Time Series Clustering of Rice Productivity Using Trimming Gaussian Mixture Models

Abstract

This study investigates the application of the Trimming Gaussian Mixture Model (TGMM) for clustering monthly rice productivity time series data in West Java from 2018 to 2023. TGMM is a robust clustering approach that reduces the influence of outliers by trimming a specified portion of the data prior to parameter estimation. The dataset, sourced from Open Data Jabar, was analyzed to identify the most representative number of clusters using the Silhouette Score. The optimal clustering solution was achieved with two main clusters (k = 2) and a trimming proportion of 15%. The results revealed three distinct regional groups: two dominant clusters characterized by moderate-stable and high-consistent productivity patterns, and a separate group of outliers marked by low and highly fluctuating productivity. Cluster stability was assessed using the Adjusted Rand Index (ARI), yielding values of 0.41 (bootstrap) and 0.545 (subsampling), which indicate a reasonably consistent clustering structure. These findings demonstrate the effectiveness of TGMM in capturing underlying productivity patterns while accounting for noise and outliers, suggesting its potential as a robust decision-support tool for data-driven agricultural planning and policy formulation.