In current attribute reduction methods for multi-granularity fuzzy rough sets, the fuzzy similarity relationship within attribute sets is often defined using a fixed threshold. However, in fuzzy decision systems, different thresholds are required for different attributes, since each attribute has a different value range. To address this issue, the attribute reduction of optimistic multi-granularity fuzzy decision rough sets based on fuzzy self-information is proposed in this paper. First, within the fuzzy decision information system, the threshold for each attribute is introduced based on the quantile to construct an adaptive-threshold-based fuzzy similarity relation. An optimistic multi-granularity fuzzy decision rough set model for a family of attribute subsets is then established. Second, fuzzy self-information is proposed by utilizing the lower and upper approximations of optimistic multi-granularity fuzzy decision rough set model and boundary domain information. Furthermore, an attribute significance measure is introduced, and an attribute reduction algorithm based on fuzzy self-information is developed. The experimental results on ten datasets demonstrate that the proposed algorithm effectively reduces the dimensionality of the original attribute set and improves the classification accuracy.

Spectral clustering is widely used in the field of unsupervised learning due to its superiority in modeling pairwise similarity structures within data. However, conventional spectral clustering methods typically rely on the assumption of clean and structurally consistent datasets, and as a result their performance degrades significantly when the samples are noisy or mismatched in real-world applications. To address these issues, an extended spectral clustering framework, noise spectral clustering based on knowledge reuse(NSCR), is proposed. NSCR effectively leverages semantic knowledge from large-scale multi-model neural networks. A pseudo-label generation module is then designed by leveraging the prior capability of multi-modal neural network models in cross-modal understanding. Reliable samples are identified by semantic-consistency verification and saliency-aware confidence modeling. Softmax entropy is introduced as an uncertainty measure to filter pseudo-labels. Semantically consistent samples with low entropy across multiple models are selected and their pseudo-labels are generated to guide the clustering. Moreover, a joint optimization objective is employed to extend the traditional spectral clustering methods. A feature alignment term and a regularization balancing factor are utilized to mitigate semantic conflicts between pseudo-label supervision and clustering objective. Experiments on public datasets demonstrate that NSCR exhibits good robustness and generalization capability.

To improve the modeling capability of image classification networks for critical information and enhance the completeness and discriminability of feature representation, a dual wavelet-enhanced network for image classification(DWENet) is proposed. First, a wavelet-gated convolution module is constructed during the shallow feature extraction stage of the backbone network. By integrating wavelet frequency domain decomposition and a gating mechanism, edge and texture information are effectively captured with key details preserved and redundant noise suppressed. Second, a wavelet kernel attention module is introduced after the max-pooling layer. Frequency awareness and large receptive field spatial modeling are integrated to enhance structural representation and long-range dependency perception. As a result, the information loss caused by pooling operations is compensated. Furthermore, a dual-path feature enhancement module(DFEM) is incorporated into the deep network layers. DFEM enhances mid-and high-level semantic representation and key region responsiveness by reorganizing features through spatial-frequency decomposition and incorporating channel attention mechanisms. While maintaining computational efficiency, DWENet effectively mitigates core issues such as shallow feature decay and limited spatial perception. Experiments on CIFAR-10, CIFAR-100, SVHN, Imagenette, and Imagewoof datasets demonstrate that DWENet improves classification accuracy.

Most clustering methods mainly focus on single-view data, while the research on multi-view clustering remains relatively under-explored. Existing multi-view clustering methods often emphasize learning inter-view information while neglecting the thorough exploitation of intra-view information. In this paper, a contrastive collaborative multi-view attribute graph clustering based on adaptive structure enhancement(ACCMVC) is proposed. First, an adaptive structure enhancement strategy is designed to generate edge weights by combining node importance and the complex relationships among node features. These edge weights are applied to construct new adjacency matrices for the views, and thereby structure-enhanced graphs are generated. Second, edge weights are introduced into neighborhood contrastive learning. Intra-view enhanced neighborhood contrastive learning is applied to views and their structure-enhanced graphs, while inter-view enhanced neighborhood contrastive learning is utilized among multiple views. Finally, considering the varying view importance, an attention mechanism is introduced to calculate the weight of each view for effective fusion. Experiments on multiple datasets demonstrate that ACCMVC achieves superior clustering performance.

Constrained differential evolution(CDE) algorithm is an effective way to solve constrained optimization problems. However, the existing research mainly focuses on constraint handling techniques, while the differential evolution algorithm itself is neglected, resulting in some problems, including unbalanced exploration and exploitation capabilities, and a low survival rate of offspring individuals of feasible solutions. To address these issues, a dynamic elite learning strategy is designed to improve the performance of the CDE algorithm. In this strategy, the individuals in the population are divided into ordinary feasible solutions, elite feasible solutions and infeasible solutions, and individualized mutation operators are adopted for each of these three types of individuals to balance the exploration and exploitation capabilities. Meanwhile, the elite feasible solutions are introduced to improve the classical mutation operators, thereby increasing the survival rate of offspring of feasible solutions. According to the characteristics of infeasible solutions, a fine-tuned feasibility rule is also designed as a constraint handling technique to better guide the population into the feasible region. Experiments on CEC2006, CEC2010 and CEC2017 test sets as well as three real-world engineering optimization problems demonstrate that the proposed algorithm achieves superior performance compared with six state-of-art constrained optimization evolutionary algorithms.

Computerized adaptive testing(CAT) is designed to achieve efficient assessment through dynamic question selection. However, existing methods still suffer from insufficient accuracy in semantic modeling and ability estimation. To address these issues, a text-semantic enhanced graph neural network based CAT approach(TECAT) is proposed. Contextual semantic representations of questions and concepts are extracted using a pretrained language model, and multi-level dependencies among questions and concepts are captured by applying graph attention networks on the question-concept and concept-prerequisite graphs. To integrate semantic and structural information, a gated fusion mechanism based on additive attention and SiLU activation is introduced for the above two types of information to be adaptively combined. As a result, more expressive node representations are obtained. The CAT process is further formulated as a multi-objective reinforcement learning task to jointly optimize question quality, diversity, and novelty. A quality reward function based on changes in ability estimation error is designed to better reflect the contribution of each question to ability diagnosis. Experiments on two real-world datasets, Eedi and Junyi, show that TECAT achieves superior ability in estimation accuracy and concept representation quality compared with the existing methods.

Collaborative inference is an effective method for deploying models and accelerating inference on resource-constrained edge devices. However, the existing operator partitioning strategies still suffer from high inter-device communication overhead. To solve this problem, an interleaved operator partitioning(IOP) collaborative inference acceleration strategy for edge intelligence is proposed. The core mechanism is to partition adjacent operators along the input and output channel dimensions, respectively. By matching the number of channels between consecutive operators, the concatenation of output activations is reduced, and thereby the time overhead of collaborative inference is decreased. First, the computation and communication costs of devices are modeled based on operator information in the model. An integer programming model is established to minimize the total inference time. Second, a heuristic operator pairing algorithm is designed and adjacent operators are enumerated in a forward order. The inference time overhead of IOP and traditional output channel partitioning(OCP) is compared. The operator pair with the highest benefit is executed. Finally, interleaved partitioning and distributed deployment are applied to the selected operator pairs. Experiments demonstrate that IOP achieves superior performance in terms of inference time, memory usage, and energy consumption, while maintaining robustness under sudden link fluctuations.