negative attributes. Thus, stronger discriminative power of the proposed model is validated. Theoretical analysis confirms that the three-way concepts are structurally independent of any existing three-way concept types, and they degenerate into three-way object-oriented concepts under binary contexts. These results collectively indicate that the proposed model completes the information dimension by incorporating negative attributes. Not only expressive power is enhanced, but also a conceptual analysis framework is provided for knowledge space theory from the perspective of completeness. 2026 Vol. 39 (8): 697-710 [Abstract] ( 9 ) [HTML 1KB] [ PDF 748KB] ( 6 ) 711 Visual-Tactile Multimodal Learning via Hierarchical Tactile-Guided Spatial Attention and Dynamic Temporal Fusion FAN Chenglong, HU Lihua, HU Jianhua Visual and tactile modalities differ in perceptual scope and information form. Existing methods fail to effectively associate tactile contact information with local visual regions. Moreover, fixed fusion strategies cannot adapt to the dynamic variations in the contributions of the two modalities across different interaction stages. To address these issues, a visual-tactile multimodal learning method via hierarchical tactile-guided spatial attention and dynamic temporal fusion(HTA-DTF) is proposed. First, tactile features are utilized as queries to guide the visual branch to focus on contact-related regions, while intermediate- and high-level visual features are integrated to capture both local texture details and high-level semantic information. Second, a dual-stream Mamba architecture is employed to model the temporal dependencies of visual and tactile sequences, and a dynamic gating mechanism is adopted to adaptively adjust the fusion proportions of the two modalities across different interaction stages. Finally, temporal attention pooling performs weighted aggregation over the fused sequence, emphasizing key interaction moments while suppressing interference from redundant time steps. Experiments on the Touch and Go material recognition dataset demonstrate that HTA-DTF achieves high material recognition accuracy. Ablation studies and visualization analyses further verify the effectiveness of the proposed components. 2026 Vol. 39 (8): 711-724 [Abstract] ( 14 ) [HTML 1KB] [ PDF 1813KB] ( 8 ) Researches and Applications 725 Traceable Evidence-Centric Generation for Digital Forensics LIN Guokai, SUN Yuanyuan, GUO Hong, Paerhati TULAJAING, YANG Liang, LIN Hongfei To address the problems of vulnerable evidence boundaries, difficulty in preserving source and location information, and insufficient verifiable support for generated results in multi-source heterogeneous electronic evidence scenarios, a traceable evidence-centric generation method for digital forensics(TEC-Gen) is proposed. First, multi-source forensic data, including chat records, file records, call records and system logs, is parsed in a unified manner. Original records are regarded as the basic units to construct evidence tuples. Within each evidence tuple, the original record text and metadata, including source files, location information, timestamps and key entities, are jointly modeled to preserve the original evidence boundary and its location anchor. Then, candidate evidence is organized according to a case analysis query. Relevant records are retrieved while structural attributes, including source files, location information and key entities, are preserved, providing a unified evidence basis for subsequent generation and verification. In the generation stage, the candidate evidence and citation rules are jointly incorporated into the prompt template, and a large language model is required to attach source markers after key factual statements. Finally, the output citations are parsed and matched with valid original positions in the candidate evidence. Explicit back-links between generated statements and original evidence are established. Experimental results show that TEC-Gen outperforms closed-book generation method and standard retrieval-augmented generation method on different large language model backbones. Stable advantages are observed in text alignment, key fact coverage, and result traceability. Methodological support is provided for case summary generation, evidence localization, and assisted analysis in digital forensics. 2026 Vol. 39 (8): 725-736 [Abstract] ( 12 ) [HTML 1KB] [ PDF 956KB] ( 9 ) 737 Deep Learning Based Multi-objective Multi-task Evolutionary Approach for Imbalanced Feature Selection WANG Haoren, YANG Jiaru, HUA Qichong, JIANG Shu, DING Weiping In data-intensive domains like medical diagnosis and financial risk control, feature selection is faced with severe challenges due to high-dimensional imbalanced data, including majority class bias, insufficient capture of high-order nonlinear feature interactions, and difficulty in collaborative optimization of conflicting multi-objectives. Traditional methods struggle to simultaneously achieve imbalanced data adaptation, high-order feature correlation mining, and multi-objective global optimization. Therefore, for supervised feature selection scenarios, a deep learning based multi-objective multi-task evolutionary approach for imbalanced feature selection(DLME) is proposed and a three-stage modular collaborative optimization framework is established. First, dynamic oversampling and preliminary feature filtering are performed by combining adaptive sampling and SMOTE based on particle swarm optimization, thereby alleviating class distribution imbalance. Then, a deep neural network is embedded. High-order nonlinear correlations among features are learned from the balanced sample set. The feature subset is subsequently refined. Finally, a grey wolf optimizer based multi-task multi-objective optimization framework is constructed. This framework integrates NSGA-II-based Pareto front search and front knowledge transfer mechanism. The optimal feature subset balancing classification accuracy and sparsity is output. Experimental results on public datasets and domain-specific datasets demonstrate the outstanding superiority of DLME under extremely imbalanced and high-dimensional scenarios. DLME exhibits excellent multi-objective optimization performance and robustness, thereby providing an effective solution for feature selection tasks on high-dimensional imbalanced data. 2026 Vol. 39 (8): 737-750 [Abstract] ( 10 ) [HTML 1KB] [ PDF 991KB] ( 8 ) 751 Neuro-Symbolic Collaborative Method for Route Planning in Remote Sensing YANG Ming, ZHOU Zhi, ZHANG Chenxi, TIAN Shiyu, YU Kunyang, LI Yufeng Task-oriented route planning from remote sensing imagery requires understanding of open-ended semantic constraints and fine-grained perception of complex land-cover environments. Existing methods relying on high-definition maps, road networks, vehicle-mounted sensors, or manually specified rules struggle to adapt to such scenarios. A neuro-symbolic collaborative route planner(NSCRP) for route planning in remote sensing is proposed. First, a multimodal large language model is utilized to translate natural-language tasks into structured traversability rules and preference representations. A stable semantic land-cover map is constructed by integrating multiple semantic segmentation models. Then, semantic constraints and environmental perception results are uniformly mapped into a pixel-level cost map. An interpretable path with the minimum accumulated cost is generated via heuristic search on the predicted cost map. Experiments based on NeSy-Route show that NSCRP outperforms existing approaches in constraint satisfaction, path quality, and overall planning ability. 2026 Vol. 39 (8): 751-760 [Abstract] ( 9 ) [HTML 1KB] [ PDF 861KB] ( 7 )

Supervised by China Association for Science and Technology Sponsored by Chinese Association of Automation NationalResearchCenter for Intelligent Computing System Institute of Intelligent Machines, Chinese Academy of Sciences Published by Science Press

Copyright © 2010 Editorial Office of Pattern Recognition and Artificial Intelligence Address: No.350 Shushanhu Road, Hefei, Anhui Province, P.R. China Tel: 0551-65591176 Fax:0551-65591176 Email: bjb@iim.ac.cn Supported by Beijing Magtech 　Email:support@magtech.com.cn