Data classification is a fundamental task in machine learning and data analysis, with applications across many fields. In practice, uncertainty often arises due to weakly discriminative features, missing values, distribution shift, class imbalance, inconsistent label spaces, and class overlap. Traditional classification methods, which typically rely on probabilistic frameworks, may not explicitly represent or reduce such uncertainty. Evidence theory (ET), also known as Dempster-Shafer theory, provides a flexible framework for modeling uncertainty and imprecision through basic belief assignments and the evidence combination rule, and has attracted growing attention in the data classification field. Consequently, data classification based on ET (DCET) has become an active research topic. This paper provides a comprehensive survey of DCET, systematically categorizing DCET methods by feature- and label-based uncertainty scenarios. We first review ET fundamentals and summarize the evidential K-nearest-neighbor classifier and its variants. We then survey ET-based methods for tabular data classification under six major uncertainty scenarios: high-dimensional features, missing attribute values, feature distribution shift, imbalanced label distribution, inconsistent label spaces, and class overlap. Next, we review ET-based methods for image (unstructured) classification, highlighting the integration of ET with deep neural networks. We further summarize representative applications, including human activity recognition, medical image segmentation, remote sensing image classification, and remote sensing image change detection. Finally, we discuss future research directions for DCET.
Multimodal recommendation aims to enrich preference modeling by leveraging visual and textual features. However, integrating high-dimensional pretrained features introduces substantial computational overhead. While knowledge distillation provides an effective compression strategy, existing frameworks face three intertwined challenges: rank bottlenecks caused by low-dimensional projections, cross-modal interference induced by shared fusion spaces, and optimization instability under static distillation temperatures. To address these issues, we propose ProMoE-DTS, a decoupled prompt-guided mixture-of-experts framework with dynamic temperature scheduling. Using an asymmetric teacher-student architecture, the teacher model leverages modality-aware soft prompts as semantic anchors to route heterogeneous features into parameter-disjoint expert networks, thereby alleviating cross-modal conflicts and resolving the rank bottlenecks. To ensure stable knowledge transfer, a feedback-driven dynamic temperature scheduler adaptively regulates the distillation intensity based on epoch-wise signals. This asymmetric design confines intensive multimodal operations to the offline teacher, leaving the online student model with a highly efficient, pure identifier-based structure. Extensive experiments on three benchmark datasets demonstrate that ProMoE-DTS improves Recall@20 by 2.24%-3.96% over state-of-the-art baselines, while requiring only 3.28%-3.55% of the teacher's parameters.
Precise depth data derived from light detection and ranging (LiDAR) have been shown to be effective in improving gait recognition performance. Those depth data, however, are coarse and lack clear segmentation boundaries, which often results in increased ambiguity in interpreting precise human motion and body shape, potentially limiting the model's ability to learn discriminative gait-related features. To address this challenge, we propose a novel silhouette-guided motion-augmentation module for depth-based gait recognition, namely, SigmaGait. SigmaGait leverages the clear boundaries and homogeneous foregrounds of silhouettes to enhance the model's awareness of part-level motions and fine-grained appearance features. Furthermore, we explore a pseudo-depth generation approach that leverages accessible red-green-blue (RGB) data for scenarios where LiDAR sensors are unavailable. By employing off-the-shelf skinned multi-person linear (SMPL) models, we synthesize pseudo-depth maps from estimated human meshes, bridging the performance gap between camera and LiDAR-reliant gait recognition. We empirically evaluate our approach on real and synthetic datasets including cloth-changing benchmark for person re-identification and gait recognition (CCPG), SUSTech1K, and FreeGait. Our model achieves state-of-the-art performance on real LiDAR data, and our pseudo-depth maps improve accuracy on 2D datasets despite their synthetic origin.
Most existing point cloud-based place recognition methods emphasize feature extraction from individual points or local regions, while largely neglecting the relational information embedded within local neighborhoods. As a result, they often fail to capture discriminative relational patterns, leading to reduced recognition accuracy in scenes containing geometrically similar structures. In this paper, we propose a novel relation-aware network (RA-Net) for place recognition. RA-Net jointly exploits local relational cues and global contextual information to learn discriminative scene representations for large-scale point cloud-based place recognition. First, a spatial relation feature extraction (SRFE) module is proposed to exploit relational information embedded within local neighborhoods. By learning relation-aware weights and adaptively aggregating neighborhood information, the proposed module captures discriminative relational patterns by jointly considering feature discrepancies and spatial offsets. Furthermore, a global feature extraction (GFE) module is introduced to aggregate global feature statistics and integrate them with pointwise representations, enabling local features to be enhanced with global contextual cues. Experimental results on four benchmark datasets demonstrate that RA-Net can generate more discriminative global descriptors and achieve promising performance. It exhibits strong generalization capabilities for unseen scenes.
Social media is an essential data source for risk perception of complex social events. However, existing early warning methods mainly rely on popularity peaks, task-specific semantic labels, or observable propagation structures, often ignoring propagation rhythm changes prior to a crisis. This paper proposes an unsupervised online framework for social media crisis early warning. Aggregating heterogeneous behaviors into a unified intensity sequence, it detects anomalously accelerated propagation phases relative to historical baselines through causal smoothing, rate-of-change extraction, rolling quantile thresholds, persistence constraints, and cooling mechanisms. Compared to intensity-driven methods, it focuses on “whether the system is accelerating toward a risk state” rather than “whether it has reached a high popularity state;” compared to semantic- and structure-driven methods, it has weaker dependencies on annotations, ontologies, and complete propagation graphs. Strict causal replay experiments on the Russia-Ukraine crisis Weibo dataset and the Douban Movie Short Comments dataset demonstrate that the framework captures aggregative rising processes before key events early with low false alarm interference, exhibiting structural reusability across scenarios. Results indicate that detecting propagation dynamics changes offers a lightweight, interpretable, and transferable path for online early warning in complex systems.
