Data exploration with learning metrics

A crucial problem in exploratory analysis of data is that it is difficult for computational methods to focus on interesting aspects of data. Traditional methods of unsupervised learning cannot differentiate between interesting and noninteresting variation, and hence may model, visualize, or cluster parts of data that are not interesting to the analyst. This wastes the computational power of the methods and may mislead the analyst. In this thesis, a principle called “learning metrics” is used to develop visualization and clustering methods that automatically focus on the interesting aspects, based on auxiliary labels supplied with the data samples. The principle yields non-Euclidean (Riemannian) metrics that are data-driven, widely applicable, versatile, invariant to many transformations, and in part invariant to noise. Learning metric methods are introduced for five tasks: nonlinear visualization by Self-Organizing Maps and Multidimensional Scaling, linear projection, and clustering of discrete data and multinomial distributions. The resulting methods either explicitly estimate distances in the Riemannian metric, or optimize a tailored cost function which is implicitly related to such a metric. The methods have rigorous theoretical relationships to information geometry and probabilistic modeling, and are empirically shown to yield good practical results in exploratory and information retrieval tasks.