Hierarchical feature selection is a specialized dimensionality reduction technique for hierarchical data. However, most existing hierarchical feature selection methods handle only hierarchical data with complete label information. In reality, semi-supervised hierarchical data (hierarchical data with missing labels) is pervasive due to collection and labeling costs. Therefore, this paper proposes a hierarchical feature selection method for semi-supervised hierarchical data based on label replenishment and fuzzy synergistic discrimination mutual information. Firstly, we employ a local structure strategy to acquire the topological relationships of label-missing samples, and converting the hard-label representation of semi-supervised data into a soft label representation. Subsequently, we utilize the hierarchical structure to compute a label similarity matrix to achieve label replenishment. Secondly, addressing the internal node importance difference caused by varying tree structures and sample distributions, we propose a novel method for internal node weight calculation. Thirdly, we define an importance metric based on hierarchical structure and fuzzy synergistic discrimination mutual information, and construct a feature selection method for semi-supervised hierarchical data. Finally, experimental results on various datasets show that our method can effectively perform feature selection on semi-supervised hierarchical data.
Electroencephalogram (EEG)-based emotion recognition is a cornerstone of affective computing and human computer interaction. However, its advancement is hindered by the intrinsic fuzziness of emotions and the complex nature of EEGsignals. To address these challenges, we propose Orthogonal Fuzzy Power-Coherence Alignment (OFPCA), a novel framework that enhances emotion-brain representation learning through three key innovations. First, OFPCA employs a Takagi-Sugeno Kang (TSK) fuzzy system to explicitly model the fuzzy nature of emotions, enabling robust feature extraction from EEG signals. Second, it integrates two complementary EEG perspectives—the local-energy view (captured via power spectral density, PSD) and the cross-regional interaction view (characterized by coherence, COH)-using contrastive learning to align their semantic representations in a unified fuzzy space. Third, we introduce an orthogonal regularization loss to minimize feature redundancy and stabilize training, further improving representation diversity. Evaluated on two benchmark datasets, SEED (three-class emotions) and CRED (five-class emotions), OFPCA outperforms existing state-of-the-art methods. Under the cross-subject setting, OFPCA achieves accuracies of 41.14% on CRED and 83.24% on SEED. Under the intra-subject setting, it reaches accuracies of 63.24% on CRED and 85.37% on SEED. These results validate that explicitly modeling emotional fuzziness and aligning multi view EEG features significantly enhance recognition accuracy. Our work establishes OFPCA as a pioneering fuzzy learning framework for multi-view EEG analysis, offering new insights into emotion-brain mapping.
Partial observability and intermittent measurements limit the reliability of point-estimate feedback for constrained stochastic nonlinear systems because estimation uncertainty and constraint-violation risk are not explicitly propagated into control decisions. This paper proposes a belief- and risk-driven interval type-2 neuro-fuzzy (BR-IT2NF) output-feedback framework. A causal Gaussian belief, constructed by fusing recurrent-predictive and model-based extended Kalman filter (EKF) moment estimates, provides the state estimate and covariance, from which Gaussian model risk is evaluated. The central fuzzy mechanism maps belief spread and risk-threshold exceedance into the footprint of uncertainty (FOU), adaptively expanding the fuzzy-output set while exactly recovering the Type-1 limit when both drivers vanish. An anchored selector blends the set projection of a nominal control with a risk-directed endpoint of the exact Karnik–Mendel (KM) output interval, followed by projection onto the admissible input set. We establish FOU validity and a linear bound from FOU activity to fuzzy-output-set diameter, and derive conditional finite- and infinite-horizon practical mean-square tracking bounds. Paired Monte Carlo studies on stochastic continuous stirred-tank reactor (CSTR) and Duffing systems show that BR-IT2NF achieves the lowest mean root-mean-square tracking error (RMSE) across all tested long-outage durations. Matched ablations and uncertainty-propagation studies further substantiate the proposed belief-to-FOU-to-control mechanism.
For unsupervised outlier detection, angle-based methods usually characterize local geometric differences through angular relations constructed from relative position vectors between samples. However, conventional methods mostly perform angular calculation at the original sample level, which leads to high computational cost. Moreover, when cluster boundaries are ambiguous or local neighborhood representations are unstable, the discriminative ability of angular measures tends to degrade. To address these issues, this paper proposes a Granular-Ball Angle-Based Outlier Detection with Uncertainty-Aware Angular Weighting (GABOD) method. GABOD first uses granular balls to compactly partition the sample space, and employs granular ball centers to represent local sample regions for angular modeling. Then, it constructs center-based k-nearest-neighbor (ckNN) neighboring-ball sets based on the distances between granular-ball centers, and evaluates the outlier tendency of each granular ball through the weighted angular variance among neighboring center vectors. Furthermore, an approximation error upper bound is introduced by transferring sample-level angular relations to granular-ball center angular relations. Based on intra-ball dispersion and inter-ball separation, an uncertainty aware angular weight is designed to reduce the angular estimation bias caused by center-based representation and boundary mixing. Finally, the granular-ball-level outlier scores are mapped back to the sample level to obtain the sample outlier ranking. Experimental results show that GABOD achieves competitive detection performance on various benchmark datasets.
This work develops a boundary-aware sliding prescribed performance control (PPC) scheme for uncertain nonlinear systems in the presence of actuator saturation and external disturbances. In contrast to traditional PPC approaches, the proposed method integrates a sliding function to eliminate the initial feasibility requirement, enabling the performance boundary to adaptively adjust to varying initial errors. Moreover, this study incorporates an error-driven mechanism (EDM) within the PPC architecture. The EDM's straightforward logic enables the adaptive relaxation of performance boundaries without requiring the implementation of auxiliary systems. A key characteristic of this approach is that boundary adjustment is not initiated immediately upon the detection of input saturation or external disturbances. Instead, relaxation is conditionally enacted only when the tracking error exceeds a predefined EDM threshold, and it is applied solely to the pertinent unilateral boundary. These distinctive attributes collectively address the susceptibility of traditional PPC to input saturation and disturbances, while also substantially reducing the performance loss typically associated with current flexible PPC methods. Simulation studies verify the proposed strategy's effectiveness and demonstrate its superior performance.
