Clustering of longitudinal curves via a penalized method and EM algorithm

TL;DR

This paper introduces FWP, a novel clustering method for longitudinal curves that combines penalized pairwise fusion with EM and ADMM algorithms, effectively incorporating prior neighborhood information for improved group detection.

Contribution

The paper presents a new penalized clustering approach for longitudinal data that integrates spatial weights and prior information, enhancing clustering accuracy and interpretability.

Findings

FWP outperforms existing methods in estimating the number of subgroups.

Incorporating spatial weights improves clustering performance.

Ignoring covariance structure reduces accuracy.

Abstract

In this article, a new method, called FWP, is proposed for clustering longitudinal curves. In the proposed method, clusters of mean functions are identified through a weighted concave pairwise fusion method. The EM algorithm and the alternating direction method of multiplier algorithm are combined to estimate the group structure, mean functions and the principal components simultaneously. The proposed method also allows to incorporate the prior neighborhood information to have more meaningful groups by adding pairwise weights in the pairwise penalties. In the simulation study, the performance of the proposed method is compared to some existing clustering methods in terms of the accuracy for estimating the number of subgroups and mean functions. The results suggest that ignoring covariance structure will have a great effect on the performance of estimating the number of groups and estimating accuracy. The effect of including pairwise weights is also explored in a spatial lattice setting to take consideration of the spatial information. The results show that incorporating spatial weights will improve the performance. A real example is used to illustrate the proposed method.