Unsupervised Learning of Finite Mixtures using Scaled Dirichlet Distribution and its Application to Software Modules categorization

We have designed and implemented an unsupervised learning algorithm for finite mixture model using the scaled Dirichlet distribution for multivariate proportional data. In this paper, the task of learning finite mixture model involves estimation of model parameters as well as inferring the hidden class information of our observed data. We made use of the expectation maximization algorithm to find the maximum likelihood estimate of our model parameters. This work, aims to address the flexibility challenge of the Dirichlet distribution by introducing a distribution that adds to it a scale parameter. This is important, because there is growing need for models that can fully describe the intrinsic nature of datasets. In addition, we applied our learning algorithm to synthetic datasets as well as to address the challenge of detecting fault prone software modules. Our proposed algorithm, makes it possible to discover these fault prone modules by harnessing their complexity-based attribute information. Finally, we compare our proposed model classification results with those from the Gaussian and Dirichlet mixture models.