Comparison Study on Multi Logit and Stepwise Classification

Abstract

Multi class classification is an important topic in real world problems. The most popular strategy in doing multi class classification is classifying all at once based on posterior probability or distance metric. Discriminant analysis, k-nearest neighbors, neural network and multi logit regression are belong to this strategy. Friedman(1996) suggested the a new intuitive approach for the multi class problems: solve each of the two-class problems and combine all the results of pairwise decisions to form a multiclass classifier. Trevor Hastie and Robert Tibshirani(1998) developed this strategy and applied it to many other areas in their studies. Linear discriminants, K nearest neighbors and support vector machine were used as classifiers. In this paper, we construct pairwise classifiers for multi class problems by using binary logit regressions and compare the results with those of classical multinomial logit model. We use the forest cover type data from US Forest Service inventory information and predict forest cover types from cartographic variables.