A Strategy of Multi-level Classification Based on SVM
Xiaoou Chen · Jisuanji gongcheng · 2005
This paper proposes a new multi-level classification strategy named 1-vs-brothers on the basis of 1-vs-rest strategy which is used to transfer multi-category problems to two category problem. Compared with the original strategy, the new strategy which is based on the selection of negative examples, only selects the example documents of brother nodes as negative examples, that cuts down the documents number needed to learn during the non-first level nodes training period. The experiment shows that on the data of this paper the algorithm based on this strategy improves the training efficiency about 60 percent, and the classification precision yet remains no change on the whole. This strategy can also be used on 1-vs-1 strategy to form 1-vs-brother strategy, which will cut down the node numbers needed to learning during the training period of multi-level classification.