Comparison of performance between model-based and K-means clustering for reliability analysis: a real-life application

Model-based clustering and K-means clustering are widely used in various data clustering fields. This paper compares the performance of model-based clustering and K-means clustering in reliability analysis using automobile component warranty claims data. The Weibull and lognormal mixture models are applied to model the usage rate variable. The Expectation-Maximisation (EM) algorithm is employed to obtain the maximum likelihood estimates of the parameters for the mixture models. It also obtains the nonparametric estimates of the reliability function of the usage rate. The 5-fold Weibull mixture model achieves 79.2% accuracy, outperforming K-means clustering (67.6% accuracy) and the 5-fold lognormal mixture model (54.7% accuracy). Simulation studies confirm the applicability of the method and the superiority of model-based clustering, particularly the Weibull mixture model. The findings will have managerial implications for accurately assessing and predicting the component's reliability, and offering a flexible two-dimensional warranty policy, which can enhance customer satisfaction and the company's reputation.