ificant. РOur micro-averaged F score (0.8567) for KNN is noticeably higher than the KNN score (0.823) reported by Joachims. We also tested the simplified KNN (following Joachims) which assign only the top-ranking category to each document, and obtained a F score of .8140. These contrasting tests suggest that this simplification is neither optimal nor necessary for KNN. РOur micro-averaged F score (0.7956) for NB is significantly higher than the NB score (0.720) obtained by Joachims. According to the analysis by McCallum at al. who implemented both models, the multinomial mixture model which we tested is better than the multivariate Bernoulli model which Joachims tested. So our experiment confirmed McCallum's conclusion. However, the better model of NB still underperforms SVM, KNN and LLSF.
全文预览