Normalized maximum likelihood methods for clustering and density estimation
Panu Luosto · Työväentutkimus Vuosikirja · 2013
The normalized maximum likelihood (NML) distribution has an important position in minimum description length based modelling. Given a set of possible models, the corresponding NML distribution enables optimal encoding according to the worst-case criterion. However, many model classes of practical interest do not have an NML distribution. This thesis introduces solutions for a selection of such cases, including for example one-dimensional normal, uniform and exponential model classes with unrestricted parameters. The new code length functions are based on minimal assumptions about the data, because an approach that would be completely free of any assumptions is not possible in these cases. We also use the new techniques in clustering, as well as in density and entropy estimation applications.