Graph contrastive learning can effectively capture the complex structural dependencies in behavior graphs, and now it is an important tool for malware detection. However, existing methods treat all samples indiscriminately or overemphasize hard samples during training, leading to training oscillation and slow convergence. To address this issue, a prototype-based graph contrastive learning method for malware detection is proposed. Curriculum learning is extended from the instance-selection level to the sample-pair level, and the contrastive difficulty of software semantics is defined by jointly considering the typicality of anchors and negative samples. A dual-threshold dynamic scheduler is designed to enable progressive contrastive training from typical cases to edge cases. Furthermore, a history-guided anchor enhancement mechanism that tracks and reweights insufficiently learned hard anchors is introduced, thereby enhancing the representation capability for hard samples. Experiments on four public malware datasets demonstrate that the proposed method significantly improves training stability and convergence speed.

The integration of concept-cognitive learning and fuzzy three-way concepts is an important research direction in uncertain data analysis. However, existing methods still have limitations in the utilization of attribute information, the characterization of the information contributions of pseudo-concepts, and the efficiency of processing relatively large-scale data. To address these issues, a fuzzy granular three-way concept-cognitive learning approach with weighted pseudo-concepts(FG3WCCL) is proposed. First, attribute granulation is introduced, fuzzy three-way attribute granular concepts are defined, and a granular concept space is built. Then, the concepts are clustered according to the similarity of concept extensions. The weights are determined by jointly considering the information contributions of different concepts, the hierarchical decay factor, and hierarchical consistency. Weighted pseudo-concepts are consequently constructed. Furthermore, a spectral decomposition-based object reduction method is proposed. Representative objects are selected through hybrid sampling to reduce the modeling cost. Finally, comparative experiments on 14 datasets show that FG3WCCL achieves high classification accuracy.

Hybrid policy frameworks combining reinforcement learning with geometric planners achieve high success rates in automatic-parking tasks. However, oscillatory gear-shifting behavior is still observed in some complex scenarios, and behavioral preferences cannot be adequately represented by manually designed scalar reward functions. To address these problems, a method for reward modeling and automatic parking policy optimization with large language model-based semantic assessment is proposed as a four-stage policy-optimization framework. First, trajectory pairs containing oscillatory negative samples are mined from the hybrid policy system, and preference labels are generated by ensemble voting among four models. Then, a two-stage reward model is trained through environment-reward pre-training and preference-residual fine-tuning. Preferences of large language models for trajectory quality are converted into sortable reward scores. Finally, elite trajectories are selected by scenario for behavior cloning. Experiments on 500 simulation scenarios constructed from perception data of real parking lots show that the proposed method improves parking success rates and greatly reduces the average number of forward-reverse gear switches during each parking process.

Visible-infrared person re-identification is a key technology for all-day video surveillance. However, constrained by modality heterogeneity, pose variations, and occlusions, existing methods still suffer from misalignment in local semantic modeling, a gap between training and inference objectives, and insufficient robustness in single-space matching. To address these issues, a dual-space constrained method based on visual-semantic dynamic anchoring(DS-VSDA)is proposed. First, a semantics-driven prototype dynamic anchoring module is designed to introduce spatial prior initialization of human body parts, thereby alleviating semantic redundancy and drift. Then, a cross-sample neighborhood affinity loss is introduced to explicitly model the topological relations among samples within a batch, bridging the training-inference gap. Finally, a dual-space semantic consistency constraint and a matching calibration mechanism are constructed to fuse shallow and deep features and optimize matching results through mutual verification. Experiments on the SYSU-MM01 and RegDB datasets demonstrate that DS-VSDA achieves stable performance in both indoor and outdoor scenarios, thereby validating its effectiveness.

In complex scenarios such as urban unsignalized intersections, autonomous vehicles interact intensively with heterogeneous agents. Synchronous trajectory prediction of heterogeneous agents is crucial for driving safety and efficiency. Existing methods neglect dynamic risk factors during interactions, resulting in insufficient prediction safety and accuracy. A heterogeneous agent trajectory prediction method based on safety-enhanced interaction modeling(SEIM) is proposed. Interaction risk, heterogeneous interaction characteristics, road environment, and individual dynamics are jointly considered. In the risk perception layer, interaction risks are characterized via the safety potential field and interaction risk features are extracted. In the heterogeneous interaction layer, two heterogeneous graphs are constructed and a heterogeneous interaction encoder is designed to capture heterogeneous interaction features. Safety potential field-based risk values are introduced as edge attributes of the safety- enhanced directed edge-feature heterogeneous graph to enhance edge features of the graph structure. In the road environment layer, road structure features are extracted from the directed graph. In the individual dynamic layer, individual motion features are captured. The four types of features are fused and fed into trajectory decoders to predict future trajectories of heterogeneous agents. Experiments on the inD dataset demonstrate the effectiveness and accuracy of SEIM in heterogeneous agent trajectory prediction.

Concept drift is a critical factor that limits the performance improvement of models in data stream mining tasks in open environments. Due to the continuous evolution of data distributions in data streams, the discrepancies between historical and current data distributions result in difficulties in fully exploiting historical knowledge, thereby undermining model stability and adaptability to drifting environments. To address this issue, a multi-stage perception guided adaptive ensemble for concept drift(MPAE) is proposed. Temporal profiling based on model performance over consecutive data blocks is performed, and the key historical data consistent with the current distribution are selected accordingly. Loss sequences and statistical significance testing are further combined to adaptively filter current data. Dynamic sample reconstruction is conducted based on temporal perception. In addition, a dynamic ensemble stopping criterion is designed based on the loss decay rate, allowing drift-perception adjustment of the ensemble process. According to the estimated drift intensity, an adaptive ensemble model is constructed to achieve effective coordination between the data and the model. Experiments demonstrate that MPAE exploits the historical data effectively, improves the drift adaptability of the model, and achieves favorable generalization performance.