POLL : multiclass classification from binary classifiers through random sampling
msra
摘要
We apply a sampling technique in combinatorial optimizations to the problem of multiclass classification. Various methods have been used to construct a multiclass classifier from a set of binary classifiers. Suppose that there are n classes. A simple majority vote (known as max-win), while giving very reliable answers, need a quadratic number of comparisons. Other methods, by cascading binary classifiers, reduce the number of comparison to O(n) but suffer from the build-up of the failure probability. This implies that they do not scale well with n. We show that by using random sampling, we need only O(n log n) comparisons while retaining the good performance of the majority method.
更多查看译文
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络