Model Dependency of Performance in Generative Local Metric Learning

[ Information ]

한국정보과학회 학술발표논문집, vol. 38, no. 2B, pp. 307-310, 2011.

[ Authors ]

Yung-Kyun Noh, Frank C. Park, and Daniel D. Lee

[ Abstract ]

Nearest  neighbor  classification  with  generative  local  metric  (GLM)  learning  is  a  hybrid  method  of  the  discriminative and  generative  approaches.  A  discriminative  nearest  neighbor  classifier  does  not  consider  any  model  of  data,  while  a generative classifier unavoidably adopts a particular form of the probability density function. In this work, we illuminate how these discriminative and generative approaches have different advantages and show how the advantages of both can be  resolved  into  a  GLM  method.  We  present  various  examples  that  clearly  show the different  regimes  where the discriminative and generative  approaches should  outperform each  other.  In  these  examples,  we  show  that  the  GLM  is robust to the usage of an incorrect model, enhancing nearest neighbor classifier performance even when the model is not exact.