Effective construction of compression-based feature space
Hisashi Koga, Yuji Nakajima, Takahisa Toda · International Symposium on Information Theory and its Applications · 2016
This paper investigates how to construct a feature space for compression-based pattern recognition which judges the similarity between two objects x and y through the compression ratio to compress x with y('s dictionary). Specifically, we focus on the known framework called PRDC which represents an object x as a compression-ratio vector (CV) that lines up the compression ratios after x is compressed with multiple different dictionaries. For PRDC, the dimensions, i.e., the dictionaries determine the quality of CV space. This paper presents a practical technique to modify the chosen dictionaries which improves the performance of pattern recognition substantially by making them more independent.