k-Medoids Clustering algorithm is one of the most commonly used clustering methods. However, it does not consider the weight of each feature, and the selection of center points is random, which may affect the clustering results. Granularity computing analyzes knowledge from various perspectives, considering the level of detail and completeness of that knowledge. This approach helps identify which features have the most significant impact on clustering outcomes, allowing for a focus on key features and the selection of more effective initial clustering centers. Based on this, we propose an Effective Intelligent k-Medoids Clustering Scheme Based on Granularity Rough Entropy (IKCS). Specifically, the granularity roughness entropy is first defined based on indistinguishable relationships and roughness, and feature weights are determined based on the roughness entropy. In addition, feature weight coefficients are introduced to explain the similarity function between samples, and they are integrated into k-Medoids. In the similarity calculation of clustering algorithms, the samples are subjected to coarse-grained clustering based on the above similarity function to obtain a set of coarse clusters. Subsequently, the algorithm uses the maximum minimum distance method to select initial cluster centers from the coarse clustering set. Experimental results on both artificial and well-known UCI datasets effectively demonstrate the superiority of the proposed IKCS. On artificial datasets with complex cluster distributions, IKCS achieves an Adjusted Rand Index (ARI) of 0.7582, outperforming the second-best algorithm (fast k-Medoids) by 8.1%. Meanwhile, running on 12 UCI datasets with high dimensionality, IKCS achieves an average accuracy that is 13.89% higher than the PAM algorithm and 7.19% higher than the FKM (fast k-Medoids) algorithm. Moreover, it shows strong performance in other key metrics, consistently achieving a lower clustering Sum of Squared Errors (SSE) and a higher average ARI compared to the baseline methods.
Leakage Detection and Localization (LDL) is one of the greatest challenges to achieving sustainable management in Water Distribution Systems (WDS). As an essential element for the evaluation of LDL approaches, we recently developed LeakG3PD, a simulated dataset that represents WDS behavior in various leakage scenarios, providing all variables and network structure data. Unlike most works in the literature, this article explores the feasibility and advantages of using water demand Automated Meter Reading (AMR) devices, rather than pressure sensors, as the main data source for LDL. To evaluate such an approach, a derived dataset is generated from LeakG3PD using WDS inlet flows and pressures, along with, mainly, consumption demand data. As an introductory work, intuitive and transparent algorithms are applied for LDL. Calculated loss demand is used for detection based on historical thresholds and localization output consists of an ordered list of the pipes in which the insertion of the estimated leak demand causes minimum flow residuals in WDS sources (reservoirs and tanks). A one-step hydraulic simulation is performed to determine the residuals for each candidate pipe. The main difficulties are discussed to provide future improvements. Even though the fundamental step of model calibration is not executed, results indicate promising localization performance for future works. The derived dataset LeakG3PD-OD and the LDL Python programs are made available online for the academic community to contribute with more sophisticated algorithms.
Attraction-Repulsion Optimization Algorithm (AROA) is a recently proposed meta-heuristic algorithm known for its simplicity, ease of implementation, and robustness. However, AROA may converge to local optima when applied to complex optimization problems. To address this limitation, we propose an enhanced version called the Differential Cauchy Tangent Attraction-Repulsion Optimization Algorithm (DCTAROA). First, we propose a mutation operator based on a tangent flight mutation strategy and a dimension decision mechanism using the inverse cumulative distribution function of the Cauchy distribution. The tangent flight mutation enhances the local search capability and accelerates convergence, while the dimension decision strategy of the Cauchy distribution inverse cumulative function increases population diversity and improves exploration efficiency. Subsequently, we integrate Differential Evolution (DE) as a local search mechanism to strengthen the global optimization performance of AROA. To evaluate the proposed algorithm, we compare it with 15 state-of-the-art algorithms on 29 CEC2017 benchmark functions across various dimensions. Experimental results demonstrate that DCTAROA outperforms the compared algorithms in terms of solution accuracy, stability, convergence speed, and statistical significance based on the Wilcoxon rank-sum test. Furthermore, we apply DCTAROA to three practical engineering design problems. The results confirm that DCTAROA effectively explores the search space and yields competitive solutions, thereby validating its practical applicability.
Bayesian networks based on mixtures of polynomials (MoPs) are used to model multivariate hybrid or continuous probability distributions. MoPs provide a flexible yet simple model to deal with probabilistic reasoning, approximating complex or empirical probability distributions. We review the main models based on MoPs in the literature, introducing an alternative MoP approach: the mixtures of polynomials with tails (tMoPs). Also, we present procedures for learning tMoP marginal and conditional densities from data. This learning algorithms are tested with many well-known probability distributions. In these experiments, the tMoP models yield satisfactory results in comparison to other techniques, including other available MoPs alternatives. We also propose a meta-model that can directly, using interpolation, provide a tMoP expression for a specific density within a given family. This idea is tested with several probability distributions of one and two parameters, giving promising results.
Facial expression recognition (FER) is a challenging task that involves inferring human emotions from facial images entangled with various attributes. Although FER models based on Transformer have recently reported impressive performance, two major types of bias continue to degrade accuracy—namely, age-related facial attributes such as wrinkles and skin texture, which introduce confusion in recognizing elderly emotions, and imbalanced training data, which limits the model's ability to generalize across age groups. This paper proposes a novel bias-mitigation method that decouples and regularizes age- and emotion-related components within the self-attention mechanism of Transformer. The proposed method separately regularizes the value vectors that encode texture information correlated with age, and the query-key matrices that focus on facial landmarks crucial for emotion recognition. It encourages intra-class compactness of facial landmarks for emotion representation while minimizing age interference in features unrelated to emotion. To suppress age information and preserve emotional features, an age discriminator is employed to guide the value vectors in eliminating age cues, while an emotion classifier restores discriminative information for classification. In addition, the proposed method leverages triplet learning on the query-key space to enhance age-invariant emotion separation. Experiments conducted on four widely used FER benchmarks demonstrate that our method notably reduces age-related bias while maintaining or exceeding state-of-the-art performance.
