A randomized framework for discovery of heterogeneous mixtures

Mixture models are the term given to models that consist of a combination of independent functions creating the distribution of points within a set. We present a framework for automatically discovering and evaluating candidate models within unstructured data. Our abstraction of models enables us to seamlessly consider different types of functions as equally possible candidates. Our framework does not require an estimate of the number of underlying models, allows points to be probabilistically classified into multiple models or identified as outliers, and includes a few parameters that an analyst (not typically an expert in statistical methods) may use to adjust the output of the algorithm. We give results from our framework with synthetic data and classic data.