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.