Application of Dimensionality Reduction Technique Based on Fast Dynamic Clustering Idea

This study proposes a fast and dynamically convergent dimensionality reduction method to enhance prediction accuracy. By processing and decomposing multidimensional data, we capture the seasonality, time trends, and detailed changes of sorting center shipments. Using K-means clustering, the data are categorized into four classes, with three metrics introduced to mitigate short-term shocks. Five accuracy metrics are used to select the optimal prediction model for each class. Sensitivity analysis of cluster numbers and algorithms verified the stability of K-means clustering results. Temporal variations in profile coefficients, intra-cluster variance, core parameter configurations, and noise amounts demonstrated that multi-period clustering stabilizes over time. The fast and dynamically convergent clustering method proposed in this study effectively solves the time-series data prediction problem in the presence of a large number of individuals with significant heterogeneity, and provides a data-driven cost-saving and open-source strategy for small and medium-sized enterprises (SMEs) with limited budgets in the era of big data, which has a wide range of application potentials.