Remote sensing image classification (RSIC) plays an important role in interpreting surface features for applications such as urban planning, disaster monitoring, and agriculture. Despite advances in classification models, feature selection remains a key step that affects model performance. Traditional feature selection methods often involve high computational costs and do not explicitly account for uncertainty regions. This paper introduces TwGrey, a feature selection algorithm that integrates grey relational analysis (GRA) and three-way decision (TWD) theory. GRA is used to evaluate feature importance with low computational complexity, while TWD handles uncertainty by dividing features into positive, boundary, and negative regions. An adaptive threshold calculation method based on shadowed set theory is proposed to reduce the reliance on manually specified thresholds. Experiments on the NWPU and UCM datasets show that TwGrey achieves classification accuracy comparable to or higher than competing methods in most settings while selecting substantially fewer features. When applied post-hoc to frozen Triplet-256 embeddings, it yields consistent accuracy improvements over the full-dimensional baseline across all evaluated settings in our experiments. Notably, this is achieved while reducing the feature count by 38% – 79% , including on the full 45-class NWPU benchmark. Statistical tests and stability analyses further indicate that the observed improvements are statistically significant in several settings. Furthermore, while TwGrey incurs higher computational latency than non-iterative filter methods, it achieves favorable speedups over comparable three-way wrapper baselines.
Artificial neural networks have become one of the most extensively studied machine learning algorithms in the field of artificial intelligence in recent years, primarily due to their ability to simulate the biological nervous system, one of the most complex systems in nature. The increasing computational power of modern hardware and the improved understanding of the biological structure of the nervous system have enabled the development of algorithms capable of modeling this complexity more effectively. However, in conventional networks, the architecture is fixed in advance and only the weights are learned. While there are studies in which architectural components are optimized using evolutionary algorithms, existing approaches mostly focus on a single component or maintain a layered network structure. This study proposes a framework in which the network is represented as a distributed graph, and its connection topology, the aggregation and activation function of each neuron, and the connection weights are all determined by the Artificial Bee Colony (ABC) algorithm. The search is organized in two nested stages: an outer ABC selects the topology and the neuron functions, and, for each candidate structure, an inner ABC optimizes the weights of the active connections. The framework was evaluated on the Pima Indians Diabetes, Hepatitis and Heart Disease datasets over ten independent runs and compared, using Wilcoxon and Friedman tests, with two fixed-architecture baselines and three ablation variants under the same neuron budget and number of objective function evaluations. The results show that, on the Pima Indians Diabetes dataset, the proposed method achieved significantly higher F1-score (0.777 vs. 0.683 for backpropagation) and recall than the alternative approaches. On the Hepatitis and Heart Disease datasets, however, no statistically significant differences were observed between the methods. Moreover, the selected networks retained approximately half of the connections of the fully connected architecture without any loss in accuracy. These findings indicate that, despite requiring higher computational cost than gradient-based training, the joint optimization of network architecture and weights can provide competitive performance while producing more sparse network structures.
Handling imbalanced datasets remains a major difficulty in many machine learning tasks. In such cases, learning algorithms are typically biased toward the majority class, leading to skewed decision boundaries and insufficient representation of minority samples. The universum twin support vector machine (UTSVM) is also susceptible to this limitation, as it often produces models dominated by the majority class, causing minority samples to be incorrectly treated as noise and thereby reducing predictive accuracy. Moreover, UTSVM is sensitive to the presence of noise and outliers in the training data. To address these issues, and inspired by the integration of prior knowledge with intuitionistic fuzzy modeling, this study proposes a novel framework termed the Universum Relative Density Intuitionistic Fuzzy Kernelized Weighted Extreme Learning Machine (URDIFKWELM). The proposed method integrates Universum samples as prior information and applies an Relative Density Intuitionistic fuzzy membership mechanism to control the influence of uncertain, noisy, and irregular samples. The Universum data embed additional domain knowledge into the learning process, guiding the classifier toward more reliable decision boundaries. This approach enhances robustness to noise, improves the modeling of minority class instances, and ultimately strengthens the overall generalization performance of the model. Extensive experimental evaluations on real-world imbalanced datasets, conducted under both clean and Gaussian noise-corrupted settings, show that the proposed approach consistently achieves superior performance compared with several state-of-the-art methods.
Air quality forecasting is crucial for environmental monitoring and public health. While recent studies focus on short-term forecasting using long short-term memory (LSTM) or convolutional neural networks (CNN) models, long-term forecasting remains underexplored despite its value in supporting policy and resource planning. To address this gap, we propose a channel-independent model tailored for long-term air quality forecasting. The model performs patching on each univariate time series, uses dilated depthwise separable convolution to capture intra- and inter-patch dependencies, and introduces a sparse mixture of experts (MoE) to enhance representation and prediction. Experiments on eleven datasets show that our model achieves average improvements of 1.7, 2.1, 4.2, and 4.6
Granular computing theory offers a well-established framework for measuring data uncertainty. However, in practical applications involving multi-density data, disparities in data distribution significantly affect uncertainty measurements, often causing conventional neighborhood methods to yield substantial errors. To address this issue, this study proposes a novel uncertainty measurement and data reduction framework tailored for multi-density data from the perspective of variable granularity, establishing the theoretical foundation for Variable Granularity Rough Sets (VGRSs). For samples with high classification uncertainty, the VGRS model assigns an initial granule to each sample, identifies internal class-label conflicts, and dynamically adjusts the granule. Whether a sample belongs to the lower approximation of a decision class is then determined by its adjusted neighborhood granule. Building upon the VGRS model, feature selection algorithms are further developed. Comprehensive numerical experiments demonstrate the superior performance of VGRS-based algorithms, validating the feasibility, effectiveness, and generalization capability of the proposed model.