Robotic manipulators often suffer from insufficient positioning accuracy due to dynamic uncertainties, while existing control algorithms struggle to simultaneously achieve fast response and high tracking precision and often rely on complex dynamic models that demand excessive computational power, hindering practical onboard deployment. To address these issues, we propose an advanced real-time capable onboard tracking controller for robotic manipulators with dynamic uncertainties, which can balance the overall control performance, computational complexity, and parameter setting. The controller is suitable for direct implementation on embedded hardware with limited computational resources, without requiring heavy dynamics computation or learning approximation. First, a model-free adaptive time-delay estimator is designed to estimate and compensate for the lumped dynamic uncertainties of the robotic manipulator, where the gain matrix is directly updated by the magnitude of the sliding mode variable of tracking error for excellent overall performance. Then, an adaptive fixed-time nonsingular terminal sliding mode controller with prescribed performance is developed to stabilize the tracking errors of the robotic manipulator in a predefined fixed time. The controller parameters determine an upper bound of the fixed convergence time through analytically derived expressions, where the actual settling time is automatically ensured regardless of the initial conditions. The fast response and accurate steady-state tracking are also guaranteed by the prescribed performance, such that the processing quality and efficiency of the robotic manipulator during real-time operation tasks can be improved. Moreover, the adaptive updating law for the tracking controller gain is designed to achieve fixed-time stability of the entire closed-loop states without the requirement of an exact upper bound of dynamic uncertainties or estimation errors. The proposed tracking control scheme for robotic manipulators is easy to implement owing to the convenient parameter setting and low computational burden, and its effectiveness is verified by simulations and experiments.
With the widespread application of UAVs, ensuring their safety and reliability is increasingly crucial. Data-driven anomaly detection methods demonstrate outstanding performance in unmanned aerial vehicle (UAV) health monitoring. Previous studies extract spatiotemporal features from high-dimensional flight data, but emphasize intricate multidimensional correlations, neglecting the intrinsic characteristics of single-dimensional data. This oversight leads to insufficient supervision of intrinsic univariate information. Consequently, subtle deviations in individual flight variables may be masked by dominant multidimensionality, limiting the model's ability to distinguish fine-grained anomalies from normal flight states. To address this problem, we propose F-UAD, a novel dynamic trend perception framework for fine-grained UAV anomaly detection. F-UAD is designed to coordinate univariate feature preservation, multidimensional coupling extraction, and dynamic trend perception, enabling fine-grained anomaly detection of subtle deviations in UAV flight data. During training, the dynamic trend perception module incorporates (1) multiscale spatiotemporal feature reconstruction to balance independent univariate feature preservation with comprehensive multidimensional coupling extraction and (2) dynamic trend forecasting based on decomposed flight trends to improve the perception of complex temporal evolution. During testing, reconstruction consistency and dynamic trend consistency are jointly used to compute anomaly scores, thereby enhancing the separation between normal and abnormal flight states. Experiments demonstrate that our F-UAD framework outperforms state-of-the-art methods across six AirLab failure and anomaly (ALFA) real flight dataset subsets, improving UAV reliability.
Stacked intelligent metasurfaces (SIMs) have recently emerged as a promising hardware architecture for integrated sensing and communication (ISAC), owing to their capability of performing low-power wave-domain signal processing. This paper investigates an SIM-aided multi-static ISAC system, where multiple SIM-enabled base stations (BSs) are coordinated by a central processing unit (CPU) to support downlink communication and target sensing. In the considered architecture, the BSs are configured as either transmitting or receiving nodes, enabling joint user service, target illumination, and cooperative echo processing. To characterize the interplay between communication and sensing, we formulate a joint design problem that maximizes the communication sum rate subject to target-specific sensing quality constraints, transmit power constraints, and SIM hardware constraints. The resulting non-convex problem involves the coupled optimization of transmit power allocation, transmit and receive SIM phase shifts, and centralized sensing receive beamformers. To solve this problem, we develop an efficient alternating optimization algorithm by combining weighted minimum mean-square error reformulation, generalized Rayleigh quotient maximization, and Riemannian manifold optimization. Numerical results demonstrate that the proposed design achieves a favorable communication-sensing trade-off and outperforms representative benchmark schemes.
We propose the expected graph Hilbert transform (EGHT), a robust extension of the Hilbert transform, for graph signals under topological uncertainty. By modeling the shift operator as a random variable and using its expected value as a stable spectral reference, the EGHT applies a π/2 phase shift to non-direct current (DC) graph Fourier modes, enabling the construction of analytic signals on uncertain graphs. The definitions of EGHT for undirected and directed graph ensembles are given, along with key properties such as linearity, anti-involution, and orthogonality; moreover, the expected graph analytic signal (EGAS) for node-wise modulation analysis is introduced. Experiments show that the EGHT outperforms conventional graph Hilbert transforms (GHTs) on noisy realizations, recovering coherent oscillations, producing smooth phase fields, and resisting topological perturbations. As such, it enables reliable time-frequency analysis in stochastic networks.