DOI:
Abstract:
Many problems in machine learning and data mining depend on distance estimates between observations, e.g., instance-based classification, clustering, information retrieval, and record linkage in databases. However, the appropriate notion of similarity can vary depending on the particular domain, dataset, or task at hand. Consequently, a large number of functions that compute similarity between objects have been developed for different data types, varying greatly in their expressiveness, mathematical properties, and assumptions. Additionally, there exists a substantial body of research on feature space transformations that attempt to provide a more salient representation of data than the original feature space, e.g. Principal Component Analysis and Locally Linear Embedding. All of these techniques make certain assumptions about the optimal representation of data and its influence on computing similarity which may or may not be applicable for specific datasets and tasks. Therefore, it is desirable to learn accurate similarity functions from training data to capture the correct notion of distance for a particular task at hand in a given domain. Recently, several approaches have been suggested for training such functions using pairwise relations between instances, e.g. pairwise equivalence, common cluster membership, and relative comparisons. These approaches have shown improvements over traditional similarity functions for different data types such as vectors in Euclidean space, strings, and database records composed of multiple text fields. While these initial results are encouraging, there still remains a large number of similarity functions that are currently unable to adapt to a particular domain. In our research, we attempt to bridge this gap by developing both new learnable similarity functions and methods for their application to particular problems in machine learning and data mining.