Abstract

We have designed and implemented an unsupervised learning algorithm for a finite
mixture model of shifted-scaled Dirichlet distributions for the cluster analysis of multivariate
proportional data. The cluster analysis task involves model selection using Minimum
Message Length to discover the number of natural groupings a dataset is composed of.
Also, it involves an estimation step for the model parameters using the expectation maximization
framework. This thesis aims to improve the flexibility of the widely used Dirichlet
model by adding another set of parameters for the location (beside the scale parameter)
We have applied our estimation and model selection algorithm to synthetic generated
data, real data and software modules defect prediction. The experimental results show
the merits of the shifted scaled Dirichlet mixture model performance in comparison to
previously used generative models.