Non-Metric Similarity Measures
Abstract
This was an extension of mass theory to non-metric similarity measures. The non-metric similarity measures were created as a generalization of mass estimation from a unary function to a binary function. Unlike the traditional similarity measure that is based on geometric difference between two instances, the mass-based measure takes data distribution into account. It is demonstrated that the new measure results in better performance in applying to information retrieval task. A derivative of mass measure called relative mass was also investigated using three implementations. The research in relative mass was expanded to two tasks: In anomaly detection relative mass is used to overcome one weakness of current mass-based anomaly detectors using a tree-based approach and a nearest-neighbor-based approach; in clustering, relative mass is used to recondition density-based clustering algorithms to successfully find clusters with varying densities.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 26, 2015
- Accession Number
- ADA622081
Entities
People
- Kai M. Ting