This study proposes a multi-strategy improved Aquila optimizer (MIAO) to address the key limitations of the original Aquila optimizer (AO). First, a phasor operator is introduced to eliminate excessive control parameters in the X2 phase, transforming it into an adaptive parameter-free process. Second, a flow direction operator enhances the X3 phase by improving population diversity and local exploitation. The MIAO algorithm is applied to optimize Long Short-Term Memory (LSTM) hyperparameters, forming the MIAO_LSTM model for monthly railway freight forecasting. Comprehensive evaluations on 15 benchmark functions show MIAO’s superior performance over SOA, PSO, SSA, and AO. Using freight data (2005–2021), MIAO_LSTM achieves lower MAE, MSE, and RMSE compared to traditional LSTM and hybrid models (SSA_LSTM, PSO_LSTM, etc.). Further, Grey Relational Analysis selects high-correlation features (≥0.8) to boost accuracy. The results validate MIAO_LSTM’s effectiveness for practical freight predictions.

It is becoming a mainstream feature fusion approach for speech-visual emotion recognition (SVER) by directly using neural networks to fuse the extracted speech and visual features. However, the heterogeneity between speech and visual modalities usually results in a distribution gap and information redundancy between the extracted speech and visual features, thus affecting the performance of the SVER. To this end, this paper proposes a SVER method based on the modal decomposition learning. It leverages the shared, private and reconstructed modal learning with a specifically designed loss to decompose the extracted speech and visual features into the shared and private subspaces to obtain the shared and private features, which effectively reduces the distribution gap and information redundancy between the extracted speech and visual features. Experiments on the BAUM-1s, RAVDESS and eNTERFACE05 datasets also show that the proposed method achieves a better result.

Outlier mining constitutes an essential aspect of modern data analytics, focusing on the identification and interpretation of anomalous observations. Conventional density-based local outlier detection methodologies frequently exhibit limitations due to their inherent lack of data preprocessing capabilities, consequently demonstrating degraded performance when applied to novel or heterogeneous datasets. Moreover, the computation of the outlier factor for each sample in these algorithms results in considerably higher computational cost, especially in the case of large datasets. This paper introduces a local outlier detection method named FOLOF (FCM Objective Function-based LOF) through an examination of existing algorithms. The approach starts by applying the elbow rule to determine the optimal number of clusters in the dataset. Subsequently, the FCM objective function is employed to prune the dataset to extract a candidate set of outliers. Finally, a weighted local outlier factor detection algorithm computes the degree of anomaly for each sample in the candidate set. For the analysis, the Golden Section method was used to classify the outliers. The underlying causes of these outliers can be revealed by exploring the anomalous properties of each outlier data point through the outlier factors of each dimension property. This approach has been validated on artificial datasets, the UCI dataset, and an NBA player dataset to demonstrate its effectiveness.

Instance selection is an important preprocessing step in data mining and machine learning, particularly for large-scale datasets where instance redundancy and noise can degrade computational efficiency and generalization performance. Motivated by physical interaction principles, this paper proposes a Local Coulomb Force-based Instance Selection method (LCFIS). In the proposed framework, each data point is modeled as a charged particle, with its local density—estimated by combining reverse nearest neighbors and Gaussian kernel density estimation—serving as the charge magnitude. Neighborhood interactions are explicitly captured through the resultant Coulomb force acting on each point. We show that, as the neighborhood size increases, the rate of change of the resultant force differs significantly among boundary and core points. Utilizing this character, LCFIS effectively identifies boundary instances. A compact and representative training set is constructed by combining selected boundary points with randomly sampled instances to preserve data diversity. Extensive experiments on multiple public benchmark datasets using two standard classifiers demonstrate that LCFIS achieves a favorable trade-off between compression and classification performance, outperforming nine representative instance selection methods across accuracy, reduction rate, effectiveness, and computational cost.

Existing boundary point detection methods struggle to generalize across diverse data distributions due to their reliance on parameter sensitive, single-perspective (density or geometry) measures. To overcome the limitations of existing approaches—which often struggle to generalize across diverse data distributions and rely heavily on parameter‑ or threshold‑sensitive density‑ and geometry‑based measures—this paper introduces a fusing structural similarity and angular information for boundary point detection in diverse data (FSABD). For each data point, the algorithm first constructs a k-nearest neighbor local structure. It then defines a charge based on structural similarity and employs a Coulomb force model to simulate forces among neighboring points, thereby computing a structural boundary factor that incorporates both the magnitude and directional consistency of the resultant force. Next, depending on the dimensionality of the data, the algorithm applies one of two angular‑analysis strategies to derive an angular boundary score that captures the concentration characteristics of the neighbors’ angular distribution. The structural and angular components are subsequently fused through an adaptive weighting mechanism to form a comprehensive boundary factor. A threshold is automatically determined from the mean and standard deviation of this factor, and points exceeding the threshold are identified as boundary points. On 20 diverse benchmark datasets (UCI, image, text, scRNA-seq), FSABD consistently outperforms nine state-of-the-art methods. When integrated with k-means clustering, it improves average accuracy (ACC) by 10.09%, Adjusted Rand Index (ARI) by 11.32%, and Normalized Mutual Information (NMI) by 7.26%. In 3D point cloud hole detection, FSABD achieves >99% accuracy, demonstrating strong cross-task generalization.