Improved model-based clustering performance using Bayesian initialization averaging

TL;DR

This paper introduces Bayesian initialization averaging (BIA), a novel ensemble method that improves the starting point quality for EM and variational Bayes algorithms, leading to better clustering results across various data types.

Contribution

The paper proposes BIA as a new ensemble initialization technique for EM and VB algorithms, enhancing convergence to higher likelihood solutions in mixture modeling.

Findings

BIA outperforms traditional initialization methods in likelihood maximization.

BIA achieves higher-quality clustering solutions across diverse data sets.

The method is computationally efficient and adaptable to different models.

Abstract

The Expectation-Maximization (EM) algorithm is a commonly used method for finding the maximum likelihood estimates of the parameters in a mixture model via coordinate ascent. A serious pitfall with the algorithm is that in the case of multimodal likelihood functions, it can get trapped at a local maximum. This problem often occurs when sub-optimal starting values are used to initialize the algorithm. Bayesian initialization averaging (BIA) is proposed as an ensemble method to generate high quality starting values for the EM algorithm. Competing sets of trial starting values are combined as a weighted average, which is then used as the starting position for a full EM run. The method can also be extended to variational Bayes (VB) methods, a class of algorithm similar to EM that is based on an approximation of the model posterior. The BIA method is demonstrated on real continuous, categorical and network data sets, and the convergent log-likelihoods and associated clustering solutions presented. These compare favorably with the output produced using competing initialization methods such as random starts, hierarchical clustering and deterministic annealing, with the highest available maximum likelihood estimates obtained in a higher percentage of cases, at reasonable computational cost. The implications of the different clustering solutions obtained by local maxima are also discussed.