Locally adaptive metric nearest-neighbor classification

• DOI: 10.1109/tpami.2002.1033219
• Publication Date: 2002-10-18T21:39:07Z
• Authors: C. Domeniconi; Jing Peng; D. Gunopulos
• Keywords: Nearest neighbor graph; Best bin first; Feature (linguistics); Nearest-neighbor chain algorithm

Nearest-neighbor classification assumes locally constant class conditional probabilities. This assumption becomes invalid in high dimensions with finite samples due to the curse of dimensionality. Severe bias can be introduced under these conditions when using nearest-neighbor rule. We propose a adaptive method try minimize bias. use chi-squared distance analysis compute flexible metric for producing neighborhoods that are highly query locations. Neighborhoods elongated along less relevant feature and constricted most influential ones. As result, probabilities smoother modified neighborhoods, whereby better performance achieved. The efficacy our is validated compared against other techniques both simulated real-world data.