我們研究了高維資料中聚類變數的選取問題。為此，我們將原有的柴比雪夫貪婪演算法（CGA）改良為更具貪婪性的 ggCGA 演算法，大幅提升了在實務資料處理上的效能與抓取高度相關變數的能力。基於 ggCGA 下，我們進一步提出了計算效率高的演算法 UL-MLR 來逐步選取變數，並結合了高維資訊準則（HDIC）與修剪機制來強化變數選取的效果。在模擬情境下，UL-MLR 具有優異的表現，相較於其他聚類變數選取方法擁有更高的計算效率。最後，我們將 UL-MLR 應用於四筆單細胞測序數據以及氣象資料，最終結果顯示 UL-MLR 選出的聚類變數能夠有效提升整體的聚類效果。;This study investigates the problem of selecting clustering variables in high-dimensional data. To this end， we modify the original Chebyshev Greedy Algorithm (CGA) into a more aggressive version， termed the ggCGA algorithm， which significantly enhances practical data processing efficiency and the ability to capture highly correlated variables. Building upon the ggCGA framework， we further propose UL-MLR， a computationally efficient algorithm for stepwise variable selection. This method integrates the High-Dimensional Information Criterion (HDIC) and a pruning mechanism to reinforce the effectiveness of the selection process. In simulation studies， UL-MLR demonstrates superior performance and higher computational efficiency compared to alternative clustering variable selection methods. Finally， we apply UL-MLR to four single-cell sequencing datasets as well as meteorological data. The results indicate that the clustering variables selected by UL-MLR effectively improve overall clustering performance.
